git-aiGit extension that records which agent, model and prompt wrote each line2.9k starsApache-2.0last commit Oct 2026
37
100
88
70
75 out of 100score
#106
native-subtitle-quote-imageAgent skill that stitches video frames and subtitle strips into long images2.7k starsMITlast commit Oct 2026
35
100
100
40
75 out of 100score
#107
agentskillsOpen specification for SKILL.md folders that give AI agents reusable instructions26.5k starsApache-2.0last commit Aug 2026
85
100
39
70
74 out of 100score
#108
agent-skills-for-context-engineeringCollection of 18 Agent Skills on context and harness engineering for coding agents18.2k starsMITlast commit Oct 2026
81
100
41
70
74 out of 100score
#109
loop-engineeringPattern library and CLI for running scheduled agent loops around a codebase11.6k starsMITlast commit Oct 2026
75
100
61
30
74 out of 100score
#110
skills-managerDesktop app and CLI that manage agent skills across coding tools6.8k starsMITlast commit Oct 2026
64
100
82
0
74 out of 100score
#111
skills.NET team's skills and plugins for Copilot, Claude Code, Cursor and Codex5.7k starsMITlast commit Oct 2026
60
99
77
30
74 out of 100score
#112
cc-skills-golangAgent Skills that teach coding assistants idiomatic, production-grade Go3.6k starsMITlast commit Oct 2026
47
100
88
40
74 out of 100score
#113
dashi-taskboardLocal issue board embedded in the Codex desktop app, with a CLI and Skill3.4k starsApache-2.0last commit Oct 2026
45
100
93
30
74 out of 100score
#114
cc-safety-netPre-execution hook that blocks destructive commands and secret access by coding agents1.8k starsMITlast commit Oct 2026
20
100
99
85
74 out of 100score
#115
ag-kitRules, skills, agents and workflows for Google Antigravity projects8.2k starsMITlast commit Aug 2026
68
100
62
40
73 out of 100score
#116
skillsTrail of Bits plugin marketplace of security and testing skills for Claude Code7.8k starsCC-BY-SA-4.0last commit Oct 2026
67
100
52
70
73 out of 100score
#117
n8n-skillsClaude Code skills and hooks for building n8n workflows through n8n-mcp6.5k starsMITlast commit Oct 2026
64
100
67
40
73 out of 100score
#118
agents-cliCLI and skills that teach coding agents to build and deploy ADK agents6.2k starsApache-2.0last commit Oct 2026
62
100
81
0
73 out of 100score
#119
apmDependency manager for agent skills, plugins, instructions and MCP servers4.1k starsMITlast commit Oct 2026
52
100
91
0
73 out of 100score
#120
cc-sddSpec-driven workflow skills for Claude Code, Codex, Cursor and other coding agents3.7k starsMITlast commit Sep 2026
48
100
72
70
73 out of 100score
#121
ripwireCode-context CLI giving coding agents a ranked call graph and edit checks2.5k starsApache-2.0last commit Oct 2026
31
100
83
85
73 out of 100score
#122
ffmpeg-skillAgent Skill and MCP server that edits video with local FFmpeg2.3k starsMITlast commit Oct 2026
28
100
99
45
73 out of 100score
#123
brooks-lintCode review skills that cite classic engineering books for each finding1.5k starsMITlast commit Oct 2026
14
100
100
85
73 out of 100score
#124
superclaude_frameworkSlash commands, agents and modes that add structure to Claude Code23.9k starsMITlast commit Sep 2026
84
97
30
85
72 out of 100score
#125
memuShared memory and skill store for coding agents across sessions and devices14.6k starscustom licenselast commit Oct 2026
78
97
55
30
72 out of 100score
#126
autoresearchSkill that runs coding agents in a measured keep-or-revert loop6.7k starsMITlast commit Aug 2026
64
100
60
45
72 out of 100score
#127
lazycodexOmO agent harness packaged as a plugin for OpenAI Codex3.8k starsMITlast commit Oct 2026
49
100
82
30
72 out of 100score
#128
one-skill-to-rule-them-allClaude meta-skill that logs corrections and proposes updates to your other skills3.5k starscustom licenselast commit Oct 2026
46
100
94
0
72 out of 100score
#129
pstack-claudeSkill stack that routes coding-agent tasks through planning, delegation and verification3.4k starsMITlast commit Oct 2026
44
100
95
0
72 out of 100score
#130
stop-that-shitSkill and guard that stop coding agents from scope creep and needless defensive code3.1k starsMITlast commit Oct 2026
41
100
98
0
72 out of 100score
#131
jevgrepCLI that answers repo questions with files and source excerpts for coding agents2.9k starsMITlast commit Oct 2026
37
100
89
45
72 out of 100score
#132
claude-blogClaude Code skill suite for writing, auditing and localizing SEO blog posts2.5k starsMITlast commit Oct 2026
31
100
95
40
72 out of 100score
#133
skills-hubDesktop app that installs AI agent Skills once and syncs them to coding tools1.8k starsMITlast commit Oct 2026
21
100
95
70
72 out of 100score
#134
dsh-routing-suitePlugins for the DSH agent: runtime injector, routing presets, graded task protocol7k starsMITlast commit Sep 2026
65
100
72
0
71 out of 100score
#135
claude-code-localAnthropic-API MLX server that runs Claude Code on local models on Apple Silicon3.4k starsMITlast commit Oct 2026
44
100
92
0
71 out of 100score
#136
lintESLint and Oxlint rules that enforce Tailwind design systems with fix guidance for agents3.3k starsMITlast commit Oct 2026
43
100
95
0
71 out of 100score
#137
rulerCLI that syncs one set of Markdown rules to many coding agents' config files3k starsMITlast commit Oct 2026
38
100
90
30
71 out of 100score
#138
unity-skillsUnity Editor plugin exposing 839 REST skills for AI coding agents to drive scenes1.9k starsMITlast commit Oct 2026
23
100
89
70
71 out of 100score
#139
pg-aiguidePostgres skills and doc-search MCP server for AI coding agents1.9k starsApache-2.0last commit Oct 2026
22
100
99
40
71 out of 100score
#140
mexGit-shared project memory for engineers and their coding agents1.8k starsMITlast commit Oct 2026
20
100
90
85
71 out of 100score
#141
skillsAgent Skills for Google Cloud, Firebase, Ads and Analytics21.5k starsApache-2.0last commit Oct 2026
83
100
50
0
70 out of 100score
#142
vibe-skillsSkill router that plans tasks and assigns local agent Skills to each step3.8k starsApache-2.0last commit Oct 2026
49
100
84
0
70 out of 100score
#143
memsearchPersistent semantic memory for Claude Code, Codex, OpenClaw and OpenCode agents2.8k starsMITlast commit Oct 2026
36
100
85
40
70 out of 100score
#144
token-optimizerPlugin that trims wasted context and keeps state across compaction in coding agents2.7k starscustom licenselast commit Oct 2026
skill-codexClaude Code skill that delegates tasks to the Codex CLI1.5k starsMITlast commit Oct 2026
13
100
20
0
40 out of 100score
#292
app-store-preflight-skillsAgent skill that checks iOS/macOS projects for App Store rejection causes1.4k starsMITlast commit Oct 2026
10
100
16
0
38 out of 100score
#293
evoPlugin that runs parallel optimization experiments on a codebase via coding agents1.5k starsApache-2.0last commit Oct 2026
13
100
11
0
37 out of 100score
#294
harmonistHook-enforced agent protocol pack for Cursor, Claude Code and other coding assistants2.3k starsMITlast commit Jun 2026
28
88
0
0
35 out of 100score
#295
claude-code-proxyProxy that lets the Claude Code CLI use OpenAI-compatible APIs2.8k starsMITlast commit Mar 2026
36
51
11
40
33 out of 100score
#296
myclaudeClaude Code slash-command workflows that delegate coding to Codex, Gemini or OpenCode2.8k starsAGPL-3.0last commit May 2026
35
74
0
0
33 out of 100score
#297
cli-continuesCLI that moves a coding session between 16 AI coding tools1.6k starsMITlast commit May 2026
16
71
0
70
33 out of 100score
#298
paulPlan-Apply-Unify workflow commands for structured development in Claude Code1.3k starsMITlast commit Aug 2026
5
97
7
0
33 out of 100score
#299
agent-toolkitClaude Code skills, sub-agents and slash commands, installable as plugins2.6k starsMITlast commit Mar 2026
33
48
0
70
31 out of 100score
#300
skillkitPackage manager that installs and translates agent skills across 46 coding agents1.6k starsApache-2.0last commit Jun 2026
15
86
4
0
31 out of 100score
#301
guard-skillsReview-pass skills that check AI-written code, tests and docs before merge1.3k starsMITlast commit Jul 2026
5
92
0
0
29 out of 100score
#302
continuous-claude-v3Claude Code add-on for session handoffs, memory, and a library of skills and agents3.9k starsMITlast commit Jan 2026
51
34
0
15
27 out of 100score
#303
taches-cc-resourcesClaude Code commands, skills and auditor subagents for planning and debugging2k starsMITlast commit Apr 2026
25
58
0
0
25 out of 100score
#304
mcphub.nvimMCP client plugin for Neovim that connects MCP servers to chat plugins1.8k starsMITlast commit Jan 2026
21
29
17
0
20 out of 100score
#305
tutor-skillsClaude Code skills that build Obsidian study vaults and quiz you on them1.4k starsMITlast commit Feb 2026
8
46
0
0
16 out of 100score
#306
vibesec-skillSecurity skill that steers AI coding agents away from common web vulnerabilities1.4k starsApache-2.0last commit Feb 2026
8
42
0
0
15 out of 100score
#307
vibekitCLI that runs coding agents in sandboxes with secret redaction1.9k starsMITlast commit Nov 2025
22
12
0
0
10 out of 100score
#308
sudolangPseudocode language for writing structured prompts that LLMs interpret directly1.4k starsMITlast commit Dec 2025
8
22
0
0
9 out of 100score
Momentum, Verified build, Docs and Privacy are not measured yet; their weight goes to the signals shown. A dash means the signal is not scored for that kind of project. Hover a number for its rating in words.
Impeccable installs a single `/impeccable` skill into coding agents such as Claude Code, Cursor, Codex CLI, Gemini CLI and GitHub Copilot. It adds 24 commands (audit, critique, polish, typeset, layout and others), a live browser variant mode, and 61 deterministic detector rules for common AI-generated design tells. `/impeccable init` writes a PRODUCT.md that later commands use as context.
Strengths
Installs into many agents: Claude Code, Cursor, Codex, Gemini CLI, Copilot, Grok Build, Hermes and more
61 detector rules run in the CLI and browser extension with no LLM or API key
Edit hooks surface detector findings back to the agent on Claude Code, Copilot, Codex, Cursor and Grok
Apache-2.0, with npx, git submodule, plugin and manual copy install options
Weaknesses
Skill runs through a separate binary downloaded on first run into ~/.impeccable/bin/
Design roll requests go to impeccable.style and telemetry is on unless you opt out
Hooks need per-harness trust approval, and Hermes and Veto get no edit hook
Writes many working files under .impeccable/ that need a manual .gitignore block
no GPU
Needs Node.js (only for the npx installer), An AI coding tool such as Claude Code, Cursor or Codex CLI, impeccable.style roll API (telemetry can be disabled)
Caveman is a CLI and skill set that reduces token use in coding agents such as Claude Code, Codex, Gemini CLI and OpenCode. It shortens agent replies, compresses logs, JSON, CSV and test output before the model reads them, and can route each request to a model and effort level through a free account. Reported results include 33.2% fewer input tokens over 54 Claude Code runs and an interim 54% cost cut for Auto on 85 tasks.
Strengths
One npx command sets up 30+ agents; runs locally with your own keys
README lists losing cases: HTML input 9.9% worse, grep-defs loses to bare grep
Each part can be switched off separately; the skill installs alone
Cited by Adobe Research paper, JetBrains A/B on 86 tasks, Elastic blog
Weaknesses
Auto routing needs a Caveman account; its benchmark is interim and half complete
Terse output saves nothing on per-request billing such as GitHub Copilot
Rules cost about 1,160 input tokens; output benchmarks used older 3.1.0 rules
No HTML compressor yet; Blocks is pre-release (v0.1.0-rc.5), Python only
no GPU
Docker + Compose
Needs Node.js 22.13+, Caveman account (Auto only)
Models: Claude Code, Codex, Gemini CLI, OpenCode, Aider
Agent skills for planning, TDD, debugging and code review
302.1k starsMITlast commit Oct 2026
A collection of Markdown skill files for coding agents, split into engineering and productivity groups. It covers grilling sessions that build a project glossary and ADRs, spec and ticket generation, red-green-refactor TDD, a bug diagnosis loop, and diff review. Installs as a plugin for Claude Code, Codex and GitHub Copilot, via gemini skills install for Gemini CLI, or as editable files through npx skills.
Strengths
Small, composable skills that can be edited; README says they work with any model
Plugin install for Claude Code, Codex and Copilot, with automatic updates
Setup skill configures issue tracker (GitHub, GitLab, local files), triage labels and doc locations
Separates user-invoked orchestration skills from model-invoked discipline skills
Weaknesses
Manual updates for Gemini CLI and npx installs
Installing both plugin and npx copies gives every skill twice
Must run /setup-matt-pocock-skills once per repo before the engineering skills
Architecture skill is a survey only and will not untangle an old codebase
no GPU
Needs Issue tracker (GitHub, GitLab or local files)
Agents, skills, hooks and rules bundle for Claude Code, Codex and other harnesses
287.5k starsMITlast commit Oct 2026
ECC installs a plan, test, implement, review, verify workflow into coding agents. It ships 71 agents, 311 skills, 97 legacy command shims, hooks, selectable rules, session memory and the AgentShield scanner for prompts, hooks, MCP config and permissions. Native plugin paths exist for Claude Code and Codex; Cursor, OpenCode, Gemini, Zed and others get capability-limited adapters.
Strengths
Large bundle: 71 agents, 311 skills, hooks and rules in one install
Guided setup wizard for Claude Code, Codex and Kimi Code with dry-run mode
Includes AgentShield scanning of MCP config, permissions, hooks and secrets
MIT licensed; npm releases pinned per version tag
Weaknesses
Feature parity varies; non-Claude harnesses are capability-limited adapters
Claude plugins cannot ship rules, so rule packs need a manual copy
Stacking install methods duplicates skills and hooks; README warns against it
Private-repo GitHub App is paid (from $19/seat/month)
no GPU
Docker + Compose
Needs Node.js 18+, Git, Claude Code 2.1+ (for plugin setup)
Models: Claude Code, Codex, Kimi Code, Cursor, OpenCode
Multi-agent orchestration plugin and CLI for Claude Code
40.2k starsMITlast commit Oct 2026
Adds staged team pipelines (plan, PRD, exec, verify, fix), autopilot and persistent execution loops to Claude Code, as a plugin and as an `omc` npm CLI. `omc team` can launch tmux worker panes running claude, codex, gemini, antigravity, grok or cursor-agent CLIs. Configuration lives in `.claude/omc.jsonc`.
Strengths
Installs as a Claude Code plugin or as an npm CLI (`omc`)
Team workers can be claude, codex, gemini, antigravity, grok or cursor CLIs in tmux
Workers spawn per task and exit when done, so nothing idles
Includes a deep-interview skill and a HUD statusline
Weaknesses
Requires Claude Code; in-session team mode needs an experimental flag enabled
Slash-command skills such as /autopilot and /ralph cannot run in CI
Named autopilot workflow profiles need Linux with flock
npm package name (oh-my-claude-sisyphus) differs from repo name; install prints a deprecation warning
no GPU
Needs Claude Code, Node.js/npm, tmux (for omc team), codex, gemini, agy, grok or cursor-agent CLIs (optional)
Design-system skill for AI coding assistants, with searchable styles, palettes and fonts
139.6k starsMITlast commit Oct 2026
Installs as a skill or plugin for coding assistants such as Claude Code, Cursor, Windsurf, Codex CLI and Gemini CLI. A Python script using BM25 search and 192 reasoning rules returns a design system for a given product type: landing pattern, style, colors, typography, effects, anti-patterns and a pre-delivery checklist. The data covers 79 UI styles (50 active), 192 palettes, 74 font pairings and guidelines for 22 stacks.
Strengths
Installs into 20+ assistants via the uipro CLI, or as a Claude Code plugin
Search script uses only the Python standard library and makes no network calls
Output includes anti-patterns and an accessibility-oriented checklist, not only palettes and fonts
Covers 22 stacks, from React and Vue to SwiftUI, Flutter and WPF
Weaknesses
Requires Python 3 on the machine to run the search script
Brand identity, logo and asset generation are in the paid premium version only
Only 50 of 79 styles are active in default recommendations
Output quality depends on the host assistant following the skill; no tests or benchmarks stated
no GPU
Needs Python 3, Node.js and npm (for the uipro CLI)
Converts books and document folders into on-demand agent skills with per-chapter files
35.9k starsMITlast commit Oct 2026
A Python extractor turns PDF, EPUB, DOCX, Markdown, HTML, RTF or MOBI files into clean text. The agent then follows SKILL.md to generate a skill: a core SKILL.md, one file per chapter, a glossary, patterns and a cheatsheet. The skill installs under ~/.agents/skills/ and loads chapters on demand in Claude Code, Copilot CLI, Amp, Codex, Hermes Agent, OpenCode or OpenClaw. The README reports 24x–51x fewer tokens than dumping the book into context.
Strengths
Reads PDF, EPUB, DOCX, MD, HTML, RTF and MOBI, with fallbacks to stdlib parsers
One SKILL.md output works across seven agent hosts that follow the Agent Skills format
Chapter files load only when asked, keeping the base skill near 4,000 tokens
Extraction runs locally; `extract.py --check` reports which extractors are missing
Weaknesses
Scanned PDFs are rejected; OCR (for example ocrmypdf) must be run first
Generation quality depends on the host agent and its model, not the extractor
MOBI/AZW needs Calibre installed separately
Sharing skills built from copyrighted books can infringe; README advises keeping them private
Markdown engineering skills and slash commands for AI coding agents
110.1k starsMITlast commit Oct 2026
A pack of 25 SKILL.md workflows covering spec, plan, build, test, review and ship, plus 9 slash commands, 4 reviewer personas and reference checklists. Each skill has steps, verification gates and an anti-rationalization table. It installs through the skills CLI (`npx skills add`) or natively in Claude Code, Codex, Gemini CLI, Antigravity and Command Code.
Strengths
Installs into 70+ agents with one `npx skills add` command
Skills are plain Markdown, usable with any agent that reads instruction files
Persistent cross-session memory plugin for Claude Code and other coding agents
102.6k starsApache-2.0last commit Oct 2026
Claude-mem hooks into Claude Code's lifecycle events (SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop, SessionEnd) to capture tool-usage observations, summarize them, and inject relevant context into later sessions. Data is stored in SQLite with FTS5 and Chroma for hybrid search, served by a local Bun-managed worker with a web viewer. Search is exposed through three MCP tools (search, timeline, get_observations). Installers also exist for OpenCode, T3 Code, Antigravity, OMP, Pi and OpenClaw.
Strengths
Local SQLite plus Chroma storage with a worker HTTP API and web viewer
Three-step search/timeline/get_observations flow filters before fetching full details, cutting tokens
Installers for several harnesses besides Claude Code, including OpenCode, T3 Code and OpenClaw
`<private>` tags exclude sensitive content from storage
Weaknesses
Installer prompts for sign-in to a hosted observer with a 30-day trial; opt-out needs flags
Depends on Node 20+, Bun and uv, auto-installed if missing
`npm install -g claude-mem` gives only the SDK, not the hooks or worker
Some features, such as TeamBrain and cloud sync, rely on the hosted cmem.ai service
Ruflo wraps Claude Code and Codex with a harness: swarm coordination, vector-backed memory, hooks that route tasks, and an MCP server. It installs either as Claude Code plugins (35 listed) or via `npx ruflo init`, which adds 98 agents, 60+ commands, 30 skills and a daemon. A console (Claude Code plugin or `npx ruflo tui`) shows agents, cost and missions.
Strengths
Two install paths: zero-file Claude Code plugins or full CLI init with hooks and daemon
Exposes an MCP server, so Codex and other MCP clients can use it
Standalone terminal UI runs without Claude Code (`npx ruflo tui`)
Console Claude-control defaults to read + ask, with confirmation for spending and destructive actions
Weaknesses
Large surface (314 MCP tools, 26 CLI commands, 35 plugins) is hard to audit
Plugin path gives only a subset; full loop needs the CLI install that writes files into the repo
Mods are not sandboxed and run with your account's permissions
Hook and console support depends on client; Claude Code 2.1.287 or later required for plugins
Plugin of 36 skills for coding agents: plan, build, review, capture learnings
25.7k starsMITlast commit Oct 2026
A plugin of 36 skills for AI coding agents that runs a loop of brainstorm, plan, work, simplify, review and compound. The compound step writes learnings to docs/solutions/, which later brainstorm and plan runs read as context. It installs on 14 agent hosts, including Claude Code, Cursor and Codex, and includes an autonomous /lfg pipeline.
Strengths
Runs on 14 agent hosts, including Claude Code, Cursor, Codex, Copilot and OpenCode
Captured learnings land in docs/solutions/ as plain files in the repo
Covers planning, multi-agent code review, git and PR workflow, and browser testing
Artifact folder location is configurable through the docs_root setting
Weaknesses
Invocation syntax differs by host ($skill-name in Codex, /skill:name in omp)
Pi needs companion packages such as pi-subagents for subagent workflows
Compound Packs for org-wide rules are marked experimental
Some skills use Claude-style allowed-tools names Devin does not map, so it prompts
no GPU
Needs An AI coding agent host (Claude Code, Cursor, Codex, etc.)
Agent skill that makes coding agents write less code
175.1k starsMITlast commit Oct 2026
Ponytail is a single prompt (SKILL.md, with a compact AGENTS.md version) that tells a coding agent to reuse existing code, the standard library or native platform features before writing new code. It adds slash commands for diff review, repo audit and deferred-shortcut tracking. In Claude Code it also shows a cost and token band and a local board. It installs as a Claude Code or Codex plugin, or as a rules file for other agents.
Strengths
README badge says it works with 20 agents; plugin installs for Claude Code and Codex, rules file for others
Own benchmark of 39 tasks, 5 runs each, reports -53% code and 98% of risky logic shipping with a test
Keeps validation, security and accessibility; leaves one small test for logic with branches
No config file required; level set with /ponytail lite, full, ultra or off
Weaknesses
Benchmark is the author's own, run only in Claude Code with Opus 5.5
Slash commands need a skill-capable host; Cursor rules, Windsurf, Cline and Copilot get only the ruleset
Band and board need Claude Code 2.1.290+ and do not show in the VS Code panel or claude -p
Behavior depends on the agent following a prompt, so results are not deterministic
no GPU
Needs Claude Code, Codex or another supported coding agent
Models: Claude Code, Codex, Devin CLI, OpenCode, Gemini
Spec-driven development toolkit and CLI for AI coding agents
142.6k starsMITlast commit Oct 2026
Spec Kit installs a Python CLI (`specify`) that scaffolds a project with slash commands or skills for your coding agent. The core flow is constitution, specify, plan, tasks, implement, converge; bug-fixing and idea-assessment ship as optional extensions. Behavior is customized through extensions, presets, workflows, bundles and catalogs.
Strengths
Works with multiple coding agents through integrations; GitHub Copilot is the documented example
Extensions, presets, workflows and bundles customize the process without forking
Bug-fix and assessment extensions write reports to .specify/ directories in the repo
Supports air-gapped use via local sources and self-hosted catalogs
Weaknesses
Needs a separate supported coding agent; it does not run models itself
Steps are run one at a time in agent chat, with manual review
Invocation syntax differs per agent, so examples need translating
Community catalog components are unaudited and need manual source review
no GPU
Needs Python 3.11+, uv, a supported AI coding agent
Archify is a skill for coding agents (Cursor, Claude Code, Codex CLI, OpenCode) that generates self-contained interactive HTML diagrams from a text description or from a repository the agent reads. It also builds one-page explainer answers with tables, steps and validated diagrams. Output has light and dark themes, route and reach tracing, and PNG, WebP, JPG or PDF export.
Strengths
Output is one standalone HTML file; viewing needs no Archify install
Installs with one command: npx skills add tt-a1i/archify -g
Works from a description alone, or from a repo with source excerpts pinned to a commit
Exports PNG, WebP, JPG and PDF, plus route and reach share cards
Weaknesses
Needs a separate AI coding agent to generate diagrams; not standalone
Optional update check contacts a remote manifest unless ARCHIFY_UPDATE_CHECK_DISABLED=1
README carries many sponsor and API-relay promotions
Diagram quality depends on the host agent's model; no benchmarks given
no GPU
Needs Node.js (npx), Claude Code, Codex CLI, Cursor or OpenCode
A catalog of installable SKILL.md playbooks for coding agents such as Claude Code, Codex CLI, Cursor, Gemini CLI, Kiro and OpenCode. AAS Core is a local MCP plus an `aas` CLI that lets the agent search the catalog, record chosen skills in an aas-stack.json manifest and preview a plan. An npx installer copies selected skills into a host's skill directory.
Strengths
2,690+ skills, each a plain SKILL.md file, with a machine-readable skills_index.json
Installer targets many hosts: Claude Code, Cursor, Codex, Gemini CLI, Kiro, OpenCode, Copilot
Install and stack plan steps support dry-run preview before any target is changed
MIT license; 13 domain plugins for Claude Code and Codex
Weaknesses
Core does not rank or recommend skills; the agent must choose
Core apply and recovery are experimental and outside the supported preview
Validation checks structure only, not fit, compatibility or safety of a skill
Installing too many skills can overload Antigravity's context
no GPU
Needs Node.js/npm (npx installer), Codex or Claude Code for AAS Core MCP
Drives NotebookLM (now Gemini Notebook) through its undocumented web endpoints, with an opt-in Android gRPC backend. It exposes notebooks, sources, cited chat, research agents and artifact generation (audio, video, slides, quizzes, mind maps) via an async Python client, a CLI, an MCP server, a REST server and an installable agent skill. Artifacts download in bulk as MP3, MP4, PDF, PPTX, CSV, JSON or Markdown.
Strengths
Covers notebooks, sources, chat, notes, research, sharing and all Studio artifact types
Bulk artifact download in MP3, MP4, PDF, PPTX, CSV, JSON and Markdown
Ships CLI, MCP server (stdio or remote), REST server and Claude Code/Codex skill
Master-token auth refreshes sessions unattended for servers and CI
Weaknesses
Relies on undocumented Google APIs that can break without notice
Not affiliated with Google; heavy use may be rate limited
Needs a Google account; source counts per notebook depend on account tier
Default login downloads Chromium (~170 MB) via Playwright
no GPU
Docker + Compose
Needs Google account, Google NotebookLM / Gemini Notebook, Playwright Chromium (for browser login)
Agent skill that renders styled PDF, PNG and PPTX documents from templates
12.1k starsMITlast commit Oct 2026
Kami is a skill for coding agents (Claude Code, Codex, Cursor) that supplies layout rules and eight document templates: one-pager, long doc, letter, portfolio, resume, slides, equity report and changelog. It renders through WeasyPrint to PDF, with python-pptx for editable PowerPoint and a Marp variant for slides. It also includes a zero-dependency MCP server exposing render, check and screenshot tools.
Strengths
Eight document templates plus a landing page system, in English, Chinese and Korean variants
Installs via npx skills, host plugin marketplaces, or a Claude Desktop zip
Includes a zero-dependency MCP server with render, check and screenshot tools
Structured content schemas and coverage checks run before final output
Weaknesses
Windows requires WSL; helper scripts are Bash and WeasyPrint needs GTK/Pango
Default Chinese font TsangerJinKai02 is free for personal use only
Japanese and Korean get only font fallbacks and layout checks, not full support
Runs a daily version check against GitHub's latest release
Claude Code status line plugin showing context, rate limits, tools and todos
28.8k starsMITlast commit Oct 2026
Claude HUD is a Claude Code plugin that registers a status line command. It reads session data Claude Code passes on stdin (model, context window, cost, rate limits, prompt cache) and prints context and usage bars below the input. Optional tools, agents and todos lines are read from the session transcript and enabled through /claude-hud:configure.
Strengths
Runs as Claude Code's status line; no tmux or separate window needed
Default two-line view shows model, git branch, context bar and rate-limit usage
Many config options: layout, thresholds, colors, element order, cost, git and jj status
Guided /claude-hud:configure flow previews changes; labels in English and Chinese
Weaknesses
Works only with Claude Code, not other coding agents
Needs a JavaScript runtime; Windows users may have to install Node.js LTS
Third-party plugin updates are not automatic until auto-update is enabled
Library of 398 agent skills and plugins for Claude Code and 12 other tools
28.6k starsMITlast commit Oct 2026
A collection of SKILL.md instruction packages, each paired with stdlib-only Python CLI scripts and reference docs. Domains include engineering, product, marketing, compliance, C-level advisory and research. It installs as Claude Code marketplace plugins, and scripts convert or sync the skills for Codex, Gemini CLI, Cursor, Aider, Windsurf and others.
Strengths
Installs per domain or per skill through the Claude Code plugin marketplace
Python tools are stdlib-only, so no pip installs are needed
convert.sh and install.sh target Cursor, Aider, Windsurf, Kilo Code, OpenCode and Augment
Covers non-code domains such as regulatory, finance and research alongside engineering
Weaknesses
README skill and tool counts disagree (345 vs 398 skills, 706 vs 766 tools)
Hermes Agent and Mistral Vibe need a manual local sync script
Windows needs symlink-enabled clone and PYTHONUTF8 or mirror trees break
Quality of individual skills is unverified; much of the catalog is outside coding
no GPU
Needs Python, Claude Code or another supported coding agent
Rust CLI that compresses shell command output before coding agents read it
84.4k starsApache-2.0last commit Oct 2026
RTK is a single Rust binary that rewrites shell commands such as git, pytest, cargo test and docker ps to filtered equivalents, so the agent reads grouped, truncated or deduplicated output. It installs hooks or rules for 18 coding tools including Claude Code, Cursor, Gemini CLI and Codex. An optional BUSL-licensed proxy daemon extends the same pipeline to provider API traffic.
Strengths
Single binary with 100+ command filters and stated overhead under 10ms
Per-agent setup via `rtk init`, covering 18 tools including Claude Code, Cursor and Codex
`rtk gain` and `rtk discover` report savings and missed opportunities
Telemetry is off by default and sends nothing until opted in
Weaknesses
Hook only fires on Bash tool calls; built-in Read, Grep and Glob bypass it
Token counts are estimated as bytes/4; only percentages are reliable
Mixed licensing: engine and proxy are BUSL-1.1, not Apache-2.0
Full and oss editions share the binary name and cannot be installed together
Spec-driven workflow layer that adds proposal, spec and task files to AI coding
72.9k starsMITlast commit Oct 2026
OpenSpec is a Node.js CLI that installs slash commands (/opsx:explore, /opsx:propose, /opsx:apply, /opsx:archive) into 30+ AI coding assistants. Each change gets a folder with proposal.md, specs, design.md and tasks.md in plain Markdown, which you review before code is written. A beta "stores" feature keeps specs and changes in a separate repo shared across several code repos.
Strengths
Works with 30+ coding assistants through generated slash commands, not a dedicated IDE
Specs and changes are plain Markdown in the repo, reviewable in git
Artifacts can be edited at any time, with no enforced phase gates
Stores (beta) share one plan across multiple repos via git push
Weaknesses
Output quality depends on the model; README recommends Codex 5.5 or Opus 4.7
Stores for cross-repo planning are still in beta
Collects anonymous command and version telemetry by default; opt-out required
Slash command spelling differs per tool, so instructions vary by assistant
no GPU
Needs Node.js 20.19.0 or higher, A supported AI coding assistant
Models: Tool-agnostic; recommends Codex 5.5 and Opus 4.7
Repomix reads a local or remote Git repository and writes a single file, XML by default or Markdown, JSON or plain text, with a directory tree and file contents. It reports token counts, respects .gitignore, .ignore and .repomixignore, and uses Secretlint to leave out files matching known credential formats. The --compress option uses Tree-sitter to keep code structure while cutting tokens.
Strengths
One command (npx repomix@latest) produces the packed file with no install
Token counts per file and for the whole repo help fit context limits
--compress uses Tree-sitter to reduce tokens while keeping code structure
Packs remote repos by URL, branch or commit, and file lists piped via --stdin
Weaknesses
Output is a static snapshot; you paste or upload it to the model yourself
Secret detection only covers known credential formats via Secretlint
VSCode extension is community-maintained, not part of the core project
Whole large repos can exceed model context unless filtered or compressed
no GPU
Docker
Needs Node.js
Models: Model-agnostic: output is a file for Claude, ChatGPT, Gemini, Llama, Grok and others
Agent skill that keeps task plans in markdown files on disk across /clear
27.6k starsMITlast commit Oct 2026
Skill and hook set for coding agents that stores the plan in task_plan.md, findings.md and progress.md in the project. Lifecycle hooks re-inject selected plan context each turn, and a completion gate can hold the agent's stop until the plan reports complete. Installs through `npx skills add` for the Agent Skills standard, with native plugins for Claude Code, Codex CLI, Qoder CLI, Pi, Hermes Agent, OpenCode and DeepSeek Harness.
Strengths
Plan state is plain markdown on disk, so it survives /clear, compaction and crashes
Works across many hosts: native plugins for 7 agents, Agent Skills standard for the rest
Completion gate can block the agent from stopping until phases are marked complete
Parallel tasks get isolated .planning/YYYY-MM-DD-slug/ directories (v2.36.0+)
Weaknesses
Plan files are gitignored and overwritten by the next task unless you promote them
Hook features vary per host; Qoder CLI needs Bash and Python 3.10+
Completion gate matches literal English status tokens via grep -F
Benchmark evidence is small: 30 assertions and 3 blind A/B comparisons, run by the author
Shared long-term memory server for coding agents, stored as git-backed markdown
13.1k starsMITlast commit Oct 2026
ai-memory is a Rust server that records coding-agent sessions through lifecycle hooks and MCP, consolidates them into markdown wiki pages in git, and injects a brief into the next session. Handoffs let you move from Claude Code to Codex or another supported agent, across machines or team members. Capture, FTS search and handoffs work with no LLM call; vectors and LLM consolidation are optional.
Strengths
Markdown wiki in git is the source of truth; the database index can be rebuilt
Default path needs no LLM or API key: capture, FTS search, handoffs
Over twenty agent harnesses supported, with claim-once typed handoffs between them
Multi-user auth, attribution and audit log included, not a paid tier
Weaknesses
Native Windows support is experimental; WSL2 is the supported route
Docker quick-start runs with no authentication unless a token is added
Compaction, dedup and LLM dream pass are off by default and need opt-in
Several agents (Claude Desktop, Zed, VS Code Copilot) are MCP-only, without hooks
no GPU
Docker + Compose
Needs SQLite, git, systemd or launchd (native service)
Models: Anthropic (optional LLM), OpenAI (optional embeddings), local embeddings (optional)
Desktop app that switches API providers and manages MCP and Skills for coding CLIs
148.2k starsMITlast commit Oct 2026
CC Switch is a Tauri 2 desktop app for Windows, macOS and Linux, written in Rust. It switches API providers for Claude Code, Claude Desktop, Codex, Gemini CLI, Grok Build, OpenCode, OpenClaw, Hermes Agent, Pi and MiniMax Code, so you don't hand-edit their JSON, TOML or YAML config files. It also manages MCP servers, Skills and Prompts from one place.
Strengths
Covers ten coding agents and CLIs from a single provider list
Removes manual edits of JSON, TOML and YAML config files
Also manages MCP servers, Skills and Prompts in the same app
Native installers for Windows, macOS and Linux; MIT licensed
Weaknesses
README opens with a long sponsor list of API relay services
Desktop GUI only; no Docker image or headless server mode
Config-file format details and proxy behavior are not described in the README excerpt
Changes the config files of each supported tool, so a bug can affect all of them
no GPU
Models: Claude Code, Claude Desktop, Codex, Gemini CLI, Grok Build
Skill that maps a codebase and docs into a queryable knowledge graph
134k starsApache-2.0last commit Oct 2026
Graphify installs as a skill in AI coding assistants (Claude Code, Cursor, Codex, Gemini CLI and 20+ others). Running /graphify builds a knowledge graph of code, docs, PDFs, images and video, written to graph.html, GRAPH_REPORT.md and graph.json. Code is parsed locally with tree-sitter across about 40 languages, and the graph is queried with query, path and explain commands instead of grepping files.
Strengths
Code parsing uses tree-sitter AST locally: no LLM, nothing leaves the machine
Every edge is tagged EXTRACTED or INFERRED, so provenance is checkable
One install registers the skill for 20+ assistants on macOS, Linux and Windows
Optional extras add an MCP stdio server, Neo4j and FalkorDB push, and watch mode
Weaknesses
Docs, PDFs, images and video need an LLM backend or API key for the semantic pass
Cross-repo links, formal verification and Sentry/Jira connectors need paid Graphify Cloud
PDF, video, MCP and database support are separate extras to install
Benchmark figures are self-reported; the README only links to BENCHMARKS.md for details
no GPU
Docker
Needs Python 3.10+, uv or pipx
Models: Claude, OpenAI and compatible APIs, Gemini, Ollama, AWS Bedrock
Agent Skill that makes coding agents design distinctive websites and app screens
101.7k starsMITlast commit Oct 2026
Taste Skill is an Agent Skill (design-taste-frontend) that gives a coding agent an art direction, open-licensed photographs with credits, and a real-browser check of its own output. A script deals three of 20 researched directions, the agent builds, then inspects the page at three widths and fixes the top three problems over up to three rounds. Since v2.1 it also handles dashboards and app screens with separate style and checks.
Strengths
Framework-agnostic: rules target design, so React, Vue, Svelte and plain HTML all work
Installs via npx skills, or as a Claude Code or Codex plugin
Checks pages in a real browser at three widths and fails on generic output
Images come from open collections (Wikimedia Commons, Openverse, Met) with credits placed on the page
Weaknesses
A build takes roughly 30 to 45 minutes and a few hundred thousand tokens
Requires Node 22+ and a local Chrome, Edge or Chromium
Picture search needs network access
Output quality depends on the host coding agent; no benchmarks given
no GPU
Needs Node 22+, Chrome, Edge or Chromium, Network access for image search
Agile planning and delivery skills for AI coding tools
54.5k starscustom licenselast commit Oct 2026
BMad Method installs a set of skills and agent workflows into your AI coding tool, covering clarification, planning, architecture, build, verification and retrospectives. Process depth scales with the change: small changes go straight to build, larger ones get planning documents that carry forward as context. Stable v6.12 installs via npx; the v7 preview alpha installs via the Skills CLI or Claude Code and Codex plugin marketplaces.
Strengths
Process sizes to the change: direct build for small edits, deeper planning for large work
Works on new projects and on existing codebases
Official modules for builder, test architecture, game dev and unattended epic runs
Installs through npx, the Skills CLI, or Claude Code and Codex plugin marketplaces
Weaknesses
V7 is preview alpha; README warns of breaking changes between updates
V7 requires an AI coding tool that supports skills
Needs uv and Python for setup and scripts, plus Node.js for the stable installer
Detected license is NOASSERTION while the README states MIT
no GPU
Needs uv, Node.js 20.12+, Python 3.10+, AI coding tool with skills support
Routes AI coding agents to reverse-engineering and pentest playbooks
43.2k starsMITlast commit Sep 2026
reverse-skill is a skill/routing package for AI coding agents (Claude Code, Codex, Cursor, OpenCode) that, when the agent hits an APK, binary, obfuscated JS, CTF challenge or pentest target, picks the right methodology and tool (jadx, apktool, Frida, IDA, radare2, Ghidra, BurpSuite) instead of guessing commands. Routing is driven by a single JSON config (43-44 rules) validated by a 173-case regression suite and CI on Windows and Ubuntu, covering mobile, firmware, malware, pwn, EDR bypass and API security.
Strengths
Client-neutral routing core works with Claude Code, Codex, Cursor and OpenCode
Single routing.json source of truth with a 173-case regression benchmark run in CI
Workflow layer for OpenAI Codex CLI with skills, roles, and durable state
33.6k starsMITlast commit Oct 2026
oh-my-codex (OMX) wraps the Codex CLI and keeps Codex as the execution engine. It adds skills such as $deep-interview, $ralplan, $ultragoal and $team, specialist roles, scoped AGENTS.md setup, and plans, logs and state stored under .omx/. It launches sessions in tmux with a HUD and can run each session in its own git worktree.
Strengths
Staged flow from requirements interview to planning to durable goals, each stage usable alone
Named git worktree launches isolate concurrent sessions
Plans, logs, memory and mode state persist in .omx/ in the project
Ships a Codex plugin layout alongside the npm CLI
Weaknesses
Only works on top of Codex CLI, which must be installed and authenticated
Tuned for macOS/Linux; native Windows and Codex App get less support
Recommended --madmax flag bypasses Codex approvals and sandbox
Setup scope and AGENTS.md merge policy rules are involved
bd (Beads) is a Go CLI that stores issues as a dependency graph in a Dolt database, so coding agents can list unblocked work with `bd ready`, claim tasks atomically, and close them. It uses hash-based IDs to avoid merge collisions, and syncs across machines with `bd dolt push` and `bd dolt pull`. `bd init` and `bd setup` install guidance and hooks for Claude Code, Codex, Cursor and others.
Strengths
Hash-based IDs (bd-a1b2) avoid merge collisions across agents and branches
Embedded Dolt by default; no separate database server needed for single-writer use
JSON output and `bd ready` give agents unblocked tasks without parsing markdown plans
`bd setup` installs integrations for Claude Code, Codex, Cursor, Factory and more
Weaknesses
Embedded mode allows a single writer; concurrent writers need an external dolt sql-server
Upgrades can need schema migration and a designated clone on remote-backed databases
HTTP backend is experimental and diverges from local mode in documented ways
Older binaries refuse to open a database migrated by a newer schema
Local proxy that lets Codex, Claude Code and Grok Build use any LLM
17.8k starsMITlast commit Oct 2026
opencodex is a local proxy that translates Codex's Responses API into whatever a provider speaks, covering streaming, tool calls, reasoning tokens and images. It also fronts Claude Code, Claude Desktop and Grok Build, ships 40+ built-in providers plus any OpenAI-compatible endpoint, and serves a web dashboard on port 10100. It adds a ChatGPT account pool with quota-aware routing and combos for failover across providers.
Strengths
40+ built-in providers plus any OpenAI-compatible endpoint, chosen with provider/model syntax
One proxy covers Codex, Claude Code, Claude Desktop and Grok Build
Combos give one virtual model id with failover or weighted round-robin
`ocx stop` restores Codex's original configuration; caches bounded by a 256 MiB default budget
Weaknesses
Desktop app is beta; Windows .msi code signing is still rolling out
Cursor provider is marked experimental
ChatGPT account pooling carries provider-policy and suspension risk, per the README's own note
JEV Auto routing is tested only against a mocked endpoint, not a live account
Generates a linked Markdown wiki of your repo for coding agents
17.4k starsMITlast commit Oct 2026
OpenWiki writes a Markdown wiki into `openwiki/` in a repository, tracks each fact back to source evidence, and updates only what changed. It installs as a skill plus MCP tools into coding agents such as Codex, Claude Code, Cursor and Copilot CLI, or runs as a standalone CLI with its own Deep Agents-based writer. Agents query it through local, model-free `openwiki_search` and `openwiki_read`.
Strengths
Installs into ten coding agents, reusing the host's authenticated model session
Claims tracked to source evidence; clean updates skip model calls
Search and read tools are local and read-only; workspaces span linked repos
Resumable page queue; scheduled CI examples for GitHub, GitLab and Bitbucket
Weaknesses
Agent-hosted runs cover repository wikis only; personal wikis need the standalone CLI
Connector-sourced context, including LangSmith, is not yet supported in host runs
Visualizer loads graph and Markdown libraries from a public CDN, so it needs internet
Native CLI needs a separately configured model provider and credentials
no GPU
Needs Node.js 22.22.0 or newer, A coding agent or a model provider
Models: Fourteen providers, including hosted models, Local OpenAI-compatible endpoints
Eight agent skills for planning, debugging, review and writing
7.4k starsMITlast commit Oct 2026
Waza is a set of eight agent skills (/think, /ui, /check, /hunt, /write, /learn, /read, /health), each a folder with reference docs, helper scripts and gotchas. It installs with `npx skills` into the shared ~/.agents/skills directory, as a Claude Code or Codex plugin, or as a Claude Desktop ZIP. Extras include a Claude Code statusline and optional persistent rules.
Strengths
Small fixed set of eight skills, each with one trigger and one job
One install covers Claude Code, Codex, Cursor and other agents reading ~/.agents/skills
Also ships as Claude Code and Codex plugins, a Claude Desktop ZIP and a Pi npm package
Optional rules and a quota statusline are separate installs, not forced on
Weaknesses
Reflects one author's workflow; the README describes no per-skill configuration
Clarity rule is only on main and not in the current release
Extras install by running downloaded shell scripts; the README asks you to review first
Markdown slash-command skills that give Claude Code planning, review, QA and ship roles
137.4k starsMITlast commit Oct 2026
gstack is a set of Markdown skills, exposed as slash commands such as /office-hours, /plan-ceo-review, /review, /qa, /cso and /ship, that split a coding session into roles: product, architecture, design, review, QA, security and release. Setup is a git clone into ~/.claude/skills plus a Bun-based ./setup script that also builds a bundled browser for QA. Claude Code is the full-tier host; Codex, OpenCode, Cursor, Kiro and others are experimental.
Strengths
Claude Code is the only full-tier host; safety hooks like /careful and /freeze are enforced there
Skills are plain Markdown, so they can be forked and edited
Team mode auto-updates for shared repos with no vendored files
Installs for multiple agents (Codex, Cursor, OpenCode, Kiro, Copilot CLI) via --host flags
Weaknesses
Non-Claude hosts are experimental or instruction-only, and safety skills only advise there
Outside reviews need a second CLI (Codex or Claude Code) installed and signed in
/cso needs a recent Bun with specific build flags plus a native C toolchain
README leads with the author's productivity claims, not behavior of the skills
no GPU
Needs Claude Code, Git, Bun 1.3.3+ (1.4.2 tested), Node.js (Windows only), Codex CLI (optional, outside reviews), OpenAI API key (optional, design binary)
Models: Claude (Opus 5.5), OpenAI GPT-6 Astra, OpenAI GPT-6.1 Sol, OpenAI gpt-5.5 and gpt-image-2 (design binary)
Plannotator hooks into coding agents (Claude Code, Codex, Copilot CLI, Gemini CLI, OpenCode, Kiro, Droid, Amp, Pi) and opens their plans, markdown, HTML and diffs in a local browser UI. You annotate, approve or deny, and the structured feedback goes back to the agent. Code review covers local changes and GitHub/GitLab PRs, with Git, GitButler, jj and Perforce support.
Strengths
Covers nine agents through hooks, plugins or skills; one installer auto-detects and configures them
Reviews local diffs and GitHub/GitLab PRs; supports Git, GitButler, jj and Perforce
No usage telemetry; plans, diffs and annotations stay local by default
Plan mode opens the review UI automatically via ExitPlanMode and equivalent hooks
Weaknesses
Release check against GitHub on each page load has no opt-out setting
Droid integration is commands only, with no plan interception
Link sharing is moving to deprecated support; team sharing is a hosted Workspaces product
Setup differs per agent, with manual steps for Amp, Claude Code, Copilot and OpenCode
no GPU
Needs A supported coding agent CLI, Git, GitHub CLI or GitLab CLI for PR review
npx installer that configures Claude Code and Codex with workflows, MCP and API presets
6.1k starsMITlast commit Aug 2026
ZCF is an npx-run CLI that sets up Claude Code and Codex: it installs the tools, writes workflow and agent configuration, and configures API providers or CCR and MCP services. It has an interactive menu plus non-interactive flags such as `npx zcf i -s -p 302ai -k "sk-xxx"`, and supports English, Chinese and Japanese interfaces. Detailed usage lives in the external docs site.
Strengths
Single `npx zcf` entry point, no manual install step
Non-interactive mode with provider preset and key flags for scripted setup
Covers both Claude Code and Codex, including workflow updates via `zcf u`
Interface available in English, Chinese and Japanese
Weaknesses
README gives little detail on what workflows or agents are installed
Requires Node.js and your own API key or provider account
README is mostly sponsor promotions for API relay services
Full usage and options only in external documentation
no GPU
Needs Node.js (npx), Claude Code or Codex, API provider key
Skill set that gives coding agents a spec, plan, TDD and review workflow
304.4k starsMITlast commit Oct 2026
Superpowers is a set of composable skills plus a bootstrap instruction (using-superpowers) that makes a coding agent brainstorm a spec, create a git worktree, write a task plan, then run subagent-driven development with TDD and code review. It installs as a plugin or extension in many agent harnesses, including Claude Code, Codex, Cursor, Gemini CLI, Copilot CLI, OpenCode and Pi. Skills trigger automatically rather than by command.
Strengths
Installs through native plugin commands in about 16 harnesses, including Claude Code, Codex and Cursor
Covers the full cycle: brainstorming, worktrees, plans, subagent execution, TDD, review, branch finishing
Skills trigger automatically, so no slash commands are needed in normal use
Includes a diagnosing-superpowers skill that reads session transcripts to explain misfires
Weaknesses
Must be installed separately in each harness you use
Workflows are mandatory and opinionated; TDD deletes code written before tests
Maintainers generally do not accept contributions of new skills
Optional visual companion loads a vendor logo that reports version, unless disabled by env var
CLI that installs reusable agent skills into Claude Code, Cursor and others
36.3k starsMITlast commit Oct 2026
An npx-run CLI that finds, installs, updates and removes SKILL.md-based agent skills for about 79 coding agents, including OpenCode, Claude Code, Codex and Cursor. Skills install per project by default, or globally with -g. It also generates a prompt for one-off skill use, scaffolds new skills with init, and has an experimental sync for skills bundled in npm dependencies.
Strengths
Supports 4 named agents plus 75 more, with detection of installed agents
Full lifecycle commands: find, add, list, check, update, remove, init
`skills use` runs a skill without installing it
Per-project install by default, global with -g
Weaknesses
Requires Node.js 22.20 or later
npm dependency sync is experimental and may change
Collects anonymous telemetry unless DISABLE_TELEMETRY or DO_NOT_TRACK is set
Electron desktop workbench for Claude Code with diff review and multi-session tabs
15.1k starscustom licenselast commit Oct 2026
cc-haha is a macOS, Windows and Linux desktop app built around a Claude Code-based runtime, with session tabs, per-file diff review, an in-app browser preview and graphical permission approval. It connects to Claude, ChatGPT and Grok accounts, third-party APIs such as DeepSeek, Kimi and GLM, and local LM Studio or Ollama endpoints. It also covers MCP and SubAgent management, Computer Use, IM bridges and scheduled tasks. A CLI can be started from source with Bun.
Strengths
Installers for macOS, Windows and Linux; CLI also runs from source with Bun
Model choice spans Claude, ChatGPT, Grok, preset third-party APIs and local Ollama or LM Studio
GUI management of MCP servers (STDIO, Streamable HTTP, SSE) and SubAgents
Remote control through a phone H5 page and IM apps such as Telegram, Slack and Feishu
Weaknesses
Community Source License, not OSI open source; commercial use and paid services need written permission
README is mostly in Chinese, with the English version in a separate file
Sponsor and API reseller promotions take up much of the README
Spec-first workflow engine that interviews you, then runs and evaluates coding agents
6.3k starsMITlast commit Oct 2026
Ouroboros adds an `ooo` command set to coding-agent CLIs. It runs a Socratic interview that produces an immutable Seed spec with an ambiguity score, then executes, evaluates through a 3-stage gate, and evolves the result, recording actions in a replayable ledger. It registers as an MCP server and supports Claude Code, Codex CLI, OpenCode, Hermes, Gemini, Kiro, Copilot and others.
Strengths
Works across about 14 agent hosts, including Claude Code, Codex CLI, Kiro and Copilot CLI
Interview step yields a locked Seed spec with a measurable ambiguity score
Ships as a Claude Code and Codex plugin, pip package, Homebrew tap or one-line installer
MIT licensed, with recent commits and releases
Weaknesses
Native Windows is experimental, and Codex CLI on Windows needs WSL 2
Python 3.12 or newer required; LiteLLM profiles limited to 3.12-3.13
Many install profiles and MCP v1/v2 splits make setup choices confusing
Needs an existing coding agent or LLM backend; it does not run models itself
no GPU
Needs Python >= 3.12, uv or pipx, A supported coding-agent CLI or LLM backend
Models: Claude Code, Codex CLI, OpenCode, Hermes, Gemini
An Agent Skill that has your coding agent write a diagram as one self-contained HTML file with inline SVG, with no build step or JavaScript. It covers 45 diagram types, each in minimal light, minimal dark and full-editorial variants. It can pull palette and fonts from a site, import Mermaid, draw.io and Excalidraw files, and export to SVG or PNG.
Strengths
Output is one HTML file with inline SVG; opens by double-click, no JavaScript or build step
45 diagram types, each with light, dark and editorial variants
Onboarding step pulls palette and fonts from a URL so diagrams match your brand
Installs via npx skills on Cursor, Claude Code, Codex, Copilot, Gemini CLI and others
Weaknesses
Output quality depends on the model; the same prompt can lay out differently twice
Not an editable source format: no text to diff and no canvas to drag boxes on
Slash commands like /export-diagram and /doctor only come with native plugin packages, not npx install
README says Mermaid is a better fit for docs that change every sprint in Git
no GPU
Needs An Agent Skills-compatible coding agent (Claude Code, Cursor, Codex, etc.), Node.js (npx) for the skills CLI install
Models: Model-agnostic: uses whichever model the host agent runs
Lets coding agents drive your logged-in Chrome over CDP
18.6k starsMITlast commit Oct 2026
Browser Harness connects Claude Code, Codex or Cursor to your running Chrome through the DevTools Protocol, so the agent can open pages, click, type, read and download while logged in as you. A SKILL.md teaches the workflow, and the agent writes missing helpers into a local workspace while the core package stays protected. It also ships a stdio MCP server, browser-harness-mcp.
Strengths
Drives your real, logged-in Chrome rather than a clean profile
Ships both a skill (SKILL.md) and a stdio MCP server
Lowest median input tokens and model cost in its own 100-task eval
Agent-written helpers persist in a local workspace for reuse
Weaknesses
Gives the agent access to the whole browser, including all logins
Needs Chrome 144+ with remote debugging enabled and an Allow click
Eval score not distinguishable from Playwright MCP or Chrome DevTools MCP
Benchmark run with Codex only; other agents not measured
no GPU
Needs Chrome 144+, uv, Python 3.12
Models: Claude Code, Codex, Cursor, any MCP client
Multi-agent orchestration plugin for OpenCode with per-agent model routing
9.8k starsMITlast commit Oct 2026
A plugin for OpenCode that adds an Orchestrator agent plus specialists (Explorer, Oracle, Council, Designer, Fixer). The Orchestrator plans work, dispatches specialists as background tasks, and reconciles their results. Each agent can use a different model, set through presets in ~/.config/opencode/oh-my-opencode-slim.json, with project-local overrides.
Strengths
Per-agent model and variant config, with presets that can extend a base preset
Project-local overrides for prompts, custom agents and delegation rules
Optional tmux, Zellij, kitty or WezTerm panes to watch background agents live
Council agent runs several models in parallel and synthesizes one answer
Weaknesses
Works only inside OpenCode; not a standalone tool
Council is the highest-cost path and triggers only when requested
Bundled example presets go stale as model lineups change, per the README
Requires provider logins and manual model selection per agent
no GPU
Needs OpenCode, Bun or Node.js, LLM provider accounts
Models: OpenAI, OpenCode Go, Claude (via ACP), any provider OpenCode supports
Git worktree manager CLI for running AI coding agents in parallel
9.4k starscustom licenselast commit Oct 2026
Worktrunk is a Rust CLI that wraps git worktree so worktrees are addressed by branch name, with paths computed from a configurable template. Commands like `wt switch -c -x claude feat` create a worktree and launch an agent in it, and `wt merge` squashes, rebases, merges and cleans up. It also offers hooks, an interactive picker, CI status in `wt list --full`, and copy-on-write sharing of build caches on APFS, btrfs and XFS.
Strengths
Create a worktree and start an agent in one command (`wt switch -c -x claude`)
Hooks on create, pre-merge and post-merge automate per-worktree setup
Shares target/ and node_modules/ across worktrees via copy-on-write on APFS, btrfs, XFS
Installs through Homebrew, Cargo, Winget, pacman and conda; shell integration for several shells
Weaknesses
Needs a shell integration step (`wt config shell install`) before it can change directories
On Windows, `wt` clashes with Windows Terminal; Winget installs it as `git-wt`
Cache sharing only works on APFS, btrfs and XFS filesystems
LLM commit messages and summaries need a separately configured model; supported providers unknown from README
Simplified Chinese translation of Matt Pocock's engineering agent skills collection
5.2k starsMITlast commit Oct 2026
Third-party zh-CN localization of mattpocock/skills, a set of small agent skills for grilling interviews, TDD, bug diagnosis, spec and ticket writing, code review and codebase architecture review. Install with `npx skills@latest add`, as a Claude Code plugin, or as a DeepSeek Harness plugin invoked via `/zh-<name>`. Upstream is synced periodically, with translations done by Codex and reviewed by the maintainer.
Strengths
Skills are small, editable files that can be copied into your repo via skills.sh
Works with Claude Code, Codex and other Agent Skills-compatible harnesses
Covers a full flow: grill, spec, tickets, implement with TDD, code review
Directory names, commands and identifiers stay unchanged, so installs match upstream
Weaknesses
Unofficial fork; content depends on periodic syncs from upstream
Translations are machine-produced by Codex, with human review only on PRs
Needs a one-time /setup-matt-pocock-skills run per repo for tracker and labels
No native Codex plugin; plugin install is read-only, so customizing requires skills.sh copy
no GPU
Needs Node.js (npx skills), Claude Code, Codex or another Agent Skills harness, GitHub or Linear issue tracker (optional, local files also supported)
Marketplace of coding-agent plugins, subagents, skills and commands for seven tools
41.6k starsMITlast commit Oct 2026
A Markdown-sourced catalog of 94 plugins (202 agents, 185 skills, 105 commands) covering Python, JavaScript, code review, testing, infrastructure and security. It installs natively into Claude Code, Codex CLI and Cursor, and generates artifacts for OpenCode, Antigravity CLI, GitHub Copilot and Pi via make targets. Individual skills can be installed with gh skill or npx skills.
Strengths
One Markdown source adapted to seven harnesses, with native marketplaces for Claude Code, Codex and Cursor
Large catalog: 94 plugins, 202 agents, 185 skills, 105 commands
Skills can be installed singly via gh skill or npx skills without cloning
Repo includes validation targets and a static skill scorer (plugin-eval) that needs no model calls
Weaknesses
Feature support differs per harness; Pi agents need a separate subagent extension
OpenCode, Antigravity, Copilot and Pi installs require cloning, uv and Python 3.12+
plugin-eval LLM judge and Monte Carlo layers are experimental, not validated against human labels
Pensyve memory entry is external, and its hosted cloud closed on October 1, 2026
no GPU
Needs Claude Code, Codex CLI, Cursor, OpenCode, Antigravity CLI, GitHub Copilot or Pi, uv and Python 3.12+ (clone installs), GitHub CLI 2.90+ (gh skill), Node.js and npm (npx skills)
Claude Code plugin with 67 skills and 9 Jira-linked workflow commands
12k starsMITlast commit Oct 2026
A Claude Code plugin that installs 67 skills across 12 categories, covering languages, backend and frontend frameworks, infrastructure, testing, DevOps and security. Skills activate from the request text and load reference files (371 in total). It also adds 9 workflow commands for epics, which need an Atlassian MCP server, and a /common-ground command that surfaces Claude's assumptions about the project.
Strengths
Installs with two /plugin commands from the Claude Code marketplace
67 skills with 371 reference files loaded on demand per request
Skills chain into flows, e.g. Debugging Wizard to Test Master to Code Reviewer
MIT licensed, with a docs site, changelog and contribution guide
Weaknesses
Built for Claude Code only; other agents are not mentioned in the README
Workflow commands require a separately configured Atlassian MCP server
README gives no measure of skill activation accuracy or output quality
Full skill list lives in a separate guide, not the README
no GPU
Needs Claude Code, Atlassian MCP server (workflow commands only)
Curated skill registry and CLI installer for AI coding agents
7.1k starscustom licenselast commit Oct 2026
A catalog of SKILL.md-based skills (spec-driven planning, AWS advice, Playwright automation, Figma to code, security review) with an npx CLI that installs them into agents such as Claude Code, Cursor, Copilot, Windsurf and Cline. Skills are fetched from a CDN, cached in ~/.cache/agent-skills, and tracked with a lockfile and audit log. A companion MCP server exposes search and read tools for the catalog.
Strengths
Installs to about 20 agents from one interactive or scripted CLI, copy or symlink
Skills scanned with Snyk Agent Scan; lockfile, content hashing and audit log
Optional MCP server with search_skills and read_skill for progressive loading
Global or project-local install, with offline cache of downloaded skills
Weaknesses
Mixed licensing: MIT code, CC-BY-4.0 skills, third-party skills keep own licenses
Attribution to Tech Leads Club is required when using the skills
Skills are downloaded from a CDN, so first install needs network access
Local control plane that keeps goals and evidence for long-running coding agents
6.4k starsApache-2.0last commit Oct 2026
LoopX installs as a skill for Codex, Claude Code and DeepSeek Harness. It stores goals, Todos, decisions and evidence outside the chat session, so an agent can continue across runs through a heartbeat or loop driver and ask the owner when a gate is hit. It includes a local browser/PWA dashboard (`loopx dashboard`) and a CLI (`loopx status`, `loopx doctor`).
Strengths
Works with Codex, Claude Code and DeepSeek Harness without replacing their runtime
Goal state, decisions and evidence persist across sessions and restarts
Local-first dashboard shows what is running, waiting on you, or scheduled
Publishes LHTB results: 0.4948 mean reward vs 0.4218 for plain Codex on 46 tasks
Weaknesses
LHTB uses one trial per task and unequal budgets; not an isolated causal estimate
Strict pass rate (7/46) matches plain Codex; the gain is mostly partial credit
Needs Python 3.11+ and Node.js 22.22.3+ installed
Desktop App is ad-hoc signed, not notarized; Windows preview updates are manual
Needs Python 3.11+, Node.js 22.22.3+, Codex, Claude Code or DeepSeek Harness
Models: Codex, Claude Code, DeepSeek Harness, direct-model sessions
Claude Code plugin for multi-model review, debate and research workflows
4.3k starsMITlast commit Oct 2026
Claude Octopus is a plugin for Claude Code that adds /octo:* commands, skills and personas, which dispatch work to up to twelve external providers such as Codex, Copilot, Ollama, Perplexity and OpenRouter. A 75% consensus gate flags disagreements between models. Workflows include Double Diamond phases, a multi-LLM council and parallel git worktrees. It stays dormant until you run a command explicitly.
Strengths
Works with Claude alone; external providers are added one at a time when detected
Dormant until a /octo:* command is run, so ordinary prompts are not routed
Consensus gate and council modes compare several models' opinions before shipping
Also installs for Codex CLI, and for Cursor as an MCP server
Weaknesses
Requires Claude Code v2.1.14+; Windows needs WSL or SSH, Git Bash unsupported
Some providers need paid subscriptions or API keys; Qwen free OAuth tier ended
Large surface (54 commands, 63 skills, 31 personas) takes time to learn
Optional auto-router in invoke mode can send prompt context to external providers
Local router that adds Kimi, DeepSeek, Claude and other models to Codex
4k starsMITlast commit Oct 2026
Codex Router runs a per-user background service that adds external providers (Anthropic, Kimi, DeepSeek, xAI, Z.ai, Qwen plan, Ollama Cloud, OpenRouter and others) to the Codex App and CLI model picker. Credentials are entered in hidden local prompts or reused from official CLI OAuth sessions. It also ships an Electron Control Center, a tray app and a macOS desktop panel, and can serve Gemini CLI, Claude Code, Cursor and OpenClaw.
Strengths
Long model table across many providers, with separate global and China endpoints for Kimi and StepFun
Catalog is credential-aware: only models with a stored key or valid OAuth session appear
Credentials stay local; prompts are hidden and other CLIs' OAuth tokens are not copied
Installers for macOS, Linux and Windows, a Homebrew tap, and a credential-free idle install
Weaknesses
Independent project, not affiliated with OpenAI or the model vendors
No prebuilt macOS .dmg; it builds locally and needs the full Xcode app
Subscription agent bridges are experimental and reject permission requests by default
Homebrew install is CLI-only, has no Control Center, and builds Python deps from source
no GPU
Needs Codex App or CLI, Node.js 22.19+, uv or Python 3.10+, Git, Xcode (macOS Control Center build)
Models: Anthropic, Kimi, DeepSeek, xAI Grok, Z.ai GLM
Personal context, memory and skills layer for AI coding harnesses
19.6k starsMITlast commit Sep 2026
LifeOS installs as a single skill bundle on top of an AI coding harness, adding persistent memory, personal context files (goals, contacts, preferences), skill routing and hooks. It is written in TypeScript and Bash, runs on bun, and is most tested on Claude Code. Install is done by prompting your AI to read the install page, or with a curl shell script on macOS/Linux.
Strengths
Installer merges settings and leaves your USER/ customizations untouched on upgrade
Bundles a skill library covering research, security, writing and art
Built on hooks, skills and context files, not one vendor's features
Docs site and walkthrough video available
Weaknesses
Needs a capable AI coding harness; Claude Code is the only well-tested path
Local model support (Ollama, llama.cpp) is still on the roadmap
Install is delegated to an AI agent or a curl-piped shell script
Public repo is generated from a private tree; PRs are ported, not merged
no GPU
Needs bun, Claude Code or another AI coding harness
Agent skill that renders diagrams, diff reviews and tables as styled HTML pages
10.6k starsMITlast commit Oct 2026
An agent skill that makes a coding agent write a self-contained HTML page and open it in the browser, instead of printing ASCII diagrams and pipe tables. It ships command templates such as /diff-review and /plan-review, plus a local stdio MCP server, a native Pi tool, and optional PPTX and MP4 export. It installs as a Claude Code plugin or by copying the skill directory for Codex, Cursor, OpenCode and Antigravity.
Strengths
Normal skill use needs only a browser: no build step, no runtime dependencies
Works across many harnesses: Claude Code, Codex, Cursor, OpenCode, Pi, Antigravity
MCP server is local stdio only, no LLM calls, writes confined to an output directory
Bundled commands for diff review, plan review, slides, fact-check and project recap
Weaknesses
Non-Claude harnesses need manual copy installs; prompt support varies by build
Video export needs ffmpeg, Chrome and playwright-core installed separately
PPTX export is a static utility for simple HTML slide decks only
Output quality depends on the agent model following the skill instructions
no GPU
Needs Web browser, Node.js (optional MCP and PPTX utilities), ffmpeg, Chrome, playwright-core (optional video export)
Agent skill that generates and syncs editable draw.io diagrams from prompts and code
10.3k starsMITlast commit Oct 2026
An Agent Skills package for Claude Code, Cursor, Copilot, Codex and similar agents. It writes .drawio XML from natural language, Mermaid, images, Terraform, Kubernetes, SQL DDL, OpenAPI and code import graphs, then exports through the draw.io desktop CLI and checks the rendered PNG. A diagramctl CLI adds sync that keeps manual layout, architecture rule checks for CI, and an optional MCP server.
Strengths
16 importers, including Terraform, Kubernetes, SQL DDL, OpenAPI and Python/JS/Go/Rust import graphs
Sync updates changed nodes while preserving manually tuned coordinates and styles
Resolves 10,000+ official draw.io shapes and 321 AI/LLM logos instead of guessed styles
Core workflows are stdlib-only and offline; architecture rules run in CI via a GitHub Action
Weaknesses
Requires the draw.io desktop CLI; version 30 or later needed for Mermaid and ELK layout
Auto-layout needs Graphviz; headless Linux export needs xvfb
AI logos load from the unpkg CDN by default, so rendering needs network unless embedded
Many scripts and flags make it harder to learn than a plain diagram prompt
Approval-gated delivery workflows for AI coding assistants
5.5k starsMIT-0last commit Oct 2026
AI-DLC installs a harness-neutral workflow engine into Claude Code, Kiro CLI and IDE, Codex CLI, Cursor, opencode and GitHub Copilot. A single /aidlc command picks one of 11 workflow profiles, runs 33 stages across 5 phases with 14 agents, and stops at human approval gates. State, a 119-event audit trail and learned rules persist per project.
Strengths
Same engine across seven harnesses, configured with `aidlc config --harness`
Human approval gates and a 119-event audit trail
11 workflow profiles, including bug fix, infrastructure, security and proof of concept
Native installer; Bun and Node.js not needed for normal install
Weaknesses
Needs a separate harness and model provider; AI-DLC supplies neither
Install is a curl-to-shell script; manual path requires Bun
Large surface (33 stages, 14 agents, 93 tools) means a learning curve
Recommended model is Claude Opus 4.8; quality on other models unknown
no GPU
Needs Claude Code, Kiro, Codex CLI, Cursor, opencode or GitHub Copilot, Bun (manual-copy install only), Amazon Bedrock (optional)
Models: Claude Opus 4.8 (recommended), Amazon Bedrock, Provider set by harness
Plugin layer for Hermes Agent adding model routing, parallel coding lanes and memory
3.6k starsMITlast commit Oct 2026
oh-my-hermes (OMH) installs as a plugin and CLI over Nous Research's Hermes Agent. It scores each request and routes it to one of twelve editable model-and-effort categories, splits work into parallel git worktree units, and loads 108 specialist omh-* skills. It also adds a reviewed, file-backed long-term memory and its own TUI (bare `omh`).
Strengths
Twelve editable model-and-effort chains with fallback when a provider rejects a model
Parallel units get separate worktrees from one pinned SHA, with typed evidence states
Memory entries need review and carry provenance and review-due dates
Installs via curl script, Homebrew, Bun or npm, with `omh doctor` and `omh update`
Weaknesses
Requires an existing Hermes Agent install; not usable standalone
OMH TUI unavailable on Windows; falls back to Hermes' own TUI
Cost and speed figures come from the README's own GPT-6 Astra benchmark
Default install pipes a remote script to sh; setup needs an interactive model interview
Skill that makes coding assistants answer action-first with numbered steps
58.5k starsMITlast commit Oct 2026
A skill/plugin for coding assistants such as Claude Code that rewrites response style around 10 rules: lead with the next action, number multi-step tasks, cap lists at 5 items, give time estimates in minutes, and drop preambles, recaps and closers. Behavior is defined in skills/i-have-adhd/SKILL.md, and you can fork and edit that file, then reinstall your copy through the plugin marketplace commands.
Strengths
Behavior lives in one editable SKILL.md with 10 explicit rules
Before/after example shows the exact output change
Install is a single prompt pasted into the assistant
MIT licensed; README translated into 10 languages
Weaknesses
Only changes response style; adds no tools or capabilities
Tune-it steps shown only for Claude Code plugin commands
No tagged releases
No evidence offered that it improves outcomes beyond one example
Local gateway that routes coding agents to the model providers you choose
37.8k starsMITlast commit Oct 2026
Claude Code Router runs a local gateway on port 3456 that Claude Code, Codex, Grok CLI, Kimi CLI, Kilo Code, OpenCode and other clients point at. Providers, models, credentials and routing rules are managed in an Electron desktop app or an npm CLI with a browser UI. It adds retries, ordered fallbacks, credential pools, request logs with token and cost estimates, and tools such as web search, MCP and ToolHub.
Strengths
One local endpoint for ten named agents, so provider config lives in one place
Retries, key rotation, credential pools and ordered fallback models are built in
Logs show resolved provider, model, latency, tokens and estimated cost per request
Ships as desktop app (macOS, Windows, Linux), npm CLI or Docker Compose
Weaknesses
Docker script builds against a sibling ../../next-ai/gateway checkout when present
Windows desktop packaging must run on Windows x64 because of native better-sqlite3
Remote exposure needs care; README points to a separate Docker guide for it
Installable agents, commands, hooks, MCPs and skills for Claude Code
32.9k starsMITlast commit Oct 2026
A catalog of Claude Code add-ons (agents, slash commands, settings, hooks, MCP configs and skills) installed through an npx CLI or browsed on aitmpl.com. The CLI also ships a session analytics view, a mobile-oriented conversation monitor with optional Cloudflare Tunnel access, a health check and a plugin dashboard. Some components are bundled from other repositories under their original licenses.
Strengths
One npx command installs agents, commands, hooks, settings or MCPs into a project
Covers six component types, including skills and MCP integrations, in one catalog
Web catalog at aitmpl.com plus a separate documentation site
MIT license; bundled third-party components keep their original licenses
Weaknesses
Targets Claude Code only; no other coding agents are mentioned
Component quality likely varies because many are imported from other repos
README gives no versioning or update policy for installed templates
Requires Node.js via npx; minimum version unknown
no GPU
Needs Node.js (npx), Claude Code, Cloudflare Tunnel (optional, for remote chat monitor)
Agent skill that turns a video URL or file into timestamped frames and transcript
19.1k starsMITlast commit Sep 2026
Watch is a skill/plugin for Claude Code, Codex and other Agent Skills hosts. It converts a video URL or local file into timestamped frames and a transcript, using native captions first and optional WhisperX, Groq or OpenAI transcription otherwise. With a GEMINI_API_KEY it can instead send the video to Gemini and relay the timestamped answer.
Strengths
Installs as a plugin in Claude Code or as a skill in Codex and other hosts
Local engine works without an API key, using ffmpeg and yt-dlp
Frame detail levels, caps, start/end ranges and cue timestamps are configurable
Captions are used first; WhisperX, Groq or OpenAI only as fallback
Weaknesses
Does not work in Claude Chat or Cowork
Requires Python 3.10+, FFmpeg and current yt-dlp installed separately
YouTube also needs a JavaScript runtime such as Deno
WhisperX is verified only on Apple Silicon; other platforms untested
An agent skill that captures your real product pages, picks shots from a library of 124 recipe cards (150 shots), storyboards the film and renders it with Remotion to a 1920x1080, 30 fps MP4. Frames are drawn by code from your screenshots and copy, not generated by a video model. It also ships a browser editor (Motion Workbench) and optional JianYing draft export.
Strengths
Output is a Remotion React project, so every shot stays editable in code
Library of 124 recipe cards and 150 shots, with a searchable online Gallery
Includes 149 SFX, 5 BGM tracks, beat sync and frame-level QA steps
Motion Workbench for post-delivery editing; optional JianYing draft export
Weaknesses
Tested on macOS; headless Linux needs workarounds such as --concurrency=1
Remotion has its own license; companies may need a paid one
Bundled audio is under per-file licenses, not Apache-2.0
Spec-driven phase loop that runs AI coding agents in fresh-context subagents
10.8k starsMITlast commit Oct 2026
GSD Core installs commands and agents into coding runtimes such as Claude Code, Codex, Cursor, Copilot and Windsurf. It drives each milestone through a discuss, plan, execute, verify, ship loop, running research and execution in fresh-context subagents to limit context rot. State persists between sessions in files like STATE.md and CONTEXT.md.
Strengths
One installer targets many runtimes, including Claude Code, Codex, Cursor, Copilot, OpenCode and Kimi CLI
Persists project state in STATE.md and CONTEXT.md across sessions
Verify step walks through built work and generates fix plans before a phase closes
Docs split into tutorials, how-to guides, reference and explanation
Weaknesses
Must be installed through its installer; copying agents/ or commands/ files directly is unsupported
Fixed five-step phase loop imposes a workflow on every project
Default install path uses npx, so it needs Node.js; other runtimes need a separate guide
README gives no benchmarks for its context-rot claims
no GPU
Needs Node.js, npm
Models: Claude Code, Codex, OpenCode, Antigravity CLI, Kimi CLI
Chinese translation of the superpowers skills set for AI coding tools
8.4k starsMITlast commit Oct 2026
Installs 21 skills (brainstorming, TDD, systematic debugging, code review, parallel agents) into 26 AI coding tools, including Claude Code, Cursor, Codex CLI, Gemini CLI and Windsurf. 15 are translations of upstream superpowers, 4 are manually invoked China-specific skills, and 2 are kept from upstream history. Installed with `npx superpowers-zh`, which detects the tool in use.
Strengths
One `npx superpowers-zh` command detects the tool and installs skills, project or global scope
Covers 26 tools versus 6 in the English upstream
Also ships as a Claude Code plugin marketplace entry with hooks and bootstrap files
Adds Gitee, Coding, GitLab CN and CNB workflow skills; upstream sections are translated, not altered
Weaknesses
Skills and docs are in Chinese; English readers should use the upstream project
The four China-specific skills load only when invoked manually, e.g. `/chinese-code-review`
Manual `cp -r skills` install skips hooks and bootstrap, so skills do not auto-trigger
README is crowded with sponsor banners and promo links
no GPU
Needs Node.js (npx), An AI coding tool such as Claude Code or Cursor
Obsidian vault template giving coding agents persistent memory across sessions
5.4k starsMITlast commit Oct 2026
An Obsidian vault with hooks, slash commands and an AGENTS.md manual that let Claude Code, Codex CLI or Gemini CLI keep notes, links, decisions and brag-doc entries between sessions. Lifecycle hooks inject goals, active projects and recent changes at session start and validate markdown writes. Optional QMD adds local semantic search through an MCP server; without it the agent falls back to grep and the Obsidian CLI.
Strengths
Same hooks, commands and vault work with Claude Code, Codex CLI and Gemini CLI
Memory is git-tracked markdown in brain/, browsable in Obsidian and portable across machines
QMD search runs three local models, so no API key or per-query cost
Context injection has a byte budget and a size meter reported every session
Weaknesses
Claude Code has full support; Codex and Gemini have known gaps, and Codex needs hook approval
Requires Obsidian 1.12+ and Node 22+; hooks use the experimental strip-types flag
QMD model downloads total about 2.2GB; queries with embedding take a couple of seconds
Claude Code mod is unsandboxed code that runs with your permissions once trusted
Tutorial modules and copy-paste templates for Claude Code features
44.1k starsMITlast commit Oct 2026
A tutorial repository with 10 modules covering Claude Code slash commands, memory, skills, subagents, MCP, hooks, plugins, checkpoints, advanced features and the CLI. Each module has Mermaid diagrams and files you copy into .claude/ or ~/.claude/. It also includes a self-assessment quiz, a learning roadmap estimated at 11-13 hours, and an EPUB build script.
Strengths
Covers slash commands, memory, skills, subagents, MCP, hooks and plugins in one ordered path
Templates are plain Markdown, shell and JSON files copied into .claude/ directories
Includes /self-assessment and /lesson-quiz commands to find knowledge gaps
MIT licensed; README states it is synced with Claude Code releases (v2.1.302)
Weaknesses
Documentation and templates only; not a runnable tool or library
Specific to Claude Code; templates do not target other coding agents
Ships no plugin marketplace, so plugins load only via --plugin-dir from a clone
Needs updating as Claude Code changes hook events, flags and install methods
no GPU
Needs Claude Code 2.1+, uv (only for the EPUB build script)
Official Android development skills for AI coding agents, in SKILL.md format
9.5k starsApache-2.0last commit Oct 2026
A set of SKILL.md files from the Android team that give coding agents task-specific instructions for Android development, such as R8 analysis and Wear Compose M3. They target areas where evaluations show LLMs underperform, not basics like Jetpack Compose. Installed with the Android CLI (`android skills add`) or as a Claude Code or Codex plugin.
Strengths
Maintained by Android, grounded in developer.android.com guidance
Follows the open agent skills standard (SKILL.md), so it works across agents
Installs via Android CLI, or as a Claude Code or Codex plugin marketplace entry
Targets workflows where evaluations show LLMs underperform
Weaknesses
Public contributions are not accepted; changes go through GitHub issues only
Does not cover well-established areas such as basic Jetpack Compose practices
README does not list the full set of skills available
README disclaims that AI output can be wrong and needs review
Go installer adding memory, workflow and review rules to 17 coding agents
8k starsMITlast commit Oct 2026
Gentle-AI is a Go CLI that configures coding agents you already have (Claude Code, Codex, OpenCode, Cursor, Pi and 12 more) with persistent memory (Engram), skills, optional Context7 MCP and a credential-file deny-list. It adds two workflows: ODD, which sizes work and decides when to delegate to subagents, and RDD, where the binary freezes a change and sets review depth by risk. Configs are snapshotted before each write.
Strengths
Configures 17 agents using each agent's native configuration
Config backups before every write, plus a read-only `gentle-ai doctor` check
Review transitions come from the binary and on-disk state, not model judgment
Security deny-list blocks ~/.ssh, .env and credential files
Weaknesses
Does not install an agent; requires one already set up
Delegation and RDD review features differ per agent
Benchmarks are single runs (n=1) on the authors' own suite
Windows install is a Go source build needing Go 1.25.10+
no GPU
Needs An existing supported AI coding agent, Go 1.25.10+ (Windows source install), Engram, Context7 MCP (optional)
Run a local process against a live Kubernetes pod's traffic, files and env
5.4k starsMITlast commit Oct 2026
mirrord runs a process on your machine while routing its network traffic, file access and environment variables through a target pod in a Kubernetes cluster. It launches a helper pod on the target's node and mirrors incoming traffic to the local process. It ships as a CLI, a VS Code extension and an IntelliJ plugin, and the README says it can be used with agents such as Claude Code, Cursor and Codex.
Strengths
Test local code against real cluster services without a deploy
CLI, VS Code extension and IntelliJ plugin all available
Uses the default kubeconfig, no cluster-side install mentioned for basic use
Written in Rust, MIT licensed, with frequent releases
Weaknesses
Helper pod needs Linux capabilities such as CAP_SYS_ADMIN and CAP_NET_ADMIN
Database branching, queue splitting and preview environments need the paid Teams tier
Requires access to a Kubernetes cluster; not usable standalone
Not an AI tool itself; agent support is via external skills repo
Lint and format CLI that configures Biome, ESLint or Oxlint plus AI agent rules
3.4k starsMITlast commit Oct 2026
Ultracite is a CLI that sets up a JavaScript/TypeScript linter and formatter (Biome, ESLint with Prettier and Stylelint, or Oxlint with Oxfmt) with preconfigured rules. `ultracite init` also writes rule and context files for AI agents such as Claude Code, Copilot, Cursor and Gemini, and `fix` can hand unfixable issues to the Claude Code or Codex CLI.
Strengths
One `npx ultracite init` configures linter, formatter, editor settings and agent rule files
Supports three toolchains: Biome, ESLint/Prettier/Stylelint, and Oxlint/Oxfmt
Non-interactive init flags for CI, plus per-workspace framework settings for monorepos
`doctor` command diagnoses setup problems, including linter version mismatches
Weaknesses
Targets JavaScript and TypeScript only
Pins linter versions it supports; `upgrade` reinstalls the toolchain at those versions
Agent-assisted fixing requires the separate Claude Code or Codex CLI
README gives no rule counts or benchmark figures behind the speed claims
no GPU
Needs Node.js, Biome, ESLint or Oxlint, Claude Code or Codex CLI (optional)
Models: Claude Code, Codex, GitHub Copilot, Cursor, Gemini
Claude Code template with 49 game-dev subagents, 74 skills and validation hooks
27.3k starsMITlast commit Oct 2026
A project template that configures Claude Code with 49 subagents arranged as directors, department leads and specialists, plus 74 slash-command skills, 15 hooks, 14 path-scoped rules and 62 document templates. It guides a game project from brainstorm through design docs, stories, QA and release, with Godot, Unity and Unreal specialist agents. A `modes.rigor` setting in project.yaml (minimal, standard, full) controls how much process is enforced.
Strengths
Engine-specific agent sets for Godot 4, Unity and Unreal Engine 5
Hooks validate commits, pushes, assets and design-doc reads in Bash
modes.rigor presets default to minimal; README reports the standard tier did not improve the game
Stories touching visuals require a launched game and a retained screenshot
Weaknesses
Requires Claude Code; no other coding agent is supported
Needs Python 3 and Bash; Windows users need Git for Windows
README admits standard rigor wrote 30 docs before code and ranked last
No sample game shipped, so output quality is not demonstrated in the repo
no GPU
Needs Claude Code, Git, Python 3, Bash, jq (recommended)
A skill for Claude Code, Codex, OpenCode or Hermes that takes one reference image and produces a TypeScript THREE.Group factory built from primitives, procedural shaders and generated geometry, with pivots, sockets and colliders for animation. Python stdlib scripts run staged build passes (blockout through optimization) with validation and review gates, while the agent does the visual judgment and code. Optional plugins add CS2 weapon skins, character rigging and GLB output.
Strengths
Output is code-only Three.js with pivots, sockets and colliders, not mesh files
Quality gates block generation until a detail inventory is complete
Installs into Claude Code, Codex, OpenCode or Hermes via one npx command
Public live gallery shows generated source for each model
Weaknesses
Needs an external coding agent; installing the skill alone produces nothing
A single image cannot guarantee likeness; unseen regions are inferred
The environment plugin is private and needs authorized GitHub access
Hermes setup needs a manual symlink step
Needs Node.js >= 18, Python >= 3.10, Git, Claude Code, Codex, OpenCode or Hermes, Three.js
/brag is an agent skill for Claude Code, Codex, opencode and other Agent Skills hosts. It reads your project, plans the story and tone, then hands a brief to Hyperframes to render brag.mp4 with music, motion and share copy. A lighter /brag-slim variant skips Hyperframes and is used automatically on Opus 5.5.
Strengths
Installs through Claude Code and Codex plugin marketplaces or the skills CLI
Outputs plan, composition brief, share copy and rendered brag.mp4 in brag-output/
Tone is steerable with --tone; voiceover is optional via --voice using Kokoro
Bundled music and SFX; /brag-slim needs no Hyperframes or bundled assets
Weaknesses
Full workflow requires Node.js 22+, FFmpeg and the Hyperframes CLI
Voiceover is off by default and depends on Hyperframes with Kokoro
Bundled music is CC BY 4.0 and needs attribution
Windows needs symlink support or a manual copy of the skill directory
Needs Node.js 22+, FFmpeg, Hyperframes CLI, An Agent Skills-compatible agent
Models: Claude Code, Codex CLI, opencode, Cursor, Gemini CLI
Skill-based research workflow where one model executes and another reviews
18k starsMITlast commit Oct 2026
ARIS is a set of 82 skills that run an ML research workflow (idea discovery, experiments, paper writing) inside Claude Code, Codex CLI, Cursor and similar agents. An executor model drives the work while an independent reviewer model, Codex via an MCP bridge by default, critiques each step. It also ships a standalone Rust CLI (ARIS-Code) and plugins for Claude Code and Codex CLI.
Strengths
Works across Claude Code, Codex CLI, Cursor, Trae, Copilot CLI and others
Separate reviewer model checks the executor's output; alternative model pairs documented
Also ships as standalone CLI and as Claude Code and Codex plugins
MIT license with frequent releases; latest ARIS-Code is v0.4.28
Weaknesses
Default reviewer depends on codex-cli, which dropped `codex mcp-server` in 0.154
Focused on ML research and paper writing, not general coding
README is long and dense, with a heavy changelog and many sibling projects
Hardware needs and running costs are unknown; default setup uses paid model APIs
Needs Claude Code or Codex CLI, codex-cli, Codex MCP bridge
Models: Claude Opus 5, Claude Fable 5.1, GPT-6-Astra, ModelScope-hosted models
Skills, CLI and field-note knowledge base for AI agents modding PC games
8.8k starsMITlast commit Oct 2026
Packages Agent Skills, subagent role files, a Python `um` CLI and the fal MCP server so a coding agent can scan an installed game, pick a modding route, build a mod, generate assets with fal, test and record it. Installs as a plugin for Claude Code, Codex, Gemini CLI, Copilot, Cursor and OpenCode. A `knowledge/` folder holds per-game field notes that agents contribute by pull request.
Strengths
Installs as a plugin or skills-only for six agents, including Claude Code, Codex and Cursor
`um scan` fingerprints engine, anti-cheat, loaders and save folders for Steam, Epic and Xbox installs
Engine playbooks cover Unity, Unreal, Godot, Source, Minecraft, .NET/XNA and more
Safety rules: save backups, PID-only kills, `um publish check` blocks shipping game files
Weaknesses
Asset generation depends on a fal API key; the alternative is a local ComfyUI server for images only
Recording and input driving (`um win`) are Windows-specific tools
The mod hub is listed as coming next and is not yet available
Three-agent orchestrator team uses several times the tokens of a single agent
GPU optional
Needs Git, Python 3.10+, ffmpeg, fal API key, uv (recommended), Blender (3D to sprite renders), ComfyUI (optional)
Models: Claude Code, Codex, Gemini CLI, GitHub Copilot, Cursor
Rulebook that stops coding agents shipping generic UI, filler copy and noisy comments
7k starsMITlast commit Oct 2026
Anti Slop is a set of agent skills built around 38 mandatory rules in three tiers, plus a PASS/FAIL Delivery Gate report run before work ships. Six skills cover the core filter, UI, copywriting, accessibility, responsive layout and code comments. It filters output and does not prescribe colors, fonts or layouts; direction is expected from a user-written DESIGN.md.
Strengths
Six separate skills, so an agent loads only the concerns a task needs
Installs via npx installer, skills.sh, or plugin commands for about ten agents
Two modes: guide while building, or audit existing work with a findings list
Single antislop.md file works as a paste-in filter for any chat window
Weaknesses
Filter only: output stays plain without a DESIGN.md supplying direction
Does not update itself; each install route has its own update command
Effect is shown by screenshot comparisons, with no measured results in the README
Skills load at session start, so changes need a new session
no GPU
Needs Node.js (npx installer), An AI coding agent such as Claude Code, Codex, Cursor or Cline
Local MCP server and shell hook that trim context fed to coding agents
3.9k starsApache-2.0last commit Oct 2026
LeanCTX sits between AI coding tools and the files and command output they read. It exposes ctx_* MCP tools with 16 file read modes, compresses shell output with 85+ patterns, and keeps local session memory, with an optional proxy for model requests. It is written in Rust and set up per agent with `lean-ctx wrap cursor|claude|codex`.
CLI that syncs skills, agents, rules, MCP and hooks across AI coding tools
2.8k starsMITlast commit Oct 2026
skillshare keeps skills (SKILL.md), agents, rules, MCP connections, hooks and plugins in one source directory and syncs them to Claude Code, Codex, Pi, OpenCode, OpenClaw and other tools, using symlinks or copies. It installs from any Git host, audits skills before use, and supports per-project setups committed with the code. It ships as a single Go binary with a web dashboard (`skillshare ui`) and a separate desktop app.
Strengths
Single Go binary with no Node.js or npm runtime dependency
One source syncs skills, agents, extras, MCP and hooks to many tools
Built-in `audit` command, with auto-scan on install and update
Project mode keeps skills in `.skillshare/` committed with the repo
Weaknesses
Windows file symlinks need Developer Mode, otherwise files are copied
Plugins sync separately from `sync --all`
Desktop app is a separate repo; macOS build is Apple Silicon only
Supported tool list is only partly named in the README (`and more`)
Installed with `npx skills add`, this skill lets agents such as Claude Code, Cursor or Codex produce single-file HTML interactive prototypes, HTML decks exportable to editable PPTX, MP4/GIF animations and infographics. It includes a library of 84 HTML/CSS style specimens, a brand-asset protocol, and local scripts for linting, export and reference screenshots. The core pipeline runs locally; cloud features (TTS, AI video review) are optional.
Strengths
Exports HTML decks to PPTX with real editable text boxes via html2pptx.js
Ships 84 style specimens with source, so styles are viewed before being chosen
Core design and export pipeline runs locally with no network or API key
Works across skills-compatible agents; MIT license permits commercial use
Weaknesses
No layer-level editable export to Figma; output is HTML, not Keynote-editable
Complex 3D, physics or particle animation is outside its scope
Older skills CLI (≤1.5.15) copies only SKILL.md, breaking the install
README is Chinese-first; the English version is a separate file
Codex skill that uses ChatGPT web as planner and reviewer via read-only MCP
7.5k starsMITlast commit Oct 2026
A Codex Skill plus a local Node.js bridge that lets the ChatGPT web app plan and review work while Codex executes edits, tests and git commands. ChatGPT reads the workspace through a loopback, OAuth 2.1-protected, read-only MCP server with 10 tools, exposed via a Cloudflare tunnel. This uses a ChatGPT subscription's web quota instead of API tokens.
Strengths
Read-only MCP tools only; no write, delete or shell tools exist on the server
React toolbar for annotating page elements to give coding agents feedback
5k starscustom licenselast commit Sep 2026
Agentation is a React component that adds a toolbar to your app. You click elements, select text or drag areas, add notes, and copy markdown with selectors, positions and context so a coding agent can grep for the code. An optional MCP server (npx -y agentation-mcp server) syncs annotations to the agent over HTTP.
Strengths
Outputs selectors and positions an agent can grep, instead of prose descriptions
Works with any agent via copy-paste; MCP endpoint is optional
No runtime dependencies beyond React peer dependencies
Selects inside open shadow roots and same-origin iframes
Weaknesses
Requires React 18+; no other frameworks listed
Desktop browsers only; mobile not supported
PolyForm Shield 1.0.0 license is source-available, not OSI open source
Direct sync needs a separate local MCP server process
CLI and MCP server that sends a prompt plus selected files to a second model
4.1k starsMITlast commit Oct 2026
Oracle bundles a prompt with files you select (globs, directories, exclusions) and sends it to a model through a provider API or a signed-in ChatGPT or Gemini browser session. Results are stored as sessions under ~/.oracle/sessions and can be reattached, replayed or followed up. It also runs as an oracle-mcp stdio server and ships a skill so coding agents can ask for a second-model review.
Strengths
Render mode prints the exact prompt and files without credentials or a model call
API mode supports OpenAI, Azure OpenAI, Anthropic, Gemini, xAI, OpenRouter and compatible endpoints
--models runs several models in one session with per-model usage, cost and partial failures
Dry run reports resolved files and token estimate; bundles carry line numbers for path:line citations
Weaknesses
Requires Node.js 24 or newer
Browser mode depends on Chrome automation and a one-time login flow
Follow-ups work only for supported API or ChatGPT conversations
CLI that patches Claude Code to customize prompts, themes, and UI
2.6k starsMITlast commit Oct 2026
tweakcc patches Claude Code's minified cli.js, or the native Bun binary via node-lief, using settings from ~/.tweakcc/config.json. It edits system prompts, themes, thinking verbs and spinners, toolsets and subagent models, and adds patches such as AGENTS.md support and MCP startup optimization. Customizations are reapplied with `npx tweakcc --apply` after Claude Code updates, and custom patches can be written through an API or `adhoc-patch`.
Strengths
Supports native binaries and npm, pnpm, bun, Homebrew and Nix installs on Windows, macOS, Linux
Config in ~/.tweakcc/config.json is reapplied with --apply; a SessionStart hook can automate it
Custom patches via npm API or sandboxed `adhoc-patch` scripts, including from an HTTP URL
Restores from backup before each apply; `--restore` reverts Claude Code
Weaknesses
Verified only against Claude Code 2.1.162; other versions may break non-prompt patches
Claude Code updates overwrite patches, and native updates can replace the patched binary
Patches depend on the internal layout of a third-party minified binary
README says documentation for some features is still missing
Markdown agent persona files for Claude Code, Cursor, Codex and other coding tools
163.6k starsMITlast commit Oct 2026
A collection of role-based agent definition files (engineering, design and other divisions), each describing a specialist's identity, workflow, code deliverables and success metrics. Shell scripts (convert.sh, install.sh) generate and install them for Claude Code, Cursor, Gemini CLI, OpenCode, Copilot, Aider, Windsurf, Codex and others, with per-division or per-agent selection. A separate desktop app handles browsing and install.
Strengths
Installer supports about 14 tools, with --division and --agent filters and a --dry-run mode
Plain markdown files can be copied or adapted individually without running any scripts
Large roster: roughly 60 engineering agents alone, plus design and other divisions
MIT license; runbook roster JSON lists agent teams that install directly by slug
Weaknesses
Prompt files only; no runtime, so output quality depends on the host tool and model
OpenCode registers only about 119 agents, so a full install loses the rest
No tagged releases in this repo; the desktop app is a separate project
Persona-style prompts with no benchmarks or tests shown for the claimed outcomes
no GPU
Needs bash, python3 (optional, for runbook roster extraction)
Models: Claude Code, Cursor, Codex, Gemini CLI, OpenCode
Reference repo of Claude Code agents, commands, skills, hooks and settings
68k starsMITlast commit Oct 2026
A documentation repository that explains Claude Code features (subagents, commands, skills, hooks, MCP, settings, memory) with usage notes and working examples under .claude/. It includes a weather-orchestrator demo of a Command to Agent to Skill workflow and a table comparing community development workflows such as Superpowers, Spec Kit and gstack.
Strengths
Each concept links to a guidance note and a matching implementation file
Runnable weather-orchestrator example shows the Command, Agent, Skill pattern
Compares community workflows with agent, command and skill counts
MIT licensed and mostly Markdown, so files are easy to copy into a project
Weaknesses
Documentation and examples only; no installable tool or package
Tied to Claude Code, and beta features may change quickly
Many links point to X posts and external repos rather than local docs
Sponsor banners and trending badges sit at the top of the README
MCP server that keeps raw tool output out of the context window
32k starscustom licenselast commit Oct 2026
Context Mode is an MCP server that runs analysis in a sandbox so raw tool output stays out of the agent's context. It logs edits, git operations and tasks in SQLite, indexes them with FTS5, and retrieves relevant events via BM25 after compaction. It ships hooks and routing for Claude Code, Gemini CLI, Copilot, Cursor and others.
Strengths
Sandbox tools cut raw output; README cites 315 KB reduced to 5.4 KB
Session events stored in SQLite with FTS5/BM25 retrieval after compaction
Hooks for Claude Code, Gemini CLI, VS Code and JetBrains Copilot, Copilot CLI, Cursor
Claude Code plugin install, plus MCP-only option via npx
Weaknesses
Elastic License 2.0, not an OSI open-source license
Non-hook platforms need a manual routing file copy
Most platforms require Node.js >= 22.5 or Bun
Savings figures are README examples, not independent benchmarks
no GPU
Needs Node.js >= 22.5 or Bun
Models: Claude Code, Gemini CLI, VS Code Copilot, JetBrains Copilot, GitHub Copilot CLI
Collection of 161+ Claude Code subagent definitions in 10 categories
26k starsMITlast commit Oct 2026
Markdown subagent definitions for Claude Code, grouped into 10 categories such as core development, language specialists, infrastructure, quality and security, and data and AI. Install them as Claude Code plugins (for example voltagent-lang), by copying files into ~/.claude/agents/ or .claude/agents/, or with a shell installer script.
Strengths
Installable per category via claude plugin marketplace, so you add only the groups you need
Covers many languages and frameworks, including Django, Rails, Spring Boot, Flutter and Terraform
Interactive and standalone shell installers; standalone one needs only curl
MIT licensed; agents are plain Markdown files you can copy and edit
Weaknesses
Prompt definitions only; README shows no tests or evals of agent quality
Written for Claude Code; other agent tools would need the files adapted
Meta-orchestration agents are meant to be used with other categories installed too
No tagged releases; README carries sponsor placements
Agent skill that builds interactive isometric figures to explain technical ideas
19.8k starsMITlast commit Oct 2026
Anatomy is an Agent Skill for Claude Code, OpenCode, Codex and Antigravity. Given a concept, it invents a physical object whose behaviour models the idea and builds it as an interactive isometric figure in SVG, SVG plus WebGL, or beta 3D. Output is standalone HTML, a lone SVG, or React/Next.js components, with a zero-dependency Node kit and geometry audit scripts.
Strengths
Installs with one `npx skills add` command across four agents; no API keys needed
Kit uses only Node built-ins; React is needed only for the React components
Includes geometry audit and headless Chrome capture scripts to verify figures
Ships nine built examples, including React components copyable into a site
Weaknesses
Verification needs headless Chrome, which the Codex sandbox may block without config changes
3D mode is marked beta and only re-renders the moving parts
Raptor engine example build takes a few minutes; 3D audit is slow
Repository is about 10 MB because the CLI copies examples too
Agent skills for reviewing and building UI, typography, color and accessibility
8.2k starsMITlast commit Oct 2026
A set of 14 markdown agent skills for interface work: UI polish, typography, colors, WCAG 2.2 accessibility, layout and product copy. Some are user-invoked commands, such as reviewing a branch or PR for interface regressions, stress-testing a component in every state, or building UI from a Figma file or design image. Installed with `npx skills add` or as a Claude Code plugin.
Strengths
Separate skills for accessibility, layout, writing, typography, color and UI, plus one combined review
User-invoked skills for PR review, component state pages, variants and Figma-to-UI
Two install paths: npx skills CLI and a Claude Code plugin marketplace
MIT licensed; skills are plain SKILL.md files you can read and edit
Weaknesses
README does not list which agents besides Claude Code are supported
No tagged releases; versioning is unknown
Skills are prompts only, so results depend on the model running them
README does not document what the Figma skill needs, such as Figma access
Local proxy that renders bulky LLM request context as PNGs to cut input tokens
7.6k starsMITlast commit Oct 2026
pxpipe sits between Claude Code (or Codex, cursor-agent, and other clients) and the provider API. It rewrites the system prompt, tool docs and older history into dense PNG pages plus a bounded factsheet before the request leaves the machine. It handles Anthropic Messages, OpenAI Responses/Chat Completions and Google generateContent, and logs measured savings against a count_tokens counterfactual.
Strengths
Savings measured per request against a count_tokens probe, logged to ~/.pxpipe/events.jsonl
Reported ~59-70% lower end-to-end bill on Claude Code traffic; SWE-bench Lite 10/10 both arms
Profitability gate leaves sparse prose and recent turns as text
Also offers offline PNG export, a library API and a dashboard on 127.0.0.1:47821
Weaknesses
Lossy: exact 12-char hex recall 13/15 on Fable 5, 0/15 on Sol; misses are silent
No verbatim-risk guard yet; IDs and hashes must stay text
Default model scope is limited: Fable 5, Opus 5.5 and Gemini; others opt-in
SWE-bench Pro showed 14/19 vs 15/19 text; small samples, and the README says savings depend on workload
no GPU
Docker + Compose
Needs Node.js, Anthropic, OpenAI-compatible or Google API access
peon-ping registers hooks and adapters for Claude Code, Codex, Cursor, Copilot, Gemini CLI, OpenCode and about a dozen other agents. It plays sound-pack voice lines (Warcraft, StarCraft, Portal and others) and shows overlay banners and terminal tab titles on events such as task complete, permission needed, error and rate limit. It ships a `peon` CLI, an MCP server, SSH relay and phone notifications via ntfy.
Strengths
Adapters for 20+ agents and IDEs, plus any MCP client
Per-event sound categories following the open CESP pack spec
Per-directory and per-IDE pack bindings, path exclusions, and Focus/DND detection
Installs via Homebrew, shell script, PowerShell, or Nix with Home Manager module
Weaknesses
Windows lacks some CLI commands such as setup, rotation, preview and mobile
WSL2 on Windows 11 24H2+ needs ffmpeg to play MP3 packs
Written in Shell; installer is piped from curl by default
Overlay banners listed only for macOS, WSL and MSYS2, not plain Linux
no GPU
Needs ffmpeg (WSL2 on Windows 11 24H2+, MP3 packs), ntfy (optional, phone notifications)
Models: Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI
DeepSeek Harness plugin for inspecting agent context, token use and subagents
3.5k starsApache-2.0last commit Oct 2026
dsh-context is a plugin for DeepSeek Harness that adds a Context Insights page, Context and Fleet tabs, a right-sidebar panel and a /context command. It shows context composition, per-request trends, compaction events, file activity, token cost and subagent networks. It installs through the Add plugin wizard or the dsh CLI.
Strengths
Context browser shows each request's actual system prompt, tool schemas, messages and tool outputs
Trend chart pins compactions and prunes, with provider-reported actuals beside estimates
Fleet tab shows each subagent's context size, occupancy and cost live
Installs with one dsh command or the Add plugin wizard, no build step or restart
Weaknesses
Works only inside DeepSeek Harness, from @deepseek-ai/dsh 0.2.0-rc.2 onward
Per-category token figures are heuristic estimates, not exact counts
Cost is estimated from models.dev list prices, not actual billing
UI is available only in English and Simplified Chinese
Captures coding agent sessions and turns them into reusable memory
2.7k starsMITlast commit Oct 2026
Beacon records session history from Claude Code, Cursor, Codex, OpenCode and 20+ other agent harnesses, normalizes it into an OpenTelemetry-based event model, and writes it to local JSONL. Reviewed knowledge is exposed to later agents through MCP or Agent Skills. It includes a terminal trace browser, a local read-only dashboard, and forwarding to SIEM and storage destinations.
Strengths
Covers 20+ harnesses with one normalized trace format, including Claude Code, Cursor, Codex and Cline
Local-first: JSONL at ~/.beacon/endpoint/logs/runtime.jsonl, plus terminal and local web viewers
Forwards to Splunk, Datadog, Elastic, Sentinel, S3, GCS, CloudWatch and others
Installs via Homebrew, .deb/.rpm with checksums, or a Windows MSI
Weaknesses
Interactive setup preselects Beacon Cloud; local-only requires choosing the Local opt-out
Capture coverage varies by harness; some lack token, MCP or approval events
Custom lenses run HTML in a sandbox that the README says cannot block every data exfiltration path
Memory extraction and review workflow details are mostly in external docs, not the README
no GPU
Needs Homebrew, apt or dnf, or Windows MSI for install, Optional: Beacon Cloud account, Optional: Splunk, Datadog, Elastic, S3 or GCS for forwarding
CLI that stores specs, tasks and session memory in your repo for coding agents
15k starsAGPL-3.0last commit Sep 2026
Trellis is an npm CLI that writes a `.trellis/` directory into your repo holding specs, task PRDs and per-developer journals, then injects the relevant context into coding agent sessions. It runs a four-phase loop (plan, implement, verify, finish) using skills and sub-agents, and generates platform-specific files for 22 AI coding platforms.
Strengths
Specs, tasks and journals are plain files in the repo, reviewable like code
Same .trellis structure generated for 22 coding platforms, not just Claude Code
Verify step runs lint, type-check and tests against specs and self-fixes
Per-developer journals stay separate while shared specs live in git
Weaknesses
AGPL-3.0 license may be a problem for some company use
Requires both Node.js >= 18 and Python >= 3.9
Needs upfront effort to write or draft specs to get value
Installer that adds a reversible custom instruction set to local Codex
4.8k starsMITlast commit Sep 2026
A single-file Python script that installs an instruction prompt into a local Codex config directory (e.g. ~/.codex), so new Codex sessions follow it. It shows a dry-run plan first and writes only with --yes, and it offers --status, --restore-hooks, --uninstall and --reactivate. Per the README it does not modify Codex itself or read account credentials.
Strengths
Preview first; nothing is written until --yes is passed
Uninstall and restore commands, also preview-first
Release scripts ship with SHA256SUMS for checksum verification
Single Python file, needs only Python 3.10+
Weaknesses
Affects only new Codex sessions; old tasks must be closed
Windows fresh deployment is still a beta channel
Desktop builds are unsigned and trigger Gatekeeper or SmartScreen
Main README is in Chinese; English version is a separate file
Self-hosted registry and telemetry for internal coding agent skills, MCP servers and hooks
4.5k starsApache-2.0last commit Oct 2026
Observal is a server plus CLI that stores an organization's coding agent resources (agents, skills, MCP servers, hooks, prompts, sandboxes) in one reviewed registry. It generates per-harness config for Claude Code, Cursor, Kiro, Codex, Copilot, OpenCode and others, and captures session traces through installed telemetry hooks. It also publishes approved resources over ARD and lets agents delegate tasks over A2A.
Strengths
One install command renders an agent into each supported harness's config format
Admin review, version diffs and forking of approved releases keep registry changes auditable
Session traces and usage data feed reports on which agents and tools get used
Implements ARD v0.91 discovery and A2A v1.0 delegation with team-private visibility rules
Weaknesses
Needs a multi-service stack (API, web UI, Postgres, ClickHouse) on Docker Engine 24+ with Compose v2
Telemetry hooks are installed into each developer's harness via the CLI, adding a per-machine setup step
Value depends on an organization already having many internal agent resources to catalog
RAM and hardware requirements are not stated in the README
Docker + Compose
Needs Docker Engine 24+ with Compose v2, PostgreSQL, ClickHouse
Models: Claude Code, Cursor, Kiro, Codex, GitHub Copilot
Sidebar and bottom-panel workbench plugin for the DSH coding agent
4.5k starsMITlast commit Oct 2026
A plugin for the DSH agent host that adds a right sidebar (file tree, Git, tasks, side chat) and a self-drawn bottom panel with split panes and terminal tabs. It exposes a ctx.betterSidebar service with registerTab and registerFileViewer so other plugins can add pages and file viewers. Installed with `dsh plugin add`; requires DSH 0.2.1-rc.1 or newer for the current release.
Strengths
Editable CodeMirror file editor with Markdown, Mermaid and sandboxed HTML previews
Read-only Git commit graph plus Source Control groups and a session AI-changes view
Public registerTab and registerFileViewer API; built-in tabs use the same API
Lists 30+ third-party tab and viewer plugins on a GitHub topic
Weaknesses
Works only inside DSH; no standalone use
Current v0.28.0 is incompatible with DSH 0.2.0-rc.2, which needs v0.26.1 pinned
README states RC validation of v0.28.0 is still pending
Side Chat is beta; Git graph shows no tags and has no history-changing actions
Understand Anything is a plugin for Claude Code and about 15 other agent platforms. Its /understand command runs a multi-agent pipeline that combines Tree-sitter parsing with LLM summaries to write a knowledge graph to .ua/knowledge-graph.json. A local dashboard then offers graph navigation, search, guided tours, diff impact analysis and chat, and /understand-knowledge handles Karpathy-style LLM wikis.
Strengths
Tree-sitter extracts imports and call sites deterministically, so structural edges are reproducible
Graph is plain JSON that can be committed; teammates view it with Node.js >= 18 and no LLM
Incremental re-analysis of changed files, plus an optional post-commit auto-update hook
Installs on Claude Code, Cursor, Copilot, Codex, Gemini CLI, OpenCode, Kiro and others
Weaknesses
First /understand run on a large repo can use a significant number of tokens
Needs a host agent platform and an LLM; it is not a standalone tool
Output languages limited to en, zh, zh-TW, ja, ko, ru, vi
Graphs over 10 MB need git-lfs to commit comfortably
no GPU
Needs Claude Code or another supported agent platform, Node.js >= 18 (standalone viewer), pnpm (building from source)
Models: Whatever the host platform uses, Local models via Ollama
Chinese-language community fork of msitarzewski/agency-agents: 328 markdown agent personas across 22 departments, 67 of them original China-market roles (Xiaohongshu, Douyin, WeChat, Feishu, DingTalk and others). Shell scripts (install.sh, convert.sh) copy or convert the files for 20 coding tools, including Claude Code, Copilot, Cursor, Aider and OpenClaw. Each persona defines identity, rules, workflow and deliverables.
Strengths
328 persona files in 22 departments, each with defined workflow and deliverables
install.sh targets 20 tools; Claude Code and Copilot files can be copied directly
67 original China-market roles, such as Xiaohongshu, Feishu, DingTalk and WeChat mini programs
MIT license; plain markdown files that are easy to read and adapt
Weaknesses
Most tools other than Claude Code and Copilot need a convert.sh step first
Persona content is mostly Chinese; the English version is the upstream repo
README carries many sponsor ads and links to the author's other products
Persona prompts only; multi-agent orchestration needs the separate agency-orchestrator project
no GPU
Needs Claude Code, Cursor, Copilot or another supported AI coding tool
Models: Claude Code, GitHub Copilot, Cursor, Aider, Windsurf
Agent skill that drafts Chinese software copyright filing documents from your code
6k starsAGPL-3.0last commit Oct 2026
A skill directory for coding agents that analyzes a real project and produces materials for Chinese software copyright registration: an application-form TXT, an operation manual DOCX, and front 30 / back 30 page source code DOCX files. Code is extracted from the existing repository, not generated. DOCX output is built through a globally installed OfficeCLI 1.0.151, and the agent pauses for confirmation at each stage.
Strengths
Code pages are extracted from the real project; the README says the AI does not invent source
Applies the front 30 / back 30 page rule, and emits all code if under 60 pages
Runs locally and stops for user confirmation at business, form, code and screenshot steps
Includes a generated demo folder showing drafts and final files
Weaknesses
Requires OfficeCLI pinned to 1.0.151; no fallback DOCX writer if it is missing
Exact page count needs Microsoft Word on Windows; otherwise check manually in Word or WPS
README and workflow are in Chinese only
DOCX files must be exported to PDF manually before upload to the registry
no GPU
Needs Python 3.10+, OfficeCLI 1.0.151, Skill-capable coding agent, Node.js 18+ and @playwright/cli 0.1.20 (optional, screenshots), Chrome (optional), Microsoft Word on Windows (optional)
Multi-model workflow layer for Claude Code, delegating to Codex, Grok and Kimi
5.9k starsMITlast commit Oct 2026
CCG installs slash commands, hooks, skills and a compiled Go bridge (codeagent-wrapper) into Claude Code. Claude stays the orchestrator and picks one of 10 strategies per task, sending analysis and review to external models such as Codex, Grok, Kimi Code and Antigravity. Hooks inject task state every turn so context survives compaction, and tasks persist under .ccg/tasks.
Strengths
10 built-in strategies, from direct-fix to full-collaborate with dual-model review
Hooks re-inject task state every turn and after context compaction
Skills install individually; doctor --json outputs a versioned health report
Native plugin option installs skills only, with no binary
Weaknesses
Multi-model commands ship only via the npx install, not the native plugin
Requires Claude Code CLI; external models need their own CLIs installed
Installs a compiled Go binary into ~/.claude/bin as the model bridge
README mixes in sponsor blocks and unrelated projects
no GPU
Needs Node.js 20+, Claude Code CLI, Codex CLI (optional), Grok CLI (optional), Kimi Code CLI (optional), Antigravity (optional)
Models: Claude, Codex (OpenAI), Grok (xAI), Kimi Code (Moonshot), Antigravity
Agent skill that renders answers as one offline HTML page from Markdown
4.3k starsMITlast commit Oct 2026
An agent skill with a bundled Node CLI (`am`). The model writes a short Markdown draft, and the CLI lays out panels, flow, sequence, ER and tree diagrams, tables and themes into a single HTML file. It can also build narrated explainer videos. It works with Claude Code, Codex, Cursor, OpenCode and others.
Strengths
Model writes about 612 tokens vs 4,893 for hand-written HTML, per the README benchmark
Each page is one HTML file with no CDN links or web fonts, so it opens offline
On draft errors the CLI returns line number and a correct example for the agent to fix
Installs as a Claude Code plugin or via `npx skills add`; no npm install step
Weaknesses
Always-on mode is off by default and needs a rule added to your rules file
MP4 video export needs ffmpeg
Pages accumulate in ~/.answer-me-with-html/ until you run `am clean`
Benchmark covers 3 topics on one model (Claude Sonnet 5.5); other setups unknown
Plan, work and review skills for Claude Code, Codex CLI, Cursor and Grok
3.2k starsMITlast commit Oct 2026
A skill and plugin set that adds a plan, work, review, sync and release workflow to coding assistants. It keeps the spec and tasks in spec.md and Plans.md, runs checks, and requests a review separate from the implementation. A Go-native guardrail engine blocks operations such as secret reads, production deploys and forced pushes to main.
Strengths
Plan and task state lives in spec.md and Plans.md inside the project
Supported install paths for Claude Code, Codex CLI, Cursor and Grok
Guardrail engine is bundled Go binaries, no Node.js needed at runtime
Resumable work via /harness-work --resume latest
Weaknesses
Guardrails do not guarantee containment; coverage differs per tool
Cursor has no filesystem isolation; restrictions are applied only by Harness
Codex child agents can inherit parent permissions
More workers, reviews and retries raise costs on your model provider
no GPU
Needs Claude Code v2.1+, Codex CLI, Cursor or Grok, Git, Go 1.26.0+ (only to build from source)
A kit of instructions, DTCG design tokens, 138 design-system references and 26 skills that steer Claude Code and other agents toward accessible, token-driven UI. It installs as a Claude Code plugin, an MCP server, or via npm or Homebrew. It runs 52 gates (WCAG contrast, axe, target size, keyboard, overflow), 34 of which need a real browser through Playwright.
Strengths
Gates are runnable checks (per-state contrast, axe, target size), not just prompt guidance
Several install paths: Claude Code plugin, MCP server, npx, Homebrew, AGENTS.md for Codex and Cursor
Live demo with 49 rendered pages built from the kit's own tokens
MCP server has no SDK dependency and returns gate output with exit codes unedited
Weaknesses
34 of 52 gates need Playwright and a browser, which plugin install does not pull in
Plugin root CLAUDE.md is not loaded, so rules rely on the design-doctrine skill
Evidence of quality is small: two blind runs, and the README lists unproven areas
Built mainly around Claude Code; other agents get read-only AGENTS.md guidance
no GPU
Needs Node.js, Playwright (for browser gates)
Models: Claude Code, Codex, Cursor, Copilot, Aider
Git extension that records which agent, model and prompt wrote each line
2.9k starsApache-2.0last commit Oct 2026
Git AI is a Rust git extension that stores line-level AI attribution in Git Notes, linking each AI-written line to the agent, model and session. Coding agents report their edits through `git-ai checkpoint`, so no heuristics are used. It adds `git ai blame` and `git ai stats`, and carries attribution through rebase, squash, stash and cherry-pick.
Strengths
No git hooks and no git wrapping, so git commands are not slowed
Attribution comes from agent reports, not AI-detection heuristics
Preserves attribution across rebase, cherry-pick, stash, squash and amend
Works locally and offline with no login; editor plugins exist for VS Code and Emacs
Weaknesses
`git mv`, `git filter-branch`/`filter-repo` and `git replace` are not tracked
Squash/rebase merges on hosting platforms need CI Actions or the paid Teams product
Secure prompt store and org-wide stats are in the Teams version, not the CLI
Bash-created files may be missed if the agent is not run from the repo root
no GPU
Needs git
Models: Claude Code, Codex, Cursor, GitHub Copilot, OpenCode
Agent skill that stitches video frames and subtitle strips into long images
2.7k starsMITlast commit Oct 2026
An Agent Skill (Codex, Claude Code, other Agent Skills hosts) plus a Python CLI that samples exact frames from a local video or YouTube URL and composes a JPG with one hero frame and up to 6 subtitle strips. Native mode crops burned-in subtitles from the pixels without redrawing; script mode draws reviewed lines from a JSON file and labels them as post-production.
Strengths
Native mode crops burned-in subtitles from pixels, with no OCR or redrawing
CLI subcommands (check-source, sample, band, render, render-script) run without an agent
Rejects duplicate frames and low-resolution or broken streams before rendering
Fixed 3:4 canvas or natural-ratio layout; Chinese, English and Korean examples
Open specification for SKILL.md folders that give AI agents reusable instructions
26.5k starsApache-2.0last commit Aug 2026
Defines the Agent Skills format: a folder with a SKILL.md file holding name and description metadata plus instructions, and optional scripts, references and assets. Agents load skills in three stages (discovery, activation, execution), so only names and descriptions sit in context until a task matches. Originally developed by Anthropic and released as an open standard; the repo holds the spec and docs, not a runtime.
Strengths
Minimal format: one SKILL.md with name and description is enough
Progressive disclosure keeps context small with many skills installed
Skills can bundle scripts, references and templates in one version-controlled folder
Supported by multiple agent clients, listed in a client showcase
Weaknesses
Specification and docs only; it is not an agent or runtime
Behavior depends on how each client implements skill loading
No tagged releases; versioning of the spec is not stated in the README
Skill quality and security of bundled scripts are left to authors
Collection of 18 Agent Skills on context and harness engineering for coding agents
18.2k starsMITlast commit Oct 2026
A set of SKILL.md directories covering context fundamentals, compression, memory, tool design, multi-agent patterns, evaluation and harness design. It installs as a Claude Code plugin marketplace entry, an Open Plugins manifest for Cursor and Codex, or by copying single skill folders into .claude/skills, .cursor/skills or .codex/skills. Six worked examples and a router benchmark are included.
Strengths
Installs via one /plugin command or by copying individual skill directories
Works across Claude Code, Cursor, Codex and other Open Plugins hosts
Skill router benchmarked: 600 calls per sweep across four models, with published results
Includes six worked examples, such as an LLM-as-judge TypeScript implementation with 19 tests
Weaknesses
Skills are guidance text; scripts are Python pseudocode, not a runnable library
Router top-1 accuracy ranges 0.84 to 0.92 across tested models
Skill activation depends on the host agent's skill discovery
Some skills are labeled NEW and their maturity is unknown
no GPU
Needs Claude Code, Cursor, Codex or another Agent Skills host
Models: Model-agnostic; router benchmarked on gemini-3.1-pro, composer-2, gpt-5.5, claude-opus-4-7
Pattern library and CLI for running scheduled agent loops around a codebase
11.6k starsMITlast commit Oct 2026
Loop Engineering is a library of documented patterns (daily triage, PR babysitter, CI sweeper, dependency sweeper, changelog drafter and others) for running coding agents on a schedule against a repo. An npm CLI (@cobusgreyling/loop) scaffolds a pattern for Claude, Grok, Codex or Opencode, then offers doctor, status, audit and cost commands. Rollout is staged from L1 report-only to L3 unattended.
Strengths
CLI scaffolds a pattern with init, then checks it with doctor, audit and cost
Eight documented patterns, each with cadence, week-one level and relative cost
Staged L1 to L3 rollout; week one is report-only
Examples for Claude Code, Grok, Codex, Opencode, OpenClaw and GitHub Actions
Weaknesses
Mostly docs and templates; the agent itself comes from a separate tool
PR Babysitter and CI Sweeper are rated high or very high cost
README warns unattended loops can ship mistakes and token costs can explode
Related features live in separate companion repos
no GPU
Needs Node.js (npx), GitHub, A coding agent CLI (Claude Code, Grok, Codex or Opencode)
Models: Claude Code, Grok, Codex, Opencode, OpenClaw
Desktop app and CLI that manage agent skills across coding tools
6.8k starsMITlast commit Oct 2026
Skills Manager keeps one central skill library (default ~/.skills-manager) and deploys skills to 54 supported agents, including Claude Code, Codex, Cursor and Gemini CLI, by symlink or copy. Skills install from Git, local folders, .zip/.skill archives or the skills.sh marketplace, and can be grouped into presets. A bundled CLI and a manage-skills skill let agents drive it too, and the library can be backed up to Git.
Strengths
Deploys one library to 54 agents via symlink or copy, with per-agent toggles
Bundled CLI with --json output and --dry-run on deploy, undeploy, sync, remove and adopt
Git-backed backup with skill-aware merging, restorable snapshots and plain Git remote
Shows skills installed outside the app and tracks upstream updates for Git-based skills
Weaknesses
Windows app can only copy skills into WSL; symlinks need the Linux CLI inside WSL
Applying a preset is a one-time copy, not a live sync
Skills over 100 MB are excluded from backup; secrets and machine wiring stay local
Desktop-only Tauri app; no web or server deployment described
.NET team's skills and plugins for Copilot, Claude Code, Cursor and Codex
5.7k starsMITlast commit Oct 2026
Official .NET team marketplace of agent skills, grouped into 16 plugins covering MSBuild, NuGet, EF data access, testing, upgrades, ASP.NET Core, Blazor, MAUI, WinForms and .NET AI. It installs through the plugin marketplaces of Copilot CLI, Claude Code, VS Code, Cursor and Codex CLI. It also ships GitHub Copilot custom agents and GitHub Agentic Workflows.
Strengths
Maintained by the .NET team, with 16 plugins split by area (MSBuild, NuGet, EF, Blazor, MAUI)
Installs via marketplace in Copilot CLI, Claude Code, VS Code, Cursor and Codex CLI
Public dashboard reports token use, elapsed time and activation rates per skill
Skills follow the agentskills.io open standard, so they port across hosts
Weaknesses
Custom agents (*.agent.md) use Copilot conventions and are not installed as native Codex agents
VS Code plugin support is a preview feature and may change
No native Codex agent files (.codex/agents/*.toml) are shipped
Only useful for .NET work
no GPU
Needs Copilot CLI, Claude Code, VS Code, Cursor or Codex CLI (v0.121.0+ for plugin marketplace)
Agent Skills that teach coding assistants idiomatic, production-grade Go
3.6k starsMITlast commit Oct 2026
A collection of about 50 Agent Skills covering Go style, error handling, testing, security, concurrency, observability and performance, plus library skills for gRPC, cobra, viper, wire, fx and samber/*. Skills load on demand, so only short descriptions sit in context. Installs through the skills CLI, a Claude Code plugin, or by cloning into the skills directory of Codex, Gemini CLI, Cursor, Copilot, OpenCode and others.
Strengths
Works with Claude Code, Codex, Gemini CLI, Cursor, Copilot, OpenCode and Antigravity
Lists per-skill token weights, so context cost is visible before installing
README reports 97% eval pass with skills versus 57% without, across 3,439 checks
Skills load on demand rather than all at once
Weaknesses
Skills cross-reference each other; installing a subset gives inconsistent guidance
Go only; git, CI and PR workflow skills need a separate plugin
Local issue board embedded in the Codex desktop app, with a CLI and Skill
3.4k starsApache-2.0last commit Oct 2026
Runs a Node.js HTTP service with a React issue board backed by SQLite, and injects it into the Codex desktop app as a sidebar panel over CDP. A bundled `taskctl` CLI and the `manage-taskboard` Skill let Codex tasks read issues, move them to in_progress and in_review, and record the conversation ID. Desktop packages exist for macOS, Windows and Ubuntu 24.04 x64, and an optional Cloudflare deployment uses Workers, D1 and R2.
Strengths
Same HTTP API serves the React UI and the taskctl CLI used by the Skill
Skill moves issues to done only after explicit user confirmation
Single SQLite file at .data/taskboard.sqlite, no separate database server
Live updates to all open clients via server-sent events
Weaknesses
Embedding relies on unauthenticated CDP and a CSP bypass, so only trusted local code should run
LAN mode has no account authentication; default bind is 0.0.0.0
Linux packages cover Ubuntu 24.04 x64 only; no ARM64, Fedora or RPM
Windows CI builds are unsigned and do not auto-update
Pre-execution hook that blocks destructive commands and secret access by coding agents
1.8k starsMITlast commit Oct 2026
CC Safety Net installs as a hook in coding agent CLIs and parses each shell command or file access before it runs. It blocks things like `git reset --hard`, `git push --force`, `rm -rf` on dangerous targets, `curl | bash`, and reads of SSH keys, `.env` files and `~/.aws`, including when wrapped in `bash -c` or `python -c`. It is not a sandbox and does not restrict processes, filesystem permissions or network egress.
Strengths
Supports 16 coding CLIs, including Claude Code, Codex, Cursor and Gemini CLI, on Windows, macOS, Linux
Parses command semantics, so wrapping in bash -c or reordering flags does not bypass rules
Standard, Strict and Paranoid presets, a GUI editor, and rulebooks for Terraform, AWS, gcloud, Azure
Policy lives in a committed .cc-safety-net/ directory; a broken config never blocks commands
Weaknesses
Not a sandbox: no process containment, filesystem permissions or network egress monitoring
Requires Node.js 18 or higher
Must keep the @latest qualifier on npx, or a stale cached copy may run
Protection depends on each CLI's hook support; per-CLI requirements vary
Rules, skills, agents and workflows for Google Antigravity projects
8.2k starsMITlast commit Aug 2026
AG Kit installs an `.agents/` workspace into a project with 20 specialist agent roles, 47 skills, 13 slash-command workflows and 6 rules, plus a file-based memory directory. It registers a narrow PreToolUse hook that blocks root-filesystem deletion, drive formatting and raw-disk overwrites. Updates are merge-aware with backup and rollback, and a local Antigravity plugin bundle can be built.
Strengths
Merge-aware `ag-kit update` with --dry-run, automatic backups and `ag-kit rollback`
Safety hook blocks destructive commands but allows cleanup like deleting dist/ or node_modules/
Components carry SemVer contracts, a manifest lock and a dependency graph to detect drift
Plugin bundle includes a SHA-256 content inventory (PLUGIN_CONTENTS.json)
Weaknesses
Production support is limited to Google Antigravity; other tools are not guaranteed
Hook covers only high-confidence destructive patterns and does not replace permissions or sandboxing
Needs Node.js 22 for repo tooling and Python 3.10 for validators
Default MCP example contains YOUR_API_KEY placeholders, so the doctor warns until configured
no GPU
Docker
Needs Node.js 22+, Python 3.10+, Git, Google Antigravity
Trail of Bits plugin marketplace of security and testing skills for Claude Code
7.8k starsCC-BY-SA-4.0last commit Oct 2026
A Claude Code plugin marketplace from Trail of Bits that bundles skills for security auditing, testing and development workflows. Plugins cover smart contract review, C/C++ and Rust review, Semgrep and CodeQL, fuzzing, mutation testing, constant-time analysis and YARA rules. Codex can load it through its Claude marketplace compatibility, and a ChatGPT workspace import uses the marketplace.json.
Strengths
Installs with one command: /plugin marketplace add trailofbits/skills
Claude Code skills and hooks for building n8n workflows through n8n-mcp
6.5k starsMITlast commit Oct 2026
A pack of 14 SKILL.md skills plus an always-on router skill and hooks that guide Claude in building n8n workflows with the n8n-mcp server. Topics include expression syntax, Code node JavaScript and Python, validation errors, error handling, sub-workflows and AI agents. One skill covers deploying self-hosted n8n behind Caddy with Docker Compose.
Strengths
Covers specific gotchas, such as webhook data under $json.body and Code Tool string returns
Hooks re-inject reminders on SessionStart, before get_node calls and after validate_workflow
Installs as a Claude Code plugin; skill zips also work on Claude.ai and other SKILL.md hosts
Includes a self-hosting skill with single and queue mode Docker Compose templates
Weaknesses
Requires the separate n8n-mcp server to be installed and configured
Hooks only run in the Claude Code or Codex plugin install, not Claude.ai uploads
Self-hosting skill sits outside the workflow-building router and hooks flow
Test instructions are hedged ('if testing framework available'), so the eval setup is unclear
CLI and skills that teach coding agents to build and deploy ADK agents
6.2k starsApache-2.0last commit Oct 2026
agents-cli installs seven skills into coding agents such as Claude Code, Codex and Antigravity CLI. It also provides commands to scaffold ADK Python agent projects, run evals, and deploy to Agent Runtime, Cloud Run or GKE. The CLI also works standalone from a terminal, and local create, run and eval need only an AI Studio API key.
Strengths
Covers the full lifecycle: scaffold, eval, deploy, publish and observability commands plus matching skills
Eval commands include synthesized multi-turn datasets, failure clustering and prompt auto-tuning
scaffold enhance adds deployment and CI/CD to existing agent projects
Local create, run and eval work with an AI Studio key, no Google Cloud needed
Weaknesses
Targets ADK agents on Google Cloud and Gemini; other frameworks are not covered
Deployment and publish features require a Google Cloud project
Telemetry is on by default; opt out with AGENTS_CLI_TELEMETRY=0 or DO_NOT_TRACK=1
Requires Python 3.11+, uv and Node.js installed
no GPU
Needs Python 3.11+, uv, Node.js, Google Cloud (for deploy), AI Studio API key (for local runs)
Dependency manager for agent skills, plugins, instructions and MCP servers
4.1k starsMITlast commit Oct 2026
APM reads an apm.yml manifest and installs agent primitives (instructions, skills, prompts, hooks, plugins, MCP servers) from git hosts into the config formats of Copilot, Claude, Cursor, Codex, Gemini, Windsurf, OpenCode, Kiro and Grok Build. It resolves transitive dependencies, pins them in a lockfile with content hashes, and scans installs for hidden Unicode. An apm-policy.yml lets an org restrict allowed sources and primitives.
Strengths
One apm.yml and lockfile reproduce the same setup across nine listed agent clients
Resolves transitive dependencies from GitHub, GitLab, Bitbucket, Azure DevOps and any git host
Spec-driven workflow skills for Claude Code, Codex, Cursor and other coding agents
3.7k starsMITlast commit Sep 2026
cc-sdd installs a set of 17 Agent Skills into your project via npx. They cover discovery, requirements, design, tasks and implementation, in a style modeled on Kiro specs. With native subagents, /kiro-impl runs each task through a fresh implementer using TDD, then an independent reviewer. Hosts without subagents run the same flow inline.
Strengths
Same 17-skill set across 8 agents, installed with one npx command
Specs (requirements.md, design.md, tasks.md) carry boundary and dependency annotations
Per-task TDD, independent review and auto-debug when native subagents exist
Templates and rules under the specs settings directory are editable; 14 languages supported
Weaknesses
Only Claude Code and Codex are stable; the other agents are beta
README says installing the skills does not verify a host can run the full autonomous loop
Legacy /kiro:* command modes are deprecated, so v1/v2 users must migrate
Devin adapter has unverified authenticated task execution
no GPU
Needs Node.js and npx, a supported AI coding agent (Claude Code, Codex, Cursor, Copilot, etc.)
Code-context CLI giving coding agents a ranked call graph and edit checks
2.5k starsApache-2.0last commit Oct 2026
ripwire is a single C++23 binary that indexes a repository into a deterministic, ranked call graph and answers orientation questions (`ripwire .`, `--for`, `--callers`, `--path`, `--recall`). On edits it reports blast radius, tests that reach the change, and quality regressions via `--edit-check` and `--quality-delta`. The installer also writes skills for agents such as Claude Code, Codex, Cursor, Windsurf, Gemini, opencode and aider; an MCP server is optional.
Strengths
Single offline binary with no runtime dependencies, API key, embeddings or daemon
Indexes this repo in 0.25 s using 6.6 MB, per its own benchmark
Answers disclose truncation, label floors, and give a next= to recover cut rows
Covers about 25 languages and installs skills for seven agent CLIs in one line
Weaknesses
Benchmarks and the field report are the authors' own, not independent measurements
Contract checker arity heuristic over-counts on defaulted trailing parameters
Plain grep can still win on broad, common-word conceptual queries
MCP server schemas consume agent context every session; CLI is the preferred route
Agent Skill and MCP server that edits video with local FFmpeg
2.3k starsMITlast commit Oct 2026
Installs as a skill for Claude Code, Cursor or Codex, and also runs as an MCP server over stdio. It gives the agent 42 Python scripts wrapping ffmpeg and ffprobe for cutting, joining, captions, loudness, reframing and platform exports. The workflow is probe, edit (stream copy where possible), then check and look at a contact sheet. No cloud or API keys are needed.
Strengths
Every tool supports --dry-run, --json and --plan, so edits can be previewed before rendering
Runs offline on Python standard library; only ffmpeg and python3 are required
Contract-derived MCP schema; CI tests FFmpeg 5.1, 6.1, 7.1, 8 and 9
Stream-copies cuts and joins where possible; never overwrites input files
Weaknesses
Makes no editorial decisions; no face, subject or highlight detection, only measured proxies
Visual verification depends on the calling agent having vision
Some tools need specific FFmpeg builds; Homebrew ffmpeg lacks the subtitles filter for caption.py
Transcription needs a separate speech engine such as Parakeet or whisper installed
Code review skills that cite classic engineering books for each finding
1.5k starsMITlast commit Oct 2026
brooks-lint is a set of Agent Skills and a Claude Code plugin that reviews code against six production decay risks and six test-suite decay risks drawn from twelve engineering books. Each finding is reported as Symptom, Source, Consequence, Remedy with a severity and a 0–100 Health Score. It provides six commands (review, audit, debt, test, health, sweep) and optional SARIF export and CI gates.
Strengths
Every finding carries a book citation, severity label and concrete remedy
Installs on Claude Code, Codex, Gemini, Cursor, Windsurf, Copilot and others
Per-project .brooks-lint.yaml sets strictness, disabled risks, ignores and custom risks
Parser benchmark is reproducible locally with npm run benchmark on a frozen 30-report corpus
Weaknesses
Review quality depends on the host LLM; live evals need an Anthropic API key
Maintainer says only Codex is tested going forward; other platforms are unverified
Headline benchmark table is described as illustrative and compares against plain Claude
Does not check syntax or style; it complements a linter rather than replacing it
no GPU
Needs Claude Code, Codex CLI or another Agent Skills host, ANTHROPIC_API_KEY (live evals only)
Slash commands, agents and modes that add structure to Claude Code
23.9k starsMITlast commit Sep 2026
SuperClaude is a Python package that installs 30 `/sc:*` slash commands into Claude Code, along with 20 agent personas and 7 behavioral modes, all delivered as instruction files. It can also install 8 optional MCP servers (Tavily, Context7, Serena, Playwright and others) with `superclaude mcp`. Install is via pipx or an install script; a TypeScript plugin system is planned for v5.0 and is not available yet.
Strengths
Installs with `pipx install superclaude` and `superclaude install`; `superclaude doctor` verifies it
30 slash commands cover brainstorming, implementation, testing, git, research and session save/load
MCP servers are optional; the framework works without them
MIT licensed, with docs for commands, agents, modes and flags
Weaknesses
Targets Claude Code only; separate sibling projects exist for Gemini and Qwen
Claimed 2-3x speed and 30-50% token savings with MCPs come with no benchmark details
Plugin marketplace install is not available yet; v5.0 has no ETA
README mixes v4.1, v4.2 and v4.3 notes, and the research examples are unclear
Shared memory and skill store for coding agents across sessions and devices
14.6k starscustom licenselast commit Oct 2026
memU runs as a sidecar to desktop coding agents (Codex, Claude Code, Cursor, OpenClaw, Hermes, WorkBuddy, Cola, pi). A scheduled task reads session logs, the agent distills them into Markdown memory and skills, and memU embeds and stores them. A standing instruction in the host's instruction file makes the agent run a retrieve command before answering. It can use MemU Cloud or a local SQLite or Postgres store.
Strengths
Per-host adapters for Codex, Claude Code, Cursor, OpenClaw, Hermes and others, plus generic memu-agent detect
Local storage options: SQLite for a single writer, Postgres with pgvector for concurrent access
Core service makes no LLM calls; the agent writes the skill Markdown, which stays readable
Skill that runs coding agents in a measured keep-or-revert loop
6.7k starsMITlast commit Aug 2026
Autoresearch is a skill pack for Claude Code, OpenCode and OpenAI Codex, based on Karpathy's autoresearch. Given a goal, a metric and a verify command, the agent makes one change per iteration, commits it, runs the check, and keeps or reverts the commit. It adds 14 subcommands (plan, debug, fix, security, ship, scenario, regression and others) and, for Claude Code only, hook guardrails.
Strengths
Works with Claude Code, OpenCode and Codex; installs via npx skills, plugin or install.sh
Each iteration is a git commit with automatic revert on failure and a TSV results log
Every command has a default iteration bound; unlimited runs are opt-in
Claude Code hooks block force-push, rm -rf, .env reads and node_modules context bloat
Weaknesses
Hook guardrails exist only for Claude Code, not OpenCode or Codex
Classic loop needs a verify command that outputs a number
README calls the hooks defense-in-depth, not a security sandbox
Autonomous commit and revert loop on your repo; token spend scales with iterations
OmO agent harness packaged as a plugin for OpenAI Codex
3.8k starsMITlast commit Oct 2026
LazyCodex installs the oh-my-openagent (OmO) harness into Codex through an npx installer or the Codex plugin marketplace. It adds the $init-deep, $ulw-plan, $start-work and $ulw-loop commands, skills, hooks and sub-agent roles in ~/.codex/agents/. It also routes subtasks to different GPT models by task category.
Strengths
Installs with one npx command; uninstall and a doctor health check are included
Plan, execute, verify workflow: plans go to plans/<slug>.md, loops capped at 100 or 500 iterations
Generates hierarchical AGENTS.md project memory with $init-deep
Hooks run only after approval in Codex's startup review
Weaknesses
Codex marketplace install is labeled experimental; npx is the primary path
Core engine is a git submodule of oh-my-openagent, so this is a thin wrapper
README documents only OpenAI GPT model routing
Sub-agent roles rely on Codex spawn_agent; falls back to prompt text without agent_type
Claude meta-skill that logs corrections and proposes updates to your other skills
3.5k starscustom licenselast commit Oct 2026
task-observer is a SKILL.md bundle with reference files and five helper scripts. It runs alongside Claude sessions, records corrections and repeated manual work as observation files, and proposes edits to existing skills or candidates for new ones. Changes go to staged copies, and nothing goes live until you install the updated bundle.
Strengths
Proposals only: changes land in staged copies, installed skills stay untouched
Installs via `npx skills add` or a .skill bundle upload in the Claude apps
Writes one small observation file per entry under skill-observations/observation-log/
README reports 80+ contributors and 200+ issues and PRs closed
Weaknesses
Description matching under-triggers; needs a CLAUDE.md instruction or session-start hook
Written around Claude concepts; other platforms need adaptation
Claude.ai web and mobile get handoff documents only, with no filesystem logging
README says small skill libraries may be served as well by built-in memory
no GPU
Needs Node.js (for skills CLI install), Claude or another skills-compatible agent
Skill stack that routes coding-agent tasks through planning, delegation and verification
3.4k starsMITlast commit Oct 2026
Port of Lauren Tan's Cursor pstack skill set to Claude Code, Codex, Pi, GitHub Copilot, Muse and other harnesses. A session-start hook points the agent at poteto-mode, which picks a playbook (bug fix, feature, refactor, performance, PR maintenance) and delegates with skills like architect, arena, swarm, how and why. It also has a formal-verification skill for concurrency invariants that tests miss.
Strengths
Installs as a plugin on Claude Code, Codex, Copilot and Muse; Pi via git install
Routing hook sends tasks to poteto-mode automatically at session start
No server or telemetry; scripts run locally, PR tools use your GitHub CLI login
Per-role model and reasoning-effort settings via setup-pstack
Weaknesses
Harnesses without hooks (OpenCode, Gemini CLI) need manual symlinks and a named skill call
Codex and Muse run the hook only after you trust or approve it
Copilot build tested only on CLI 1.0.87 to 1.0.92; copilot -p lists no skills on 1.0.92
Anything the skills read, including transcripts, is sent to your model provider
no GPU
Needs Claude Code 2.1.292 or later (Claude Code install), GitHub CLI (PR tools)
Skill and guard that stop coding agents from scope creep and needless defensive code
3.1k starsMITlast commit Oct 2026
Stop That Shit is a skill plus a tool-call guard for coding agents (Codex, Claude Code, OpenCode, Hermes Agent CLI, Pi, Oh My Pi). The skill sets rules against scope creep, unused hashing, read-only violations and repeated work. The guard blocks writes in review/answer/monitor modes, writes outside files=, new dependencies, extra subagents and recognizable hash operations unless allowed with flags like deps=allow or hash=allow.
Strengths
Per-task contract with modes (review, answer, monitor, change, watch) and files=, deps=, hash=, agents= flags
Covers six agent hosts, each with install steps, plus a skill-only option
Guard checks are narrow and documented: writes, dependencies, subagents, hashes, paths
Ships 18 public Bad/Good cases and an evidence file with paired runs
Weaknesses
Guard stays inactive until a mode is used; default state is OBSERVING / unconfirmed
A deny response does not prove the host skipped the action; isolation is left to the host sandbox
Hash detection covers only recognizable operations; other over-engineering relies on skill judgment
Main README is Chinese; English and Korean versions are separate files
no GPU
Needs Node.js 18+, Node.js 22.19+ (Pi 0.84.4)
Models: Codex, Claude Code, OpenCode 1.18.18+, Hermes Agent CLI, Pi Coding Agent 0.84.4
CLI that answers repo questions with files and source excerpts for coding agents
2.9k starsMITlast commit Oct 2026
jg takes a natural-language question about a repository and prints relevant files, reading leads and verbatim source excerpts with line references to stdout. It uses the Jev model, called through a provider key, to judge relevance across folders, files and declarations. A bundled skill, installed with `jg skill`, teaches coding agents such as Claude Code, Codex and OpenCode to call it.
Strengths
Installs an agent skill for Claude Code, Codex and OpenCode with one command
Respects ignore files and skips hidden, binary and .env files by default
No Python, Bun or ripgrep install needed; runs on Node 22+
Published eval: 8/10 SWE-bench tasks solved with and without it, Sol cost down 28.6%
Weaknesses
Source content is sent to the chosen provider; secret filtering is heuristic, not redaction
Needs a provider key for Vercel AI Gateway, TypeSafe, OpenRouter or OpenCode Zen
Benchmark covers ten Python tasks; total cost including Jev fell 25.8%, no speed gain claimed
Declaration parsing covers only Python, TS/JS, Go and Rust; other text uses a fallback
no GPU
Needs Node.js 22+, API key for Vercel AI Gateway, TypeSafe, OpenRouter or OpenCode Zen
Claude Code skill suite for writing, auditing and localizing SEO blog posts
2.5k starsMITlast commit Oct 2026
A Claude Code plugin with one orchestrator and 31 sub-skills behind 30 /blog commands, covering writing, rewriting, SEO and AI-citation audits, topic clusters, translation and audio narration. Each draft must pass a 5-gate delivery contract (100-point rubric, 90 minimum, up to 3 iterations) and ships as markdown, HTML, PDF, hero image, screenshots and a review report.
Strengths
Code-enforced delivery gates check format, screenshots, review score and links before output
Desktop app that installs AI agent Skills once and syncs them to coding tools
1.8k starsMITlast commit Oct 2026
Skills Hub is a desktop app for macOS, Windows and Linux that keeps one library of agent Skills and syncs them to 48 supported coding tools, including Claude Code, Codex, Cursor and OpenCode. Skills can be installed from collections, Git repositories or local folders, scoped globally or per project, and synced across devices through a GitHub, GitLab or Gitee repo.
Strengths
Supports 48 coding tools plus custom tool directories
Imports Skills from Git, local folders or curated collections
Global or project scope, with bulk enable, disable and update
Cross-device sync via GitHub, GitLab or Gitee repository
Weaknesses
Desktop app only; no Docker or hosted server option
macOS may flag the app as damaged or unverified on install
Collections are maintained in a separate repository
AI management setup is a manual step in Settings
no GPU
Needs Git hosting repo (GitHub, GitLab or Gitee) for optional device sync
Plugins for the DSH agent: runtime injector, routing presets, graded task protocol
7k starsMITlast commit Sep 2026
Bundles three add-ons for the DSH agent (@deepseek-ai/dsh): an injector that adds dev_* tools for hot-reload without restart, router-standard and router-spec presets that pick a reasoning persona per task type, and a graded two-level task protocol with a red-team gate and audit endpoint. Installed with dsh plugin add or install.ps1. Targets DSH >=0.1.0-rc.6 <0.2.0, a developer preview with announced breaking changes.
Strengths
Three components in one repo, installable with a single dsh plugin add command
Includes measure.mjs to recompute the reported metrics on your own session logs
Experiment reports (P1-P23) and measured data are published alongside the presets
Presets replace only the persona section, so plan-mode boundaries are kept
Weaknesses
Works only with DSH, a developer preview with announced breaking changes
Install steps and scripts are PowerShell; other platforms are not documented
Main README and docs are in Chinese, with a separate English README
Anthropic-API MLX server that runs Claude Code on local models on Apple Silicon
3.4k starsMITlast commit Oct 2026
A Python server (proxy/server.py) runs open models with Apple MLX and answers Claude Code in the Anthropic Messages API format, with streaming. It converts Gemma, Qwen and Llama-style tool calls into Claude Code tool calls, repairs malformed tool JSON, and reuses the prompt cache between turns. setup.sh picks a model based on the Mac's memory and creates a Desktop launcher.
Strengths
Speaks the Anthropic API directly, with no LiteLLM or OpenAI translation layer
setup.sh picks a model by RAM, from Hermes 4 14B at 16 GB to Qwen 3.8 27B at 96 GB
Launchers disable Claude Code's background traffic; README shows an lsof check for network connections
Includes tool-call parser tests and a benchmark script for its cache numbers
Weaknesses
Apple Silicon Macs only (M1 or newer)
Under 16 GB, the README says tool use is unreliable with Gemma 4 E4B
Benchmark scores come from the Agent-12 harness, not from inside Claude Code
Browser, voice and phone modes need separate repos or tools that setup.sh does not install
ESLint and Oxlint rules that enforce Tailwind design systems with fix guidance for agents
3.3k starsMITlast commit Oct 2026
@shadcn/lint is a lint plugin for ESLint 9.30+ and Oxlint 1.80+ that checks Tailwind v4 class usage in React, Vue and Svelte. Rules cover restyling components via className, raw colors, arbitrary values, inline styles, unknown classes and dynamic class strings. Errors suggest fixes from your components, variants and theme, and can be customised with per-component contracts and message placeholders.
Strengths
Error messages name allowed sizes, variants and theme file, drawn from your own components
Per-component contracts (allow/deny patterns) without changing component code
Runs on both ESLint and Oxlint across React, Vue and Svelte
Works without shadcn/ui via componentImports settings
Weaknesses
Tailwind v4 only
Oxlint reads only script blocks in Vue and Svelte files, not templates or markup
Variants from typed props are resolved in React only
Eval results come from the authors' own runs; independent numbers unknown
no GPU
Needs Node.js 20.19+, ESLint 9.30+ or Oxlint 1.80+, Tailwind v4
CLI that syncs one set of Markdown rules to many coding agents' config files
3k starsMITlast commit Oct 2026
Ruler reads Markdown rules from a `.ruler/` directory and writes them into the native instruction files of about 30 coding agents, such as AGENTS.md, CLAUDE.md, .clinerules and .goosehints. It also propagates MCP server settings and skills, supports nested `.ruler/` directories for monorepos, and manages .gitignore entries. A `revert` command restores files from .bak backups.
Strengths
Covers about 30 agents including Copilot, Claude Code, Cursor, Codex CLI, Gemini CLI and Aider
Single .ruler/ source with ruler.toml for per-agent selection and output paths
revert restores originals from .bak backups; --dry-run previews changes
Also distributes MCP server config and skills, not just rule text
Weaknesses
README labels it a Beta Research Preview and asks users to test carefully
Nested rule loading is marked experimental and may change
Subagents support is experimental and disabled by default
Unity Editor plugin exposing 839 REST skills for AI coding agents to drive scenes
1.9k starsMITlast commit Oct 2026
A Unity package (2022.3+) that runs a REST server inside the Editor and ships a skills template for Claude Code, Codex, Cursor, Antigravity, OpenCode and Kimi Code. Agents call endpoints to edit scenes, prefabs, materials, scripts and packages. It adds Approval/Auto/Bypass permission modes, dry-run previews, a JSONL audit log and transactional batches with rollback.
Strengths
839 skills across 54 categories, including Cinemachine, ProBuilder, Netcode, XR and YooAsset
Three permission modes, per-skill allowlist, and JSONL audit log at Library/UnitySkillsAudit.jsonl
POST /skills/batch runs as a transaction with rollback; ?mode=dryRun previews changes without writing
One-click installers for six agent tools from the Unity panel; installs via UPM Git URL
Weaknesses
Unity-only; 2022.3+ and Unity 6 are the maintained baseline
Domain Reload or compilation makes the REST server briefly unavailable; retries needed
Comparison table against other tools is self-reported from a 2026-07 survey
Large skill surface; advisory modules and routing add context load for agents
no GPU
Needs Unity 2022.3+, An AI coding agent (Claude Code, Codex, Cursor, Antigravity, OpenCode or Kimi Code)
Postgres skills and doc-search MCP server for AI coding agents
1.9k starsApache-2.0last commit Oct 2026
pg-aiguide gives coding agents PostgreSQL guidance on schema design, indexing, data types, constraints and naming. It also offers a hosted MCP server with semantic and BM25 search over the PostgreSQL, TimescaleDB and PostGIS docs. It installs as Agent Skills via `npx skills`, as an MCP endpoint, or as a Claude Code plugin.
Strengths
Three install paths: npx skills, MCP URL, or Claude Code plugin
Search is version-aware for the PostgreSQL manual and also covers TimescaleDB and PostGIS
Hosted MCP endpoint means no local server is needed to start
Config snippets for Cursor, Codex, Gemini CLI, VS Code, OpenCode and Windsurf
Weaknesses
Docs search depends on a public endpoint hosted by TigerData
pgvector docs are listed as coming soon
Quality claims (4x constraints, 55% more indexes) come from a single demo prompt
Skills are opinionated; the README does not say how to customize them
no GPU
Docker + Compose
Needs Node.js (for npx skills), TigerData-hosted MCP server (mcp.tigerdata.com/docs)
Git-shared project memory for engineers and their coding agents
1.8k starsMITlast commit Oct 2026
MEX stores architecture notes, decisions and handoffs as Markdown in the repository, and links them to code symbols through a local Tree-sitter graph in SQLite. Engineers review and edit it in a local Hub on 127.0.0.1, while agents read it through the CLI and installed skills. Sharing happens through ordinary Git commits, with no hosted service.
Strengths
Canonical memory is plain Markdown in the repo, shared via normal Git push and pull
Wiki claims are grounded to code symbols, with drift flags for review
Needs no hosted service, account or model key; indexes are rebuildable local SQLite
Relays record progress, tests and next steps for handing work to a teammate
Weaknesses
Code graph covers TypeScript, JavaScript, Python and Rust; C# is partial
Requires Node.js 22.5+ with SQLite FTS5 support in that build
Team must upgrade together before sharing the 0.8.4 format
Hub is local per checkout; Relays are not delivered until pushed through Git
Agent Skills for Google Cloud, Firebase, Ads and Analytics
21.5k starsApache-2.0last commit Oct 2026
Google's collection of Agent Skills covering Google Cloud services such as GKE, BigQuery, Cloud Run, AlloyDB and Spanner, plus Firebase, Genkit, Google Ads and Analytics APIs. Skills install into coding agents with `npx skills add google/skills`. The repo also bundles plugins (skills plus MCP servers) for Claude Code, Codex and Antigravity CLI.
Strengths
Covers many Google products: GKE, BigQuery, security, Ads, Analytics, Genkit
One npx command installs; pick skills, target agents and project or global scope
Plugin bundles for Claude Code, Codex and Antigravity CLI
Maintained by Google; `npx skills update` pulls newer versions
Weaknesses
Scoped to Google products; little use for other clouds or stacks
No tagged releases; updates follow the main branch
Documented install path requires Node.js and npx
Some skills are marked preview, such as the API Gateway remote MCP server
no GPU
Needs Node.js (npx), Google Cloud account for most skills
Skill router that plans tasks and assigns local agent Skills to each step
3.8k starsApache-2.0last commit Oct 2026
VibeSkills is a Skill for coding agents that confirms the requirement, picks an L or XL workflow, then scans your local Skill folders and assigns a few relevant Skills to each subtask. It records completed, failed and blocked units and checks results against the plan before acceptance. The README reports a SkillsBench run with 195 Skills (DeepSeekV4Flash-VE, OpenHands): +21.12 pp mean reward, 29.6% fewer tokens.
Strengths
Benchmark study with public data and reproduction steps is linked from the README
Shortlists Skills from compact local indexes; reads full SKILL.md only for retained candidates
Records completed, failed and blocked work units so a later session can resume
Works over your own, team and third-party Skill folders without calling every Skill
Weaknesses
Benchmark covers one model and agent setup (DeepSeekV4Flash-VE with OpenHands)
Requires approval steps before work starts, adding overhead on small tasks
Supported agent hosts are not listed in the visible README excerpt
README is truncated here, so install requirements are unknown
Persistent semantic memory for Claude Code, Codex, OpenClaw and OpenCode agents
2.8k starsMITlast commit Oct 2026
memsearch captures each agent conversation turn as an LLM-written summary in daily Markdown files, then indexes them in Milvus with hybrid search (dense vector, BM25, RRF). Plugins exist for Claude Code, Codex, DeepSeek Harness, OpenClaw and OpenCode, and a CLI and Python API allow use in custom agents. Recall runs in three layers: search, expand, then raw transcript.
Strengths
Markdown files are the source of truth; the Milvus index is rebuildable
One memory shared across five agent platforms
Default local ONNX bge-m3 embeddings need no API key; Milvus Lite needs no setup
Also exposes a CLI and Python API for custom agents
Weaknesses
First run downloads a ~558 MB embedding model from HuggingFace
Capture summaries call an LLM on every turn, adding model usage
Codex install needs `codex --yolo` for ONNX network access
Team use needs a separate Milvus server or Zilliz Cloud
Plugin that trims wasted context and keeps state across compaction in coding agents
2.7k starscustom licenselast commit Oct 2026
Token Optimizer is a hook-based plugin for Claude Code, Codex, OpenCode, Cursor, Copilot and other coding assistants. It compresses command and search output, shows file re-reads as diffs or skeletons, checkpoints before auto-compact and restores afterward, and audits configs, skills and memory files. A local SQLite database feeds a dashboard with quality scores and dollar savings.
Strengths
Covers bash, search, JSON, file re-reads and large results, plus config and memory audits
Checkpoints before compaction and restores after, so work survives auto-compact
Zero dependencies, no telemetry, no MCP server; data stays in local SQLite
Installs on Claude Code, Codex, OpenCode, OpenClaw, Hermes, Copilot, Cursor and Pi
Weaknesses
PolyForm Noncommercial license restricts commercial use; not an OSI open-source license
Antigravity, Grok and Devin support is beta or contract-only
Claude Code cloud sessions load no plugins and need a separate setup script
Comparisons against Headroom and RTK are the author's own claims
Better Harness is a plugin and skill, run as /better-harness inside a coding agent. It collects project and session evidence, evaluates five Agent Work Loop dimensions, and writes findings with scoped repair plans and acceptance checks. Output is a self-contained report.html with report.md and findings.json, or a host-native Canvas in Qoder and Cursor. Missing evidence is marked explicitly rather than scored.
Strengths
Supports Claude Code, Codex, Qoder, Cursor, Copilot CLI, Qwen Code, Pi, Kimi Code, WorkBuddy, Grok
Each finding links to evidence, expected output, repair boundary and acceptance checks
Unobserved behavior stays explicit instead of becoming an unsupported score
Public sample report and Harness Inspector demo; MIT licensed
Weaknesses
No single entrypoint: install and invocation steps differ per host
Cursor plugin is not published; install plan is reported as unavailable
Copilot records no per-response token usage; VS Code Copilot Chat has no durable transcript
Plugin lifecycle CLI is marked Beta; history shows trends, not causal proof
Task-splitting and specialist-agent staffing for Claude Code, Codex and Cursor
1.6k starsApache-2.0last commit Oct 2026
Agentlas OS (Hephaestus) installs into coding-agent hosts and splits a request into tasks. It then sources candidate agents for each task from local profiles, a private cloud and the public Hub. Candidates are unranked and the model picks one. The chosen release is pinned by digest, and an independent verifier must pass the result. It also builds portable agents with /hep-build and reads and writes A2A v1.0 agent cards.
Strengths
Installs into many hosts: Claude Code, Codex, Cursor, Antigravity, OpenCode, Goose and others
SHA-256-checked installer, no sudo, and the README lists every path it writes
Picked agent releases are pinned by digest; an independent verifier must pass before a run is done
Uses standard A2A v1.0 agent cards, with signed-card identity checks for external agents
Weaknesses
Install is a curl-pipe-bash script that edits existing host configs such as ~/.codex/config.toml
No backup is written when it edits existing host configs, apart from Codex marketplace changes
Optional global router appends about 75 lines to each host's global instructions file
Token-savings and recall figures are the project's own measurements, not independently checked
no GPU
Needs Claude Code, Codex, Cursor or Antigravity host, Model/API access supplied by the caller
Anthropic's curated plugin marketplace for Claude Code
44.2k starsApache-2.0last commit Oct 2026
A Git repository that acts as the official plugin marketplace for Claude Code. Plugins install with `/plugin install {plugin-name}@claude-plugins-official` or through `/plugin > Discover`. The repo holds Anthropic-maintained plugins in `/plugins` and partner or community plugins in `/external_plugins`, each bundling any of MCP server config, slash commands, agents and skills.
Strengths
Installs directly from inside Claude Code with one slash command
Documents a standard plugin layout with an example plugin as reference
Plugin names are immutable slugs, with a renames map for migrations
Supports skill-bundle entries for repos that ship SKILL.md files without a manifest
Weaknesses
Only useful with Claude Code; not a standalone tool
Anthropic does not control or verify third-party plugin contents, which can change
No single license; each plugin carries its own LICENSE file
No tagged releases; plugins are pinned only by marketplace entry
Spec and CLI for describing a design system to coding agents
28.5k starsApache-2.0last commit Oct 2026
DESIGN.md is a file format that pairs YAML design tokens (colors, typography, spacing, rounded corners, components) with markdown rationale so coding agents get exact values and context. The TypeScript CLI lints files for broken token references and WCAG contrast, diffs two versions, and exports tokens to Tailwind v3/v4 and W3C DTCG. The format is at version alpha.
Strengths
Lint output is structured JSON with exit codes, so it can gate CI
Eleven lint rules, including broken refs and WCAG AA 4.5:1 contrast
Exports tokens to Tailwind v3 JSON, Tailwind v4 CSS and W3C DTCG
Also usable as a library via the @google/design.md/linter import
Weaknesses
Format is alpha; README warns the spec and schema will change
Component tokens limited to eight properties such as backgroundColor and padding
Strict section order is required; duplicate sections or keys reject the file
Windows needs the designmd alias because of the .md bin name
Research workflow skills, agents and hooks for Claude Code, with Zotero and Obsidian
5.9k starsMITlast commit Oct 2026
A bundle of skills, commands, agents, rules and Node.js hooks that structures academic research in Claude Code, from ideation and literature review through experiment analysis, reporting, paper writing and rebuttal. It integrates with Zotero via zotero-mcp and with an Obsidian vault for project notes. Separate branches cover Codex CLI, Kimi Code CLI and OpenCode.
Strengths
Covers the full path: ideation, literature, experiments, analysis, reporting, writing, rebuttal
Installer backs up overwritten files and records a manifest for safe uninstall
Installs as a Claude Code plugin or by copying individual skills
Separate branches for Codex CLI, Kimi Code CLI and OpenCode
Weaknesses
Workflow is tuned for computer science and AI research, per the README
Existing CLAUDE.md files must be merged by hand with the sidecar file
Plugin install does not distribute rules or configure settings.json
Zotero and Obsidian features need separate setup, including a zotero-mcp fork
CLI that builds a context graph of your codebase for coding agents
5.7k starsApache-2.0last commit Oct 2026
Potpie is a Python CLI with a local daemon that indexes code, source history, issues and docs into a context graph. A setup wizard installs Potpie skills and instructions into coding harnesses (Claude Code, OpenAI Codex, Cursor, OpenCode), which then query the graph with commands like `potpie resolve` and `potpie search`. A local web UI explores the graph.
Strengths
One setup wizard installs skills for Claude Code, Codex, Cursor and OpenCode
Indexes GitHub, Linear, Jira and Confluence, not only source code
Local graph explorer UI opened with `potpie ui`
Installs from PyPI with uv or pip; no manual ingest step needed
Weaknesses
README gives no RAM, GPU, database or model requirements
Only four coding harnesses and four integrations supported so far
`potpie login` enables account-backed or managed features; free scope unclear
No Docker or compose setup; relies on a local daemon
SwiftUI guidance skill for AI coding agents, with Instruments trace analysis
3.7k starsMITlast commit Oct 2026
An Agent Skills package that gives coding agents SwiftUI reference guides covering state management, view composition, performance, navigation, Swift Charts, macOS scenes, Liquid Glass and SDK 27 changes. Reference files load on demand. It also ships Python scripts that wrap xctrace to record and analyse Instruments traces for hangs, hitches and SwiftUI view updates.
Strengths
Reference files load on demand, so task context stays small
Bundled stdlib-only Python scripts record and analyse .trace files, emitting JSON and markdown
Installs via skills.sh, Claude Code, Codex and Cursor plugin manifests
Covers iOS 15 through iOS 27 API migration, plus macOS and Swift Charts
Weaknesses
Trace tooling needs Xcode's xctrace, so it works only on macOS
Focused on correctness and performance, gives no architecture or code-style guidance
SwiftUI only; concurrency, Core Data and testing are separate skills
Several reference files have no description in the README listing
Agent skill that rewrites AI-generated Japanese into plain, readable prose
2.5k starsMITlast commit Oct 2026
yomiyasu is a skill for Claude Code, Codex, Cursor and similar agents that edits Japanese text written by AI. It fixes figurative verbs, unclear subject-predicate relations and needless decoration in technical articles, specs, PR descriptions and internal reports, while keeping the original meaning. It bundles Python scripts for linting and diffing the text before and after editing.
Strengths
Bundled lint, diff and check scripts flag changed numbers, negations and swapped particles; stdlib-only Python
Installs via npx skills, gh skill, openskills, Claude Code plugin, or a ZIP for Claude.ai
12 document types and 3 reader levels, so normal conventions such as minutes bullets are kept
Rules forbid adding facts absent from the source; unresolved points go back to the writer as questions
Weaknesses
Japanese text only; the README is mainly in Japanese with a short English version
Only used when named in the request or via a standing instruction line
Scripts match fixed patterns and do not judge meaning or readability
Interactive mode, Windows, Cursor, Codex Cloud and Claude.ai ZIP are untested
no GPU
Needs Python 3.9+ (optional check scripts), Claude Code, Codex or another skill-capable agent
CLI that runs spec, plan, tasks and review workflows for AI coding agents
1.7k starsMITlast commit Oct 2026
Spec Kitty is a Python CLI that stores specs, plans, work packages and review state in the repository under kitty-specs/. It installs slash commands or skills for agents such as Claude Code, Codex, Cursor, Gemini and Copilot, and gives each work package its own git worktree so agents can work in parallel. Work moves through a spec, plan, tasks, next, review, accept, merge loop.
Strengths
Keeps specs, tasks, review and merge state in Git rather than in chat history
Isolated git worktrees under .worktrees/ per work package for parallel agents
Supports many agents, including Claude Code, Codex, Cursor, Gemini, Copilot, Windsurf, OpenCode
Local-first; README states no SaaS service is required
Weaknesses
4.x line is a release candidate; stable launch acceptance is stated as pending
Hosted auth, sync and tracker flows were removed from open-core
Overkill for one-off edits, small scripts or teams not using Git
Does not write code itself; depends on a separate coding agent
no GPU
Needs Python 3.11+, Git, An AI coding agent (Claude Code, Codex, Cursor, etc.)
Models: Claude Code, Codex, Cursor, Gemini, GitHub Copilot
Skill pack that makes AI coding agents plan, verify and report evidence
1.4k starsMITlast commit Oct 2026
Aegis is a method pack of SKILL.md-style skills that coding agents load to align with a project's baseline before editing and to attach verification evidence to completion claims. It installs through an agent prompt and a Python doctor script (scripts/aegis-doctor.py) on hosts such as Codex, OpenCode and Claude Code. It also provides optional TDD routing and an update command.
Strengths
Works across many hosts: Codex, OpenCode, Claude Code, Kimi Code CLI, Copilot and others
Publishes an A/B benchmark: contract pass rate 61.67% to 93.33% over 120 runs
TDD mode is off by default and can be set to auto via a doctor command
Install is checked by a doctor script that outputs JSON with ok and configStatus
Weaknesses
Benchmark covers one model setup (gpt-5.6-sol, xhigh); review was not independent human review
Only Codex and OpenCode have fresh host evidence; other hosts await release-level smoke tests
Skill triggering stays host-controlled and may need trigger-chain diagnosis
Not a runtime or gate: completion authority stays with the user and project rules
no GPU
Needs Python, AI coding host (Codex, OpenCode, Claude Code or similar)
Adds planner, engineer, reviewer and debugger roles to your existing coding agent
1.4k starsMITlast commit Oct 2026
oma is a CLI and portable .agents/ directory that adds 12 specialist roles, 21 workflows (/plan, /work, /orchestrate, /ultrawork, /ralph, /review, /debug) and 33 skills to coding agents such as Claude Code, Codex CLI and Cursor. Hooks keep long-running workflows going until your own test, lint or typecheck script passes. Steps write plans, results and review findings as files under .agents/ and docs/.
Strengths
One .agents/ directory is written into the config files of 14 agent runtimes
Stop gate sends the agent back to work until your test or lint script passes
Plans, ADRs, review findings and per-agent results are plain files you can open
MIT license; install via script, bun, npm or brew
Weaknesses
Installer writes outside the project: ~/.claude.json, ~/.serena, ~/.cursor and other vendor configs
Sets Claude Code defaults including skipDangerousModePermissionPrompt and telemetry opt-outs
Adds a co-author trailer and a commit-msg git hook unless disabled in config
Several workflows need the All skill preset; smaller presets lack /plan or /review skills
Hallmark is a skill for Claude Code, Codex, Cursor and OpenCode that builds self-contained HTML and CSS pages from one of 48 themes, or from a custom direction. It lints the result in a real browser at eight widths against a list of AI-design tells, then has a fresh-context subagent review it. It also offers redesign and audit commands for existing pages.
Claude Code plugin that learns, merges and prunes its own skills from sessions
10.7k starsMITlast commit Oct 2026
autoharness is a Claude Code plugin that distills skills from your real sessions into .claude/skills/, folds same-scenario skills together, and archives ones that stop being loaded. It only touches skills it wrote itself and injects a grouped index of them at session start. Reflection runs after a set number of tool calls (default 50), or on demand with /learn.
Strengths
Touches only skills it generated; hand-written and installed skills are left alone
Each create/update logs reason and evidence slice to a per-skill ledger
Archives rather than deletes; moving a skill directory back revives it
Pure Python with zero third-party dependencies; tunable via AUTOHARNESS_* env variables
Weaknesses
Requires python3 3.11+ on PATH; older interpreters turn the hooks off
Claude Code only; Linux and macOS only
README says default thresholds are placeholders pending empirical calibration
Background reflection spawns extra child sessions, adding token cost
Agent Skills for Vue, Nuxt, Vite, Vitest and pnpm, plus Anthony Fu's conventions
6k starsMITlast commit Oct 2026
A collection of Agent Skills (markdown files) covering Vue, Nuxt, Pinia, Vite, VitePress, Vitest, UnoCSS and pnpm, plus two hand-maintained skills for Anthony Fu's project conventions and PR creation. Most skills are generated from upstream documentation referenced as git submodules, and install through the skills CLI or the skills-npm pack. The README describes it as a proof of concept whose real-world performance is not fully tested.
Strengths
Generated skills reference upstream docs via git submodules, so they can be synced with source changes
Installs with one command via the skills CLI or as an npm dev dependency through skills-npm
Can be forked as a template to generate your own skills collection with pnpm scripts
MIT licensed; skills are plain markdown that any Agent Skills-compatible agent can read
Weaknesses
README states it is a proof of concept and skill effectiveness is not fully tested
Narrow coverage: eight generated skills, all in the Vue/Vite ecosystem, plus two opinionated ones
Skills load on demand, so agents may fail to pull them in when expected
No tagged releases
no GPU
Needs pnpm, skills CLI (vercel-labs/skills) or skills-npm, git submodules
An Agent Skill (build-teaql-app) that coding agents such as Claude Code, Codex, Cursor, Copilot or Gemini CLI follow to build apps model-first. The agent writes a domain model, has a TeaQL service evaluate and repair it, generates a typed library with Q query and E expression APIs, then implements and verifies the app. The README's example uses Rust and SQLite and ends with a work-complete.md evidence report.
Strengths
Works with several existing coding agents through one pasted prompt
Deterministic model evaluation with repair guidance runs before application code is written
Generated typed Q/E APIs; runtime logs SQL and audit context with query purpose and comment
Requires a work-complete.md report citing build, test and runtime log evidence
Weaknesses
Default workflow needs network access; model content is sent to TeaQL services
Evaluation, generation and Assist use hosted services, so the workflow is not offline
Ties the application to the TeaQL runtime and generated library
README gives no measured token savings; the context benefit is stated as an aim
no GPU
Needs TeaQL evaluation and generation services, Coding agent with file and terminal access, Language toolchain for the target app (e.g. Rust)
Models: Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI
MCP adapter for the Pi coding agent using one lazy proxy tool
1.6k starsMITlast commit Oct 2026
Extension for the Pi coding agent that exposes MCP servers through a single `mcp` proxy tool (about 200 tokens) instead of loading every tool definition. Servers start on first call and stop after 10 idle minutes by default, and tool metadata is cached so search works without live connections. It reads `.mcp.json`, imports Cursor, Claude Code, Codex and VS Code configs, and replaces Pi's built-in MCP.
Strengths
Lazy server start with 10-minute idle stop; README measured 0 of 100 servers running at session start
Agent skill that builds, tests and drives iOS apps in the Simulator
1.3k starsMITlast commit Oct 2026
A skill and plugin for Claude Code and Codex that wraps xcodebuild, xcrun simctl and idb in 30 Python scripts. Agents build and test Xcode projects, navigate the simulator through the accessibility tree rather than pixel taps, and run audits such as WCAG checks, visual diffs and hang capture. Output is kept to a few lines, with details fetched on demand by xcresult ID.
Strengths
Accessibility-tree navigation by text, type or id instead of fragile pixel coordinates
build_and_test.py returns a one-line summary and xcresult ID; errors and logs fetched on demand
30 scripts cover boot, privacy, push, location, status bar, a11y audit and localization
Installs as a plugin for both Claude Code and Codex; all limits tunable via IOS_SIM_* variables
Weaknesses
macOS only; requires Apple silicon, macOS 15+ and full Xcode 26+
Code graph and shared team learnings for Claude Code, Cursor and Codex
10.7k starsMITlast commit Oct 2026
Trail-code (formerly graft) parses a repo with tree-sitter into a graph of files, functions and callers, and gives coding agents matching files and edit impact through hooks and MCP tools. It also saves decisions and preferences from sessions as markdown in .trail/learnings/, shared through git, and trail check compares a branch against them before a PR. The README reports 33 of 50 SWE-bench Verified issues solved with the map versus 27 without.
Strengths
Code map built locally with tree-sitter, no model or API key, 23 languages
Learnings are plain markdown in git, reviewed in PRs, no server needed
trail check flags changes that contradict earlier team decisions before the PR
Claude Code wired via hooks; Cursor and Codex via MCP tools and instruction files
Weaknesses
Benchmark figures are the project's own; the SWE-bench sample is only 50 issues
Team workspace, skills and guardrails kinds are marked as coming soon
Day-one session mining reads Claude Code transcripts only, using Claude Sonnet
Collection of 175+ ready-made Codex subagent definitions as TOML files
6.4k starsMITlast commit Oct 2026
A library of Codex custom agent files in .toml format, grouped into 13 categories such as core development, language specialists, infrastructure, quality and security, and data and AI. You copy the files you want into ~/.codex/agents/ or .codex/agents/. Each file sets a model, reasoning effort, sandbox mode and instruction text. Codex does not spawn them automatically, so you delegate in prompts.
Strengths
Plain .toml files installed by copying into ~/.codex/agents/ or .codex/agents/
Each agent sets model, reasoning effort and sandbox mode (read-only or workspace-write)
Covers many languages and frameworks plus infrastructure, security and data roles
Project-level agents override global ones on name conflicts
Weaknesses
Codex does not auto-spawn custom subagents; you must delegate explicitly in prompts
No installer or package manager; installation is manual file copying
README gives no benchmarks of how the agents affect output quality
Model names in the files are fixed and may not match your Codex setup
Four-prompt workflow from app idea to PRD, tech design and agent files
3.2k starsMITlast commit Oct 2026
Four copy-paste prompts for a chat model: research the idea, write an MVP PRD, choose a tech approach, then generate AGENTS.md, MEMORY.md and a review checklist for a coding assistant. Each step saves a markdown document with a Handoff Context so a fresh chat can continue. An optional npx vibeworkflow CLI and Claude Code skills scaffold the same files.
Strengths
Prompts are self-contained; no install or CLI required to use them
Handoff Context in each document lets you switch chats or tools between steps
Part 4 emits AGENTS.md, MEMORY.md and tool files such as CLAUDE.md or Cursor rules
Includes a worked Reading List example and a larger Reddit to AI example
Weaknesses
Output quality depends on the chat model you run the prompts in
Research step needs a chat tool with browsing, otherwise findings stay unverified
Produces planning documents, not application code
Manual copy, answer and save loop across five steps
no GPU
Needs Chat model (ChatGPT, Claude or Gemini), AI coding assistant (Cursor, Claude Code or VS Code), Node.js for optional npx vibeworkflow CLI
Role-based agent personas and workflows for Claude Code, Codex and Gemini CLI
3.1k starscustom licenselast commit Aug 2026
AIOX is an npm CLI (`npx aiox-core init` or `install`) that writes role-based agents (analyst, pm, architect, sm, dev, qa and others) and workflows into a project. It configures Claude Code, Codex CLI, Gemini CLI, Cursor and Copilot with rules, skills and hooks. Planning agents produce PRD and architecture docs, and the sm agent turns them into story files for the dev agent.
Strengths
One npx command installs agents and config; `doctor` checks setup and `--dry-run` previews changes
Story files carry PRD and architecture context from planning agents to the dev agent
Sync and validation scripts for Claude Code, Codex CLI and Gemini CLI
Updates keep `.bak` copies of locally modified files
Weaknesses
Full hook automation only in Claude Code; Cursor and Copilot lack equivalent lifecycle hooks
License detected as NOASSERTION, though the README badge says MIT
Main README is in Portuguese; English is a separate translated file
No model runtime of its own; needs a supported IDE or CLI
no GPU
Compose
Needs Node.js >=18 (v20+ recommended), npm >=9, Claude Code, Codex CLI, Gemini CLI, Cursor or Copilot, GitHub CLI (optional)
Native macOS notch overlay that tracks and approves coding agent sessions
2.1k starsGPL-3.0last commit Oct 2026
Open Island is a SwiftUI and AppKit menu bar and notch app that shows live status for local coding agents, handles permission prompts, and jumps back to the right terminal or IDE. Agents report through hooks or plugins that forward JSON over a Unix socket to an in-app bridge. It lists 15 agents, including Claude Code, Codex, Cursor, Gemini CLI and OpenCode, and runs without a server or account.
Strengths
Covers 15 agents, including Claude Code, Codex, Cursor, Gemini CLI, Kimi CLI and OpenCode
Hooks fail open, so agents keep running if the app is not open
Jump-back works with Ghostty, iTerm2, tmux, Zellij, WezTerm, Warp and JetBrains IDEs
Distributed as notarized DMG or Homebrew cask, with Sparkle auto-update
Weaknesses
macOS 14+ only; building from source needs Swift 6.2 and Xcode
Grok Build and Gemini CLI events are observation-only, with no permission round-trip
Codex default install omits PreToolUse and PostToolUse; file-edit approval is not guaranteed
Claude Desktop usage panel only updates after a terminal claude session seeds the cache
no GPU
Needs macOS 14+, Swift 6.2 and Xcode (source builds)
Claude skill that rewrites vague requests into tool-specific prompts
15k starsMITlast commit Oct 2026
Prompt Master is a Claude skill that takes a plain request and returns a single copyable prompt tailored to a target tool such as Cursor, Claude Code, Midjourney or Stable Diffusion. It detects the target, extracts nine intent dimensions, asks at most three clarifying questions, picks one of 13 templates, and can prepend a Memory Block of prior decisions. Install by uploading the ZIP in claude.ai or cloning into ~/.claude/skills.
Strengths
Profiles for 30+ named tools, plus a four-question fallback for unlisted tools
13 prompt templates selected automatically, with a one-line strategy note per output
Asks at most three clarifying questions before writing the prompt
Installs as a skill folder via ZIP upload or git clone, no runtime needed
Weaknesses
Output quality depends on Claude; no tests or benchmarks are shown in the README
Memory Block must be re-supplied each new session; it does not persist
Tool profiles name fast-moving models, so they can go stale without updates
Six Agent Skills for design, video, image and retrieval tasks
13.2k starsMITlast commit Oct 2026
Garden Skills is a set of SKILL.md-format skills for Claude Code, Cursor, Codex and other coding agents. The six skills cover web video presentations (Vite, React, TypeScript, 23 themes, pluggable TTS), front-end design, GPT Image 2 prompting and generation, article formatting, code-rendered short video, and local knowledge retrieval. Install options include the skills CLI via npx, the Claude Code plugin marketplace, release zips, manual copy or a git submodule.
Strengths
Follows the SKILL.md spec, so it works across Claude Code, Cursor and Codex
Five install paths, including pinned per-skill release zips
web-design-engineer ships 25 style recipes and a design direction advisor
gpt-image-2 includes 79 prompt templates and three runtime modes
Weaknesses
Skills are narrow: design, video, images and articles, not general coding
gpt-image-2 local mode needs an image API; key setup not shown in the visible README
Built-in TTS providers are only MiniMax and OpenAI; others are paste-in snippets
README is truncated, so details of three skills (article, showreel, kb-retriever) are unknown
no GPU
Needs Node.js (npx), OpenAI-compatible image API (gpt-image-2), MiniMax mmx-cli or OpenAI TTS (optional)
Official Expo skills that teach coding agents Expo, EAS and React Native workflows
3k starsMITlast commit Oct 2026
A set of SKILL.md packages from the Expo team covering Expo Router, animation, native UI, SDK upgrades, native modules, and EAS build, update, hosting and workflow tasks. It installs as a Claude Code or Codex plugin, or through the skills CLI for Cursor, OpenCode, Copilot, Windsurf and others. The Claude Code and Codex plugin also bundles the Expo MCP server configuration.
Strengths
Maintained by the Expo team, with Expo docs and CLIs named as the source of truth
Installs via marketplace plugin for Claude Code and Codex, or the skills CLI for other agents
Skills are split into open-source framework and paid EAS groups, with cost notes
Plugin bundles Expo MCP config for live docs, EAS builds and simulator screenshots
Weaknesses
Only useful for Expo and React Native projects
Several skills (eas-*) target paid EAS services
Experimental skills ship in a separate expo-experiments plugin and may change or be retired
Telemetry exists, opt-in and Claude Code only, and adds a setting to manage
no GPU
Needs Node.js (npx), Expo CLI, EAS CLI, Expo MCP server (optional)
Agent Skills that teach coding agents Supabase and Postgres practices
2.9k starsMITlast commit Oct 2026
Two Agent Skills in the agentskills.io format: one covering Supabase Database, Auth, Edge Functions, Realtime, Storage, the CLI and the MCP server, and a Postgres performance skill with guidance across 8 prioritized categories. Install with `npx skills add supabase/agent-skills` or as Claude Code plugins. The README lists 18+ compatible agents, including Claude Code, Copilot, Cursor and Cline.
Strengths
Works across 18+ agents via the open Agent Skills format, not tied to one tool
Install all skills or one with a single npx command
Postgres skill ranks guidance by impact, from Critical to Low, across 8 categories
Also ships as Claude Code plugins through a marketplace
Weaknesses
Only two skills, both specific to Supabase and Postgres
Agent decides when to use a skill; README gives no way to force it
No benchmarks or measured accuracy gains in the README
Skill contents are not described in detail beyond topic lists
Claude Code plugin that turns your corrections into CLAUDE.md rules and skills
1.9k starsMITlast commit Oct 2026
claude-reflect is a Claude Code plugin that uses hooks to queue corrections you make during sessions, then writes approved ones to CLAUDE.md (global, project, subdirectory) and AGENTS.md via /reflect. /reflect-skills scans session history for repeated requests and drafts slash-command files in .claude/commands/. A separate reflect-mod variant checks each short prompt with a model and shows a save prompt in the session.
Strengths
Human review step: nothing is written to CLAUDE.md without approval
Syncs to CLAUDE.md, subdirectory CLAUDE.md files, skill files and AGENTS.md (Codex, Cursor, Aider)
/reflect --scan-history recovers corrections from past sessions; --dedupe merges similar entries
Runs on macOS, Linux and Windows with Python 3.6+; 332 tests reported
Weaknesses
Plugin capture uses regex; multi-language detection only happens at /reflect time
Depends on Claude Code session history, which is deleted after 30 days by default
Mod and plugin must not be installed together or corrections are captured twice
Mod is early access and needs Claude Code 2.1.286+ and a model call per prompt
no GPU
Needs Claude Code CLI, Python 3.6+
Models: Claude (haiku by default for reflect-mod checks)
CLI that proposes evidence-backed edits to AGENTS.md, CLAUDE.md and skills
1.4k starsMITlast commit Oct 2026
backpass reads agent session transcripts from disk for seven harnesses (Claude Code, Codex, Pi, OpenCode, Grok, Cursor CLI, Hermes), locally or over SSH. It distills them into proposed edits to your AGENTS.md or CLAUDE.md and project skills, each backed by verbatim quotes from at least two sessions. Analysis never writes; you accept or reject each edit in `backpass apply`.
Strengths
Every add, rewrite or remove edit needs quotes from at least two distinct sessions
Reads transcript stores directly from disk; no upload, and it has no API keys of its own
Supports project scope and user scope (--scope user) plus SSH collection from other machines
Analysis never writes files; changes only land after explicit review in apply
Weaknesses
Requires Node >= 22.5 and acpx on PATH; default review UI also needs lavish-axi
macOS and Linux only; Windows remotes are out of scope
User scope collects only Claude and Codex sessions by default
Model calls go through an already-authenticated harness, so usage costs are unknown
Agent skill that requires runnable acceptance gates before work is reported complete
4.1k starsMITlast commit Sep 2026
unlazy is a skill for Claude Code and Codex CLI that has the agent write an acceptance ledger (GATES.md) of CHECK/EXPECT shell commands, then runs them with a Node checker and reports only what passes. It also covers scoped multi-leaf orchestration with ownership leases, and an optional Claude Code Stop hook that blocks exit while gates are unmet. Requires Node 16+, no third-party runtime packages.
Strengths
Gates pass only on exit 0 plus EXPECT match, with definition-bound evidence
Checks need explicit per-command approval, stored outside the repository
No third-party runtime dependencies; Node 16 or newer, with a test suite
Installs through the skills CLI for several agents, plus manual Claude Code and Codex paths
Weaknesses
Proves only the command oracle you write; weak gates give false assurance
Approval is consent, not a sandbox; checks run with ambient filesystem and network access
Version 2.1.0 is unreleased; README advises pinning a commit
Earlier benchmark claims lack reproducible artifacts; the README calls them historical only
Audio analysis and sound synthesis for coding agents, as CLI, Python API and MCP
2.5k starsMITlast commit Oct 2026
earshot lets a coding agent hear audio: it describes files in words, matches them to text descriptions, finds moments in long recordings, and scores a film's sound against a profile of 7 launch films. It also synthesizes sounds from code recipes (`forge`), fits recordings to recipes (`fit`), and renders from a bank of 201 code-defined effects. It runs on CPU using CLAP and an AudioSet AST tagger.
Strengths
Exposes the same tools as CLI, Python API and MCP server for Claude Code
Runs on CPU only; models are pinned revisions with safetensors weights, loaded offline after first fetch
sfx and kit commands render from the 201-sound bank with no models or network
README states its limits, including a coin-flip blind test and Goodhart risk in the critic
Weaknesses
Each CLI call reloads models, about 20-40 s; MCP or Python needed for many calls
First use downloads about 1.1 GB of models
Forge reference recordings are licensed and not in the repo; --refset needs a local registry
No published blind-test score for the v2 sounds; the v1 test used one listener
no GPU
Needs Python 3.10+, ffmpeg, ffprobe, PyTorch (CPU), Hugging Face model download (first run)
SKILL.md collection turning business, UX and coding books into agent skills
2.5k starsMITlast commit Sep 2026
65 agent skills in the SKILL.md format: 51 frameworks drawn from books and style guides (Clean Code, Refactoring UI, Mom Test, DDIA and others) and 14 metaskills that walk through multi-phase journeys such as creating a website or removing technical debt. Metaskills ask decision questions per phase and store state in the project's docs/ folder. Installs via the Claude Code plugin marketplace, skills.sh, Codex, or Agent Plugins.
Strengths
Works with Claude Code, Codex, Cursor and other agentskills.io-compatible agents
Metaskills persist journey state in docs/, so work resumes across sessions
Install individually via npx skills add or as 10 plugin collections
Covers business, UX and marketing as well as code and architecture topics
Weaknesses
Skills are prompt content summarizing third-party books, with no executable tooling
No evaluation results given; output depends on the host agent and model
Many skills are non-coding, so a coding-only user needs a small subset
Install steps differ per client
no GPU
Needs Claude Code, Codex, Cursor or another agentskills.io-compatible agent
Oil UI is an installable skill for AI coding agents that generates several genuinely different design directions for a website, app, dashboard or component. It puts them on a comparison page, lets you pick one, then refines it against desktop and mobile screenshots. The free edition covers design direction, visuals and screenshot recreation; usability, polish and existing-project work are in the paid Pro edition.
Strengths
Installs with one command, npx skills add oil-oil/oil-ui, no configuration
Built-in comparison page with desktop/mobile switching and arrow-key browsing
Directions must differ in layout, typography, color and imagery, sharing at most one
Checks results against real screenshots at desktop and mobile sizes
Weaknesses
Usability, polish, dashboard and existing-project workflows are only in Pro ($9.99)
Independent review is limited to a score with issues listed in the free edition
Checks ui.oiloil.org for updates on use; needs Python 3 (can be disabled)
Installing alongside Pro makes both compete for the same requests
no GPU
Needs Python 3, npx (skills CLI), an AI coding agent that supports skills
CLI tools, skills, hooks and plugins that extend Claude Code and Codex
2k starsMITlast commit Oct 2026
A collection of add-ons for Claude Code and other coding agents, published as the claude-code-tools PyPI package and the aichat-search crate. It includes session search (aichat), tmux-based agent control (tmux-cli), safety hooks, a statusline, session repair, Claude/Codex session porting and voice input (voxtype). Skills install with `npx skills add`, which needs Node.js 22.20 or newer.
Git-versioned per-file codebase index that coding agents read and update via MCP
1.7k starscustom licenselast commit Oct 2026
AOCI-CODE is a Go CLI (`aoci`) and local stdio MCP server that has your coding agent write one plain-text entry per file (responsibility, dependencies, public contracts, constraints) into `aoci.code.txt` in the repo. Agents read the index before a task instead of re-reading the codebase, and the server flags entries to update after code changes. It can also index MySQL, PostgreSQL and openGauss table structure.
Strengths
Index is plain text in the repo, so Git diffs, reviews and rollbacks work
Works with any host supporting stdio MCP: Codex, Claude Code, Cursor, OpenCode
Local only: reads source and DB schema, no Internet access, no stored credentials
Optional table-level index for MySQL 5.7.6+, PostgreSQL and openGauss 6.0.5
Weaknesses
First index is slow: about an hour per 200,000 lines of code
Release candidate (v0.1.0-rc21); no stable release published
FSL-1.1-MIT source-available license, not a standard OSI open-source license
Index quality depends on the agent's model; MariaDB is not supported for database indexing
no GPU
Needs Git, MCP-capable coding agent host, Go toolchain and make (source builds only), MySQL, PostgreSQL or openGauss (optional, for database index)
Models: Any model via MCP host (Codex, Claude Code, Cursor, OpenCode)
Skills that let coding agents review KiCad schematics, PCBs and Gerbers
1.5k starsMITlast commit Oct 2026
A set of 11 agent skills plus pure-Python scripts that parse KiCad schematics, PCB layouts and Gerbers into structured data. They compute regulator outputs, filters and derating, run SPICE testbenches, and check EMC with 44 rules. It installs as a plugin for Claude Code, Codex, Copilot CLI, Antigravity and opencode, and also runs as a GitHub Action for PR reviews.
Strengths
Analysis scripts are pure Python 3.10+ with no required dependencies or KiCad install
Works across five agent CLIs and as a GitHub Action for PR checks
Covers power tree, ESD, thermal, EMC (44 rules) and SPICE with Monte Carlo
Includes supplier skills for DigiKey, Mouser, LCSC, element14, JLCPCB and PCBWay
Weaknesses
SPICE simulation needs ngspice, LTspice or Xyce; it is skipped without one
Supplier skills (e.g. DigiKey) use APIs, which likely need credentials; README excerpt is silent
Claude Code /plugin update may miss new versions; cache must be cleared manually
Only supports KiCad projects, not other EDA tools
no GPU
Needs Python 3.10+, ngspice, LTspice or Xyce (optional, for SPICE)
Models: Claude Code, OpenAI Codex, GitHub Copilot CLI, Google Antigravity, opencode
CLI that generates and syncs context files across coding agents
1.3k starsMITlast commit Sep 2026
Caliber is a Node CLI that writes and maintains CLAUDE.md, Cursor rules, AGENTS.md and Copilot instructions for a repo. A pre-commit hook refreshes them from your diff, and `caliber sync` mirrors skills, rules, plugins and MCP config between agents. It also ships a Claude Code plugin and a CLI command for Jev compaction, which replaces `/compact` with per-tool-call scoring.
Strengths
`caliber score` audits context files against the repo without an LLM
Sync is deterministic and never overwrites files you edited by hand
Every write is shown as a diff and backed up to .caliber/backups/
Covers Claude Code, Cursor, Codex, OpenCode and GitHub Copilot
Weaknesses
Jev compaction needs a Vercel AI Gateway or TypeSafe API key
Anonymous usage analytics are on by default; opt out via env var
Setup and refresh need an agent seat or your own LLM API key
On agents lacking a feature, sync degrades, e.g. skills become Copilot instruction files
no GPU
Needs Node 20+, Claude Code or Cursor seat, or Anthropic/OpenAI/MiniMax/Vertex key, Vercel AI Gateway or TypeSafe key (Jev compaction only)
Agent-ready repo map, explicit skills and a safe updater for existing repositories
1.3k starsMITlast commit Oct 2026
repository-harness installs a compact AGENTS.md entrypoint, a documentation map, optional plan and decision templates, and explicit-only skills into an existing repository. A Rust `harness` binary handles status, doctor and update, using a three-way merge against the upstream base stored in `.harness-core/`. It is not a task tracker, orchestrator or runtime.
Strengths
Updater does three-way merge, backs up files and changes nothing on conflict
Skills run only when explicitly invoked; installation activates nothing
Installs no database, credentials or background processes
Installer offers dry-run and merge modes, with Bash and PowerShell scripts
Weaknesses
Installer is run via curl piped to bash
Does not supply validation commands, runtimes or fixtures for your application
Legacy SQLite harness-cli and protocol v1 ended support on 2026-08-10
Overlapping local and upstream edits need manual resolution before update continues
Go CLI harness that wraps Claude Code with SPEC workflow and verified completion claims
1.2k starsApache-2.0last commit Oct 2026
MoAI-ADK is a Go CLI that wraps Claude Code with a plan, run, sync SPEC lifecycle and binds completion claims to the command actually run and its output. Factory Mode (v3.2) splits work across one leader session and numbered lane sessions, each carrying a card through its stages in its own worktree. It also offers a local web console, a SQLite-backed task queue, and launchers for GLM and Codex backends.
Strengths
Completion claims must cite the command run and its output
Per-SPEC git worktrees and a branch guard keep parallel agents from colliding
Factory Mode spreads cards across lanes on Claude, GLM or Codex backends
Progress files and memory let work survive /clear
Weaknesses
Lanes must be launched by hand, one terminal each
Built around Claude Code; Codex can only join as a lane
MOAI_HOME is read only by Go processes, so shell hooks and statusline can split state
Large command and concept surface (SPEC tiers, cards, lanes, runs) to learn
Harness is a Claude Code plugin that takes a domain description and writes agent definitions into .claude/agents/, skills into .claude/skills/, and a minimal CLAUDE.md pointer. It picks one of three execution modes (workflow scripts, persistent named agents with SendMessage, or one-shot sub-agents) and one of six team patterns. A /harness:evolve skill feeds usage feedback back into the generated harness.
Strengths
Generates plain files in .claude/agents and .claude/skills that you can edit and commit
Covers six team patterns, each mapped to one of three execution modes
Includes /harness:evolve and a mechanical migration path from v1 artifacts
No environment variables or experimental flags required
Weaknesses
Works only with Claude Code; no other agent host is mentioned
Quality claims come from an author-run A/B test on 15 tasks
Interactive walkthrough is in Korean only
No release published; version 2.1.0 appears only in a badge
Markdown-file long-term memory for coding agents, with local SQLite search
4.7k starsMITlast commit Oct 2026
agent-memory stores agent memories as markdown files in one store, with a rebuildable SQLite FTS5 index (BM25, optional vector plugin fused by RRF) for local ranked recall. Hooks installed for Claude Code, Codex CLI and Muse Code distil memories at conversation boundaries, and a sleep-time pass consolidates them, with deletions only as proposals. It exposes a `mem` CLI and a `mem-mcp` stdio MCP server.
Strengths
Files are the source of truth; `rm -rf .index/ && mem rebuild` loses nothing, enforced by a test
No LLM client in the library; reasoning borrows the host agent's CLI, so no API key required
One store shared across Claude Code, Codex CLI and Muse Code; any agent that runs shell commands works
Recall returns paths and abstracts, so the agent opens files only as deep as needed
Weaknesses
No PyPI release; install is from a git checkout with uv sync
Benchmark is on a bounded LongMemEval-S haystack, so scores are not comparable to published ones
`context`, `sleep` and the proposal ledger have no MCP tool yet; CLI only
Muse Code setup needs OpenRouter credentials and a separate pproxy dependency
Multi-persona deliberation skill for coding agent CLIs
4.7k starsMITlast commit Oct 2026
Adds a /council command that sends a decision to up to 18 analytical personas, such as Socrates, Torvalds and Taleb. Members analyze independently, cross-examine each other, then produce a verdict listing dissent, kill criteria and a next step. Seats can be spread across providers (native host, Codex, Gemini, Ollama, NVIDIA NIM, Cursor) so one model family does not argue both sides.
Strengths
Full, quick and duo modes plus named triads for narrower questions
Same /council command across Claude Code, Codex, Gemini CLI and OpenCode
Verdicts keep dissent and unresolved questions, with FACT/INFERENCE/ASSUMPTION/UNKNOWN labels
Provider routing with --dry-route preview and fallback to the native host
Weaknesses
No benchmarks or evals in the README; output quality depends on the underlying models
Full protocol runs several rounds across many personas, so token cost is likely high
Personas are prompt contracts, not verified expert models
Four host mirrors of SKILL.md must be kept aligned by hand
no GPU
Needs Claude Code, Codex, Gemini CLI or OpenCode, Optional: codex, gemini, ollama, cursor-agent executables or NVIDIA_API_KEY
Models: Claude Code native subagents, OpenAI via codex, Google via gemini, Ollama, NVIDIA NIM
74 game-dev Agent Skills with a router that loads the right ones per request
3.7k starsApache-2.0last commit Oct 2026
A collection of 74 SKILL.md files covering Godot, Unity, Unreal, Phaser, PixiJS, three.js, Bevy, pygame, LÖVE and Roblox, plus a router skill. The router detects the engine from project files (for example project.godot or *.uproject), reads the task, and loads only matching engine, discipline, genre and workflow skills. Installs with `npx skills add` or as a Claude Code plugin marketplace.
Strengths
Router picks skills from engine detection, so users never name skills
Covers 10 engines, each pinned to a stated version (e.g. Godot 4.7, Unity 6.3 LTS)
Plain SKILL.md files load in Claude Code, Cursor, Windsurf, Cline, Codex, Gemini CLI and others
Includes an asset art-direction workflow and a validator script for skill quality
Weaknesses
Skill depth is uneven: Unreal has 6 skills, Godot 16
Version pins will go stale as engines release; upgrade policy lives in separate docs
Demo in the README is illustrative only, not a recorded run
Only 9 genres and 4 workflows are covered
no GPU
Needs Node.js (npx skills CLI), an AI coding agent that supports Agent Skills
Token-saving extension for the Pi coding agent with four opt-in mechanisms
3.6k starsMITlast commit Oct 2026
SoL-Pi is an NVIDIA extension for the Pi coding agent that cuts token use and agent turns. It adds Action Fusion (edit plus validation in one tool call), ObservationPack (handles for repeated large outputs), an evidence-checked log reducer, and context compaction at completed plan steps. It installs on unmodified Pi 0.85.1 and every mechanism is off until enabled in sol-pi.json.
Strengths
Installs through Pi's public extension APIs; no Pi patches or vendored source
All mechanisms disabled by default; config is a single sol-pi.json
Reducer keeps a quotation only if it matches the archived source log
Includes a skill example for calling it from Claude Code or Codex
Weaknesses
Tested only against Pi 0.85.1; pi-v1 support still in development
Archives persist locally and are not deleted when the session ends
Reducer sends diagnostic log content to a remote model when enabled
Prompts and templates that set up machines and projects for coding agents
2k starsMITlast commit Oct 2026
A set of markdown prompts, docs and templates that you paste into a coding agent to configure a dev machine, bootstrap a new Next.js project, or upgrade an existing one. The agent installs and wires tools such as Superpowers, Context7, RTK and Vercel skills, writes AGENTS.md and project docs, and runs checks. No application code lives here; it is a recipe the agent executes.
Strengths
Separate prompts for machine setup, new-project bootstrap and existing-project adoption
Separate guides for Claude Code and Codex, plus shared AGENTS.md template
Prompts can be adopted one at a time, e.g. only the ESLint quality gates
Validation documented on macOS, Linux and Windows; CI runs on Ubuntu and Windows
Weaknesses
Needs an agent with file and terminal access; plain chat installs nothing
New-project path is Next.js only, with Drizzle and PostgreSQL defaults
Remote prompts are fetched from main, so results shift as the repo changes
No tagged releases; installed tools depend on third-party projects and upstream docs
no GPU
Needs Claude Code or Codex, Node.js, Next.js, PostgreSQL
Open Markdown format for giving coding agents project instructions
25.1k starsMITlast commit Sep 2026
AGENTS.md is a specification for a plain Markdown file kept in a repo that holds dev environment, testing and PR instructions for AI coding agents. The repository also contains the Next.js site served at agents.md, which explains the format and lists examples. It has no runtime or agent code.
Strengths
Plain Markdown, no tooling, schema or dependencies required
One predictable file location for agent instructions, separate from the human README
Website runs locally with pnpm install and pnpm run dev
Weaknesses
No enforced schema; README shows only a minimal example
Repo holds the website and spec, not a tool or agent
SwiftUI Pro is a skill in the Agent Skills format that gives coding assistants rules for writing and reviewing SwiftUI. The rules cover API usage, navigation, layout, animations, state management, performance, VoiceOver accessibility and deprecated API, aimed at mistakes LLMs commonly make. It installs through npx or as a Claude Code plugin and is invoked with /swiftui-pro or $swiftui-pro.
Strengths
Works with Claude Code, Codex, Gemini, Cursor and others via the Agent Skills format
Installs with one npx command or as a Claude Code plugin
Can run a partial review, such as deprecated API only or accessibility only
Contribution rules keep SKILL.md short to limit token cost
Weaknesses
Covers SwiftUI only; SwiftData, concurrency and testing are separate skills
Targets iOS 26+ and Swift 6.4+, so older projects may get mismatched advice
Install via npx requires Node, which the README installs through Homebrew
No benchmarks or measured results are given for the rules
no GPU
Needs Node.js (npx) for install, Claude Code, Codex, Gemini, Cursor or another Agent Skills-compatible agent
Agent skills that teach coding assistants current Gemini API and SDK usage
4.3k starsApache-2.0last commit Oct 2026
A set of three skills that give coding agents up-to-date context on the Gemini API: gemini-api-dev (Interactions API, Python and TypeScript SDKs), gemini-live-api-dev (realtime streaming), and gemini-omni-flash-api (video generation and editing). The README reports that adding the skill raised correct API code generation to 87% with Gemini 3 Flash and 96% with Gemini 3.1 Pro. It installs via the Vercel or Context7 skills CLIs, or as a plugin for Claude Code, Cursor, Codex and Antigravity.
Strengths
Installs through several routes: skills CLIs, Claude Code, Cursor, Codex and Antigravity plugins
Covers Interactions API, Live API and Omni Flash video in separate, individually installable skills
README cites eval results: 87% with Gemini 3 Flash, 96% with Gemini 3.1 Pro
Apache-2.0 licensed and maintained by the google-gemini organization
Weaknesses
Only helps with Gemini API code; not useful for other model providers
README states it is not an officially supported Google product
The vertex-ai-api-dev skill moved to a different repository
No tagged releases, so version pinning relies on commits
Agent Skills package for CTF challenge work in coding agents
3.5k starsMITlast commit Oct 2026
A set of Agent Skills for CTF competitions, split by category: web, pwn, crypto, reverse engineering, forensics, OSINT, malware, AI/ML and misc. It installs with `npx skills add ljagiello/ctf-skills` and works with Claude Code and other tools that follow the Agent Skills spec. It also ships an orchestrator skill, a writeup generator, and an installer script for the tools each category needs.
Strengths
Skills are split by CTF category, so the agent loads only what it needs
Installer script supports dry-run, verify, force and per-group modes
Includes an orchestrator skill and a writeup generator skill
Each skill lists its own prerequisites for on-demand installs
Weaknesses
Needs many external tools installed via apt, brew, gems, go or manual steps
Similar-writeup search skill relies on the Tavily API
No tagged release
README gives no benchmarks or solve-rate results
no GPU
Needs Agent Skills-compatible agent such as Claude Code, npx, Tavily API (ctf-similar-search skill)
Local project memory and code index shared across coding agents
2.4k starsAGPL-3.0last commit Oct 2026
OpenWolf creates a .wolf/ folder in your project holding task checkpoints, a file and symbol index, bug-fix records and session notes. It registers hooks for Claude Code and Codex CLI and a plugin for OpenCode, restoring saved context after compaction or restart. It also reports recorded token usage and API-price cost estimates by agent and model in a local dashboard.
Strengths
Memory, indexing and usage processing run locally with no extra model calls
Covers Claude Code, Codex CLI and OpenCode, with context-only files for Cursor, Gemini CLI, Antigravity
Usage report keeps missing counters as unavailable instead of reporting zero
Dashboard binds to 127.0.0.1 and requires a project token
Weaknesses
Hook coverage varies by agent and version; Bash output shortening is Claude-only
Full live Claude to Codex handover is listed as unverified in the README
AGPL-3.0 license may restrict some commercial use
Cost figures are API list-price estimates, not subscription charges
no GPU
Needs Node.js >=20, At least one supported coding agent
Models: Claude Code, Codex CLI, OpenCode, Grok Build, Cursor
Agent skill that teaches coding agents to build and deploy on EdgeOne Makers
1.8k starsMITlast commit Oct 2026
Installs a single skill, edgeone-makers-tools, into Claude Code, Cursor, Codex or CodeBuddy. Its SKILL.md routes to reference docs on demand, covering deployment, edge and cloud functions, KV and Blob storage, middleware, the CLI and agent frameworks (LangGraph, Claude SDK, OpenAI Agents, DeepAgents, CrewAI). Hooks inject the relevant skill on prompt submit and before tool use.
Strengths
One skill in the agent's list; capabilities load from references/ only when needed
Installs via npx skills, Claude Code marketplace or CodeBuddy marketplace
Plugin manifests for Claude Code, Cursor and CodeBuddy in one repo
Covers Node.js, Go and Python functions plus five agent frameworks
Weaknesses
Useful only for deploying to the EdgeOne Makers platform
Requires an EdgeOne account (China or Global site)
No tagged releases
README gives no tested agent versions or hook behavior details
CLI that generates config files for many AI coding tools from one rule set
1.6k starsMITlast commit Oct 2026
Rulesync is a Node.js CLI that keeps AI tool configuration in a single `.rulesync/` source directory and generates the files each tool expects. It covers rules, ignore files, MCP config, commands, subagents, skills, hooks and permissions for dozens of tools, including Claude Code, Cursor, Copilot, Codex CLI and Cline. It can also import existing configs or convert directly between tool formats.
Strengths
One source directory generates configs for 60+ listed tools, so rules are not duplicated
`import` pulls existing CLAUDE.md, .cursorrules or Copilot files into .rulesync/
`convert --from cursor --to copilot,claudecode` translates formats without .rulesync/ files
Installs via npm, Homebrew or a single binary
Weaknesses
Feature support varies per tool; many targets lack ignore, hooks or checks
Several targets are deprecated or end of life (Roo Code, Continue, Tabnine legacy)
The `ignore` feature is deprecated in favor of `permissions`
Plugin-packaging targets are excluded from `--targets "*"` and need explicit output roots
Coding-agent skill that runs a multi-phase security audit of a codebase
42.7k starsMITlast commit Sep 2026
A skill for coding agents that audits a repository for vulnerabilities in six phases: reconnaissance, hunting, candidate validation, structured output, record verification and reporting. Isolated sub-agents hunt from a coverage ledger, and a fresh agent tries to disprove each candidate. Results go to findings.json, checked by Node validators, and to REPORT.md. It is the single-repo starting point for Cloudflare's vulnerability harness.
Strengths
Separate agents find and validate each finding; verdicts are confirmed, needs_validation or rejected
Zero-dependency Node validators check findings.json and coverage-ledger.json against a schema
Agent Skills for Hugging Face Hub, training, evals and SageMaker deployment
11.2k starsApache-2.0last commit Oct 2026
A collection of about 24 Agent Skills folders (SKILL.md plus scripts) that teach coding agents Hugging Face workflows: Hub CLI operations, TRL and Unsloth fine-tuning, evals, Spaces, Gradio, Trackio and SageMaker deployment. Claude Code and Cursor marketplaces expose only hf-cli; other skills install with `hf skills add <name>`. A Gemini CLI extension and a generated AGENTS.md fallback are included.
Strengths
Follows the Agent Skills format; works with Claude Code, Codex, Gemini CLI and Cursor
hf-cli skill is generated from the locally installed CLI, so commands stay current
Covers datasets, training, evals, Spaces, papers and SageMaker deployment
AGENTS.md fallback bundle for agents without skill support
Weaknesses
Client marketplaces list only hf-cli; other skills need the hf CLI to install
Codex install is a manual copy or symlink into .agents/skills
Training skills target Hugging Face Jobs, a hosted service
Some skills overlap, such as huggingface-llm-trainer and trl-training
Temporal, access-controlled memory store for LLM agents, backed by SQLite
4.5k starsApache-2.0last commit Oct 2026
LongMemory is a TypeScript memory engine that ingests events, documents and coding-agent sessions into an immutable graph stored in SQLite. It offers strict, historical, associative and world-grounded recall with separate recorded and valid time. The engine is exposed as an npm library, CLI, HTTP API, MCP server, Next.js dashboard and VS Code extension, plus a zero-dependency Python HTTP client.
Strengths
Recorded time and valid time are separate, so historical recall returns superseded facts
Runs in-memory or on SQLite with no external database; reopening restores state
One engine across npm, CLI, HTTP, MCP (13 tools), dashboard and VS Code
Imports sessions from Claude Code, Codex, OpenCode, Gemini CLI, Copilot Chat and Cline
Weaknesses
Python package is only an HTTP client; the engine needs the TypeScript service
Last tagged release is 2025-12-12, well behind recent commits
Embedding providers must be configured separately (OpenAI-compatible, Gemini, Bedrock, Ollama)
Hosted deployments require setting an API key and mounting persistent storage at /data
Claude Code plugin that makes agents write current Go idioms
3.9k starsApache-2.0last commit Sep 2026
Guidelines covering Go 1.0 through 1.26 that tell a coding agent which language and stdlib features to use, such as max, slices.Contains, cmp.Or, new(42) and errors.AsType. The Claude Code plugin adds a /use-modern-go command that reads the Go version from go.mod and limits the agent to features up to that version. Junie includes the guidelines by default from version 2xx.620.xx.
Strengths
Reads the target Go version from go.mod and limits features to it
Covers Go 1.0 to 1.26, including everything the modernize analyzer targets
Installs in two commands as a Claude Code plugin
Built into Junie with a settings toggle, no manual setup
Weaknesses
Supports only Junie and Claude Code; no Cursor or Codex setup documented
Claude Code needs /use-modern-go run manually at the start of each session
Guidelines only; does not rewrite existing code, unlike the modernize analyzer
Marketplace of Claude Code plugins and agent skills with a ccpi installer CLI
2.9k starsMITlast commit Oct 2026
A catalog of 434 plugins and about 2,900 skills across 19 categories (DevOps, security, testing, SaaS skill packs, MCP servers and others), installed in Claude Code with /plugin marketplace add or the ccpi npm CLI. Claude Code is the only verified-native harness. Mirrored third-party plugins keep their upstream license and are pinned by commit.
Strengths
Large catalog: 434 plugins, about 2,900 skills, 352 agent definitions, with reproducible count commands
Installs with one /plugin command or `ccpi install <pack>` from npm
Mirrored plugins record upstream author, license and pinned commit
Browsable web catalog with search and downloadable bundles
Weaknesses
Only Claude Code is verified-native; other harnesses are unverified candidates
Certification program is not implemented; no artifact claims a quality tier
Each plugin has its own license; MIT covers only scaffolding and tooling
Repo development needs Node >=22.12.0 and pnpm >=9.15.9
no GPU
Needs Claude Code, Node.js >=22.12.0, pnpm >=9.15.9
Plugin marketplace of context-engineering skills and commands for coding agents
1.8k starsGPL-3.0last commit Aug 2026
A marketplace of plugins for Claude Code, also installable on Gemini CLI, Antigravity, Cursor, Codex and OpenCode. Plugins cover reflection loops (/reflect, /memorize), spec-driven and subagent-driven development with judge agents, multi-agent PR review, git commands, TDD and language-specific rules. Each plugin loads only its own agents, commands and skills to limit token use.
Strengths
Granular install: each Claude Code plugin loads only its own agents, commands and skills
Skills follow the agentskills.io format and install on several agents via npx skills
Review plugin runs specialized agents and has a documented GitHub Actions integration
Covers reflection, spec-driven and subagent-driven workflows in one marketplace
Weaknesses
Gemini CLI and Antigravity install every plugin as one bundle, no per-plugin selection
npx skills install does not support subagents, so non-Claude setups get a reduced experience
Reliability percentages come from the authors' own usage, not a published benchmark
CLI that configures, supervises and shares memory across coding agents
1.7k starsApache-2.0last commit Oct 2026
AI DevKit is an npm CLI that writes one .ai-devkit.json and generates setup (rules, MCP servers, skills) for coding agents such as Claude Code, Codex CLI, Gemini CLI, Cursor and opencode. It adds a TUI console for running sessions, an `agent send` command that routes prompts or stdin to live agents, and a local SQLite memory exposed over MCP and CLI. It also installs skills such as dev-lifecycle, verify and tdd.
Strengths
One .ai-devkit.json reconciles setup for 16 listed coding agents
`agent console`, `agent list` and `agent send` supervise and message running sessions
Memory is a local SQLite file, searched via MCP or CLI on demand
Runs locally; README states no telemetry and writes project files you can commit
Weaknesses
Remote control is unavailable for several agents, including Cursor, Copilot, Cline and Devin
Roo Code setup and opencode remote control are marked as testing
Phase-doc workflow (requirements to testing) adds process overhead to small changes
Does not improve model quality; README says it only raises process discipline
Output styles for Claude Code that make replies shorter and answer-first
1.5k starsAGPL-3.0last commit Oct 2026
Three markdown output styles for Claude Code (Attention-kind, Spartan, Rundown) that change how Claude words its replies, not how it codes. It ships as a plugin with a /style switcher and user-invoked skills. The README's benchmark reports about 43% shorter output and equal 97% pass rates on 12 coding tasks.
Strengths
Each style is one markdown file; plugin install is a single /plugin command
Hidden-test benchmark: 97% pass rate both with and without the style
Reported ~43% shorter output; about 650 input tokens, cached per session
Style bodies reusable in Codex AGENTS.md or GEMINI.md via a sed strip
Weaknesses
Benchmark covers only 12 coding tasks; README does not claim better answers
Built for Claude Code; other agents need a manual sed-based install
AGPL-3.0 license may matter if you redistribute modified versions
Rundown favors scannability over brevity and uses emoji checklists
no GPU
Needs Claude Code
Models: Claude Code, Codex, Devin, Antigravity CLI
Desktop app for browsing and installing AI agent skills across many agents
1.4k starsMITlast commit Oct 2026
SkillsGate is an Electron desktop app for macOS, Windows and Linux that browses the public skills.sh catalog and installs skills into the config of agents such as Claude Code, Cursor, Codex CLI and Cline. It includes a CodeMirror editor for skill files, SSH access to remote machines, and keeps settings and favorites in a local SQLite database with no account.
Strengths
Installs or removes a skill per agent, across 25 listed agents
Local-first: settings and favorites in SQLite, no account required
Built-in editor writes skill source straight to disk
Syncs and browses skills on remote machines over SSH
Weaknesses
macOS build is signed but not notarized; first launch needs Open Anyway
Terminal UI and npm packages are deprecated and unmaintained
macOS 0.7.0 and earlier need a manual update to 0.7.2
Project memory and decision records for AI coding agents, via MCP and skills
1.4k starscustom licenselast commit Sep 2026
Haft is a local Go CLI and MCP server (`haft serve`) that gives coding agents a typed project memory: problems, compared options, human decisions, evidence and stale-claim tracking. It builds on the First Principles Framework, bundling the versioned FPF source and installing skills such as /h-frame, /h-compare and /h-decide into Claude Code, Codex and other hosts. State lives in `.haft/` markdown plus a per-project SQLite database.
Strengths
MCP server validates required fields, parity gaps and missing evidence server-side, not only in prompts
`haft init` wires MCP config, skills and instruction sections for Claude Code and Codex
Local-first: markdown in `.haft/` and SQLite under `~/.haft/projects/`, no hosted service
Flags stale evidence and spec drift; binding decisions require the human
Weaknesses
Requires learning FPF concepts; the README itself calls FPF complex
Only Claude Code and Codex are stable; other hosts are experimental or legacy
License is NOASSERTION in detected metadata, so terms need checking
Schema migrations and carrier semantics add operational overhead on upgrade
no GPU
Needs SQLite, Claude Code, Codex or another MCP host, haft-embed sidecar (optional)
Git merge driver that merges functions and classes instead of lines
1.3k starsApache-2.0last commit Oct 2026
Weave registers as a Git merge driver and parses base, ours and theirs with tree-sitter, then merges at the level of functions, classes and keys. Independent additions to the same file resolve without conflict, and real collisions show the entity and the guard that refused. It also has `weave land` for verified merges by agents, an MCP server, and jj support.
Strengths
Entity-level 3-way merge across 38 languages, with line-merge fallback for the rest
Conflict markers name the entity and the guard that refused (`refused_by:`)
Reproducible benchmarks: `weave stats --bench` and replay of real merge commits
Setup per repo, per user or globally; `weave setup --off` reverts
Weaknesses
Replay on real repos showed 86 regressions where Git merged cleanly and weave conflicted
Human-match rate on wins is 33% on Flask and Go
Vue, Svelte, ERB and Haskell fall back to Git's line merge
Source install needs two binaries, `weave` and `weave-driver`, both on PATH
Agent skill that plans, approves and deploys an app to your own accounts
1.3k starsMITlast commit Oct 2026
GoLive is an Agent Skill for Codex and Claude Code that detects what an app needs, builds a plan, and applies it to your own Vercel, Netlify, Supabase, Neon, Porkbun, GoDaddy, Resend and Stripe accounts only after approval. It verifies results, checks drift with `golive status`, and removes what it created with `golive teardown`. It is early alpha (0.1.0-alpha.9).
Strengths
Every write needs an approved plan ID and --yes; DNS, destroy and live steps need extra flags
Teardown removes only resources it can prove it created and lists leftovers as handoffs
Validation doc separates live-tested journeys from mock-covered ones
No GoLive account, hosted backend or product telemetry
Weaknesses
Early alpha; built-in adapters cover only Vercel, Netlify, Supabase and Neon for hosting and data
Install verified only for Codex and Claude Code; other clients unverified
Credentials stored in a plaintext 0600 file, not a keychain
Rollback is Netlify-only and not live-validated; other providers get guided steps only
no GPU
Needs Node.js 20+, Git, Codex or Claude Code, Vercel or Netlify account, Supabase or Neon account
Claude Code plugin that prunes stale tool calls instead of summarizing history
7.8k starsMITlast commit Sep 2026
Replaces Claude Code's compaction summary with a pruned copy of the original transcript. Each tool call and result is scored by the Jev model through TypeSafe's API, and stale ones are dropped or truncated while user and assistant text stays verbatim. It also works as an npm library, and it falls back to the built-in summary on errors or small reductions.
Strengths
Never rewrites content: kept messages stay verbatim, only tool calls and results are removed
Falls back to Claude Code's built-in summary on errors or insufficient reduction
Also exported as an npm library with a pluggable JevAsker transport and building blocks
Tunable keepThreshold, preserveRecentMessages, token ceilings and truncateHeadChars
Weaknesses
Requires a TypeSafe API key and sends the conversation to an external Jev endpoint
Needs Claude Code 2.1.274+ with the early-access function hooks flag enabled
Text messages are never shortened, so long text-heavy sessions may reduce little
Token counts are estimates, and histories that cannot fit the state ceiling throw
no GPU
Needs TypeSafe API key (Jev model, api.typesafe.ai), Claude Code 2.1.274+ with CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1, Node.js and npm
Catalog of NVIDIA-maintained agent skills for Claude Code, Codex, Cursor and Kiro
3.7k starsApache-2.0last commit Oct 2026
A mirror of agent skills maintained in NVIDIA product repos, synced daily and installed with `npx skills add nvidia/skills`. The catalog covers cuOpt, cuDF, CUDA-Q, DOCA, Jetson BSP, Holoscan, BioNeMo, Dynamo, DeepStream and others, with many skills per product. Skills can be installed singly or per agent (claude-code, codex, cursor, kiro-cli, cortex).
Strengths
Install one skill or several agents in a single `npx skills add` command
Skills are maintained in product repos and synced daily, so they track upstream
Wide coverage of NVIDIA software: cuOpt, cuDF, DOCA, Jetson, BioNeMo, Dynamo
Targets several agents: Claude Code, Codex, Cursor, Kiro, Snowflake CoCo
Weaknesses
Only useful for NVIDIA software and hardware workflows
Mirror only; fixes must go to the upstream product repos
Requires skills CLI v1.5.16 or newer, or skills may not appear in Claude Code
Skills are renamed or consolidated over time, so local copies need `npx skills update`
Starter Claude Code config with CLAUDE.md memory bank templates and hooks
2.7k starsMITlast commit Oct 2026
A copy-into-your-project set of Claude Code settings: a CLAUDE.md memory bank that keeps context across sessions, three CLAUDE.md templates, hooks, slash commands and a statusline script. The README also lists MCP servers, plugin installs and git worktree launchers for parallel Claude Code and Codex CLI sessions.
Strengths
Three CLAUDE.md templates (~101 to ~153 lines) plus a migration prompt for existing files
Memory bank files are read on demand, not auto-loaded, to save context
Includes a statusline script showing context usage, cost and git branch
Lists install commands for MCP servers and plugins such as Context7 and Safety Net
Weaknesses
Personal setup; files are copied and edited by hand, not installed as a package
settings.json assumes macOS Terminal-Notifier; README says to remove it elsewhere
Requires a paid Claude plan or a Z.AI account to use Claude Code
README mixes referral links and unrelated guides, making it harder to scan
no GPU
Needs Claude Code, Claude paid plan or Z.AI, ripgrep, fd, jq, Terminal-Notifier (macOS)
CLI and agent skill that makes App Store and Google Play screenshots
2.6k starscustom licenselast commit Oct 2026
Goldie takes captures replayed by argent on an iOS simulator or Android emulator, then frames them with a device bezel, background and headline. It also joins clips into an App Store preview video and checks output against store upload rules. It installs as an npm CLI plus a skill for Claude Code, Cursor, Codex and other skills-compatible agents, with a browser studio for tweaks.
Strengths
Framework agnostic: SwiftUI, Compose, Flutter and React Native work the same
Covers iPhone, iPad 13-inch and Google Play sizes from the same flows
Studio saves choices to goldie.design.json so CLI renders match
Multi-locale rendering, with optional per-locale captures on iOS
Weaknesses
iOS screenshots require macOS with Xcode simulators
Depends on argent for capture, plus ffmpeg and Node 20 or newer
Flows break when the app UI changes and need repair or re-recording
Needs a Release build; debug builds paint LogBox banners into captures
Lets CLI coding agents message each other through a shared SQLite file
1.6k starsMITlast commit Oct 2026
agmsg is a Bash skill that lets Claude Code, Codex, Gemini CLI, Copilot CLI, Antigravity, OpenCode and other CLI agents send messages to each other through a local WAL-mode SQLite database. It installs as a slash command or `$` command and adds role switching (`actas`), spawning peer agents in tmux or OS terminals, and several delivery modes. It is not MCP and needs no broker; an optional Node-based daemon (agmsgd, beta) exists.
Strengths
Runs on bash and sqlite3 only; no MCP server, broker or network required
Works across Claude Code, Codex, Gemini CLI, Copilot CLI, Antigravity and OpenCode
Message history persists in SQLite and can be replayed into a new agent
Installs via npx, curl script, git clone or Claude Code plugin marketplace
Weaknesses
Windows requires Git Bash; no PowerShell implementation, and bare bash may hit WSL
npm package and plugin can lag main by a few fixes
agmsgd daemon is beta and needs Node >= 22.13.0
Codex without monitor or daemon only checks its inbox on its next turn
Claude Code plugin marketplace with a skill-creator fork and bundled skill suites
1.5k starsMITlast commit Oct 2026
A Claude Code plugin marketplace (named daymade-skills) that ships a fork of Anthropic's skill-creator plus suites for document conversion, macOS/iOS, Codex, Claude Code operations and financial data. Standalone skills cover GitHub operations, Terraform, Cloudflare diagnostics and YouTube downloading. Install with /plugin marketplace add daymade/claude-code-skills or the claude plugin CLI.
Strengths
skill-creator adds a security_scan.py gate with gitleaks before packaging
Suites install under one namespace, e.g. daymade-docs:doc-to-markdown
Standard Claude Code plugin install; suites also installable via the Codex plugin marketplace
Optional routing hook resolves a clash when Anthropic's skill-creator is also installed
Weaknesses
Suite skills are bundle-only; a single member cannot be installed alone
Agent skills and plugins that drive Google Stitch design and code generation
8.6k starsApache-2.0last commit Aug 2026
A set of Agent Skills grouped into three plugins (stitch-design, stitch-build, stitch-utilities) for Google Stitch. They generate and edit screens, extract DESIGN.md and static HTML from code, and convert Stitch designs to React, React Native, Vite dashboards and Remotion videos. Install via Codex marketplace, `npx plugins` for Claude Code and Cursor, or manual copy for OpenCode; all skills need the Stitch MCP server.
Strengths
Covers both directions: code to Stitch design and Stitch screens to React or React Native
Plugin install for Codex, Claude Code and Cursor; skills follow the Agent Skills standard
Each skill ships scripts, resources and examples alongside SKILL.md
Includes DESIGN.md extraction and generation to keep design tokens consistent
Weaknesses
Requires the Stitch MCP server and Google Stitch credentials; not usable standalone
Skills have inter-dependencies, so selective install can leave gaps
OpenCode needs manual copying and renaming of skills with stitch:: names
Not an officially supported Google product
no GPU
Needs Stitch MCP server, Google Stitch, Node.js (npx)
Agent skill that scaffolds a Next.js editor for App Store and Play screenshots
7.6k starsMITlast commit Oct 2026
A skill for coding agents such as Claude Code, Cursor and Codex that generates a Next.js project for designing store screenshots. The editor has a connected canvas, Apple device bezels and an inspector. It exports PNG bundles at the required App Store, Mac App Store and Google Play sizes. Deck state is saved to app-store-screenshots.json, so it can be committed and resumed.
Strengths
Exports PNGs at listed sizes for iPhone, iPad, Watch, TV, Mac, Android and feature graphic
Deck state is one JSON file plus uploaded images, so it is git-trackable
Ships 18 named visual styles with specs and a headline copy library
Supports multiple locales and RTL guidance, with export zip organized by locale
Weaknesses
Needs a coding agent to scaffold the project; not a standalone app
Apple bezels are licensed for mock-ups of Apple-platform apps only
Output quality depends on the agent and the source captures you supply
No hosted demo; you must run the Next.js dev server locally
no GPU
Needs Node.js 20.9+, bun, pnpm, yarn or npm, A skills-compatible coding agent
A showcase of reusable CSS transitions such as card resize, modal open/close, menu dropdown, icon swap and error shake. Each card gives a self-contained snippet with CSS custom properties, t-* classes and a prefers-reduced-motion guard. The same set is available through `npx transitions-dev add` and as an agent skill for Cursor, Claude Code and Codex, plus a Refine tool that suggests motion changes on a running app.
Strengths
Snippets are plain CSS with custom properties and a reduced-motion guard, no framework required
Install via `npx transitions-dev add <name>` or `npx skills add Jakubantalik/transitions.dev`
Skill files are generated from index.html, so they match the showcase site
Static HTML site runs locally with `python3 -m http.server`
Weaknesses
Custom license for the transitions: redistributing as a competing library is prohibited
Some transitions (e.g. confetti-burst) are Pro and need browser sign-in
Refine's LLM mode depends on installing the Cursor CLI
Fixed set of about eighteen transitions; no release tags
iPhone remote control for Codex and OpenCode running on your Mac
3.3k starsApache-2.0last commit Sep 2026
Remodex pairs an iOS app with a Node.js bridge (`remodex up`) on your Mac, which spawns `codex app-server` and forwards JSON-RPC over a WebSocket relay. Prompts, streaming output, git actions and OpenCode chats are handled from the phone while the coding runtime stays local. Pairing uses a one-time QR and an end-to-end encrypted channel; a local relay script is included for self-hosting.
Strengths
End-to-end encrypted session using X25519 keys; QR pairing, then trusted reconnects
Runs Codex, git and workspace actions locally on the Mac, not on a hosted service
Includes relay source and run-local-remodex.sh for self-hosting, default port 9000
Phone supports git commit/push/pull, plan mode, steering runs and queued prompts
Weaknesses
Background daemon and auto-reconnect work on macOS only; other OSes run in the foreground
Needs a signed iOS build or App Store install; no Android client
Plain LAN ws:// pairing can be unreliable on iOS; Tailscale recommended
Author warns of early-stage bugs and is not accepting contributions yet
no GPU
Needs Node.js 18+, Codex CLI (authenticated), OpenCode CLI (for OpenCode chats), iPhone with Remodex app, Xcode 16+ (to build from source), Tailscale or Cloudflare Tunnel (recommended for remote)
Agent Skill that lets coding agents write and run Playwright scripts
3.2k starsMITlast commit Aug 2026
An Agent Skill, also packaged as a Claude Code plugin, that has a coding agent write custom Playwright code for a browser task and run it through a run.js executor. It opens a visible Chromium window by default and returns console output and screenshots. SKILL.md stays short and a fuller API_REFERENCE.md is loaded only when needed.
Strengths
Agent writes real Playwright code, so loops, assertions, network interception and video are possible
Scripts are plain files you can keep and rerun
Installs via the skills CLI for Claude Code, Cursor, Copilot, Codex, Gemini CLI and others
Runs headed by default, so you can watch the automation
Weaknesses
Requires a manual npm run setup step after install to fetch Playwright and Chromium
Only Chromium is installed by default; other browsers need npm run install-all-browsers
README points to Microsoft's playwright-cli for simple interactive browsing
Helper screenshots go to the OS temp directory unless PW_ARTIFACT_DIR is set
Agent skill that turns AI video into scroll, mouse and touch-driven web animation
2.7k starsMITlast commit Oct 2026
Oil Motion is a skill for coding agents that plans an interaction animation, generates keyframes and the video between them, cleans the frames, and wires the result to scroll, mouse, drag, touch or device orientation. Output is a baked MP4, an alpha WebP atlas or a green-screen MP4 keyed in WebGL, chosen by a budget script. Generation goes through ZenMux by default, with a pilot clip approved first.
Strengths
Confirms keyframes first, then generates motion between them to limit identity and structure drift
Picks one asset format (baked MP4, alpha WebP atlas, green-screen MP4) from a budget script
Pilot approval with SHA-256 frame hashes blocks mismatched clips before any network call
API key is stored in the OS credential store, not pasted into chat
Weaknesses
Generation needs a ZenMux API key and is billed per generation
README says it is only verified on macOS
Not meant for UI transitions that CSS, SVG, Lottie or real-time 3D can handle
README is Chinese-first; the English version is a separate file
Azure skills and MCP servers for coding agents, installed as one plugin
1.6k starsMITlast commit Oct 2026
Plugin that gives coding agents Azure workflow skills (prepare, validate, deploy, diagnostics, cost, RBAC) plus the Azure MCP Server and Foundry MCP for acting on live resources. The README lists 200+ tools across 40+ Azure services. It installs into GitHub Copilot, Claude Code, Cursor, Codex CLI, Gemini CLI, VS Code and IntelliJ.
Strengths
Bundles skills and MCP config so guidance and tool execution stay aligned
Installs across Copilot CLI, Claude Code, Cursor, Codex, Gemini CLI, VS Code, IntelliJ
Covers deploy, diagnostics, cost, quotas, RBAC and Foundry model workflows
Supports sovereign clouds (China, US Government) via a --cloud argument
Weaknesses
Azure only; no use outside Azure and Foundry
Needs Node.js 18+, Azure CLI login, and azd for deployment workflows
Repo is synced from GitHub-Copilot-for-Azure; contributions go upstream
Azure MCP telemetry is collected by default and must be disabled via env var
Claude Code-style sub-agents and scripted workflows for the pi coding agent
1.4k starsMITlast commit Oct 2026
An extension for the pi coding agent that spawns sub-agents in isolated sessions, each with its own tools, system prompt, model and thinking level. Agents run in the background by default and can be steered, resumed, scheduled, or isolated in git worktrees. A SubagentWorkflow tool runs deterministic JavaScript scripts (agent, parallel, pipeline), including scripts written for Claude Code's Workflow tool.
Strengths
Background agents with queuing, default concurrency limit of 10, and completion notifications
Custom agent types in markdown with YAML frontmatter, tool denylist and skill preloading
Git worktree isolation auto-commits each agent's changes to a branch
Workflow scripts run in a node:vm sandbox and match Claude Code's Workflow tool
Weaknesses
Only works inside pi, which must be version 1.1.0 or newer
Other hosts such as DeepSeek Harness rely on untested third-party adapters
Foreground-to-background handoff is one-way and not offered for queued or nested agents
Headless pi -p does not wait for scheduled subagents
no GPU
Needs pi >= 1.1.0
Models: Any model configured in pi, selectable by fuzzy name such as haiku or sonnet
Agent skill that builds scroll-scrubbed fly-through landing pages from AI-generated video
10.4k starsMITlast commit Jul 2026
A SKILL.md-compatible skill for Claude Code, Codex and other agents. It interviews you about brand, scenes and budget, then generates isometric scene stills and Seedance or Kling camera clips with frame-locked seams. A vanilla-JS scrub engine ties the chain to scroll position, with an optional 9:16 mobile chain.
Strengths
Frame-locked connector clips give one continuous flight with no cuts between scenes
Scrub engine is framework-agnostic vanilla JS: plain HTML, Next.js, Vue or Python-served pages
Optional native 9:16 portrait chain for phones instead of a cropped landscape film
States estimated cost and asks for approval before any generation runs
Weaknesses
Needs paid services: Monid CLI plus Higgsfield credits; a 6-scene 1080p chain is about $27
Generated mp4/webp assets are not shipped; every project must render its own
Mobile chain doubles the video generation count and cost
Requires ffmpeg, Python 3 with Pillow and authenticated third-party CLIs
Visual workflow canvas that exports agent skills and commands as Markdown
5.4k starscustom licenselast commit Aug 2026
A VS Code extension with a React Flow canvas for designing multi-agent workflows. It exports them as Markdown skills, agents or commands for Claude Code, Copilot, Codex CLI, Gemini CLI, Cursor and others. The same workflow.json can also be driven through an npm CLI (ccwf) and a stdio MCP server (ccwf-mcp), so an external agent can read and edit workflows.
Strengths
Exports to eight agents' formats, e.g. .claude/commands, .codex/skills, .cursor/skills
CLI and MCP server work without VS Code, usable in CI or over SSH
MCP tools (get_workflow, apply_workflow) let agents edit the canvas
CLI and MCP packages are MIT-licensed and published on npm
Weaknesses
VS Code extension is AGPL-3.0-or-later, which may restrict some uses
Agents other than Claude Code must be enabled from the toolbar More menu
Usage examples and tutorials are listed as coming soon
Needs a separate agent CLI or extension installed to run workflows
no GPU
Needs VS Code, Claude Code or another supported agent, Node.js (for npx CLI)
Design-practice skills and commands for Claude Code and Gemini CLI
3.1k starsMITlast commit Sep 2026
A Claude Code plugin marketplace of markdown skills and slash commands for design work: research, UX strategy, UI, interaction design, design systems, prototyping and testing, handoff, and visual critique. This repo hosts 9 plugins (111 skills, 34 commands). The marketplace also pulls in four companion collections, for 273 skills and 76 commands in total. Each plugin also works as a Gemini CLI extension.
Strengths
Installs from one marketplace command, then per plugin via /plugin install
Commands chain skills into workflows, e.g. /design-research:discover runs personas, empathy map, journey map
Works in both Claude Code and Gemini CLI
/designer-toolkit:start-here routes you to a first command and what to skip
Weaknesses
Only Claude Code and Gemini CLI are supported
Content is markdown prompts only; output quality depends on the host model
Skill counts in the README are inconsistent (111 vs 107 for this repo)
PRs without a matching open issue are closed without review
macOS app to browse and edit coding-agent skills across tools
2.1k starscustom licenselast commit Aug 2026
Chops is a native SwiftUI/SwiftData macOS app that scans the skill and agent directories of Claude Code, Cursor, Codex, Windsurf and Amp, plus project-level Copilot and Aider files. It lists them in one place with full-text search, a built-in editor with frontmatter parsing, and collections that leave source files untouched. It can also browse remote servers and install skills from the skills.sh registry.
Strengths
Scans skill and agent directories for seven tools, with custom scan paths
FSEvents file watching re-scans automatically when files change on disk
Dedups symlinked skills into one entry with multiple tool badges
Installs skills from the skills.sh registry and from remote servers
Weaknesses
macOS 15 or later only
App sandbox is disabled to read dotfiles across the home directory
No automated test suite; changes are validated manually
FSL-1.1-MIT license, not a standard OSI-approved license
no GPU
Needs macOS 15+, Xcode (build from source), xcodegen (build from source), Sparkle
Code-defined workflow orchestration with approval gates for AI coding harnesses
1.9k starsMITlast commit Aug 2026
Babysitter lets you write a workflow as a JavaScript process that calls tasks, breakpoints and quality gates, then drives an AI coding agent through it step by step. Every decision is recorded in an event-sourced journal. It installs as a plugin for 12 harnesses, including Claude Code and Codex, and has an internal harness for headless CI runs.
Strengths
Workflows are plain JavaScript, with human breakpoints and quality gates enforced in code
Same processes run across 12 supported harnesses via the Adapters runtime
Internal harness runs processes headless in CI with no external coding agent
Event-sourced journal supports resuming interrupted runs and diagnosing them
Weaknesses
Install is split across several npm packages (babysitter, adapters-cli, genty-platform)
Only Claude Code and Codex have dedicated pages; Codex is beta, most others experimental
Adapters CLI needs Node 22.13+, higher than the Node 20 baseline
No hosted demo; the monorepo is large and the README is dense
no GPU
Needs Node.js 20+, AI coding harness (e.g. Claude Code)
Models: Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot
Codex profiles that set up an orchestrator, Luna subagents and a reviewer
1.7k starsApache-2.0last commit Oct 2026
Installer scripts (setup.sh, setup.ps1) copy a chosen Codex profile into a target repository as .codex/, .agents/ and AGENTS.md. Each profile defines a root orchestrator (Astra, Sol or Luna), explorer, worker, tester and researcher subagents on GPT-6 Luna, and a separate reviewer. Six profiles vary the model, reasoning level and concurrent subagent count (2 or 4).
Strengths
Six ready-to-copy profiles with pinned models and reasoning levels per role
Installer preserves existing AGENTS.md and asks before overwriting component files
Manual and global (~/.codex) setup paths are documented
Includes tests and a token_usage.py script for measuring usage
Weaknesses
Requires Codex CLI and access to the GPT-6 model family
Project-scoped config loads only for trusted projects
Skill is named astra-orchestrator even in Sol and Luna profiles
Global setup needs a manual merge into ~/.codex/config.toml
Agent skill that generates SVG, PNG, GIF and HTML diagrams from prompts
11.8k starsMITlast commit Oct 2026
An Agent Skill for Codex and Claude Code that generates technical diagrams from English or Chinese descriptions. A Python CLI renders SVG with orthogonal routing and geometry checks, then exports PNG, an animated GIF, or one offline interactive HTML file. It ships 12 visual styles and 14 diagram types, including UML and AI/agent patterns such as RAG and multi-agent.
Strengths
Same skill installs into Codex and Claude Code via skills CLI or git clone
Validates XML, markers, edge crossings and label clearance before export
Outputs SVG, PNG up to 4x, SVG-to-GIF motion and offline HTML with pan/zoom
12 styles and 14 diagram types, including all UML types and AI/agent patterns
Weaknesses
Needs an agent runtime (Codex or Claude Code); not a standalone diagram app
PNG and GIF export rely on optional dependencies; Puppeteer path needs Node.js 22.12+
GIF motion only works on generated SVGs matching one of twelve approved contracts
Visual review is skipped when the runtime cannot read images
no GPU
Needs Python 3.9+, Codex or Claude Code, Node.js 22.12+ (optional, Puppeteer PNG export)
Agent Skill that generates simple, cute animal mascot images from a brief
6.1k starsMITlast commit Aug 2026
A single-file Agent Skill (SKILL.md) that guides a compatible agent through proposing three mascot directions, then generating six separate square images with an image model. Each image follows fixed rules: 4–7 basic shapes, three colors, and the character emerging from the lower-left or lower-right corner. It installs with `npx skills@latest add s1dashu/ip-as-logo-skill`.
Strengths
Fixed, checkable design rules: 4–7 shapes, three semantic colors, 3 left / 3 right split
Installs through the Agent Skills CLI, with a global option
Works with several agents (Codex, Coze, Doubao, Manus, Gemini Apps, Replit Agent)
No scripts or dependencies; the skill is one instruction document
Weaknesses
Needs a top-tier image model (GPT Image 2, Nano Banana Pro or similar) or an API key
Each candidate is generated once, with no validation or automatic retry
Never falls back to SVG, so output is raster images only
Defaults to familiar animals; other subjects need a stated product reason
no GPU
Needs Agent Skills CLI (npx skills), Image generation model or API key
Skill that runs isolated parallel reasoning frames plus a critic pass for ideation
4.5k starsMITlast commit Oct 2026
ADHD is an agent skill, npm CLI and TypeScript library that spawns N isolated LLM calls, each under a distorted cognitive frame with no shared context. A separate critic call then scores ideas on novelty, viability and fit, flags traps, clusters them and deepens the top-K. It installs with `npx skills add` into Claude Code, Cursor, Codex, Cline, Gemini CLI and others.
Strengths
Generator and critic are separate LLM calls, so the split is mechanical, not prompt-based
Installs as a skill into Claude Code, Cursor, Codex, Cline, Gemini CLI and about 50 other agents
Also ships as an npm CLI (adhd-agent) and library returning shortlist, traps and clusters
Eval transcripts and methodology are in the repo (bench/results.json, documentation/evals.md)
Weaknesses
Benchmarks are author-run: 6 problems scored by an LLM judge
One external review measured about 2.3x time and 1.9x output versus single-shot
Aimed at open-ended ideation; little use for routine coding tasks
README does not say which model providers the CLI and library support
Agent Skills and MCP server for building on Cloudflare Workers
3.2k starsApache-2.0last commit Oct 2026
A plugin of 18 Agent Skills covering Workers, Wrangler, Durable Objects, the Agents SDK, Sandbox, Turnstile, email and Cloudflare One. Skills load automatically when a request matches their triggers. The plugin also bundles Cloudflare's remote MCP server, which gives the agent access to the Cloudflare API and current docs.
Strengths
Native plugin install for Claude Code, Codex, VS Code/Copilot and Cursor
Also installable via npx skills or by copying folders into a skills directory
Covers many products: Workers, Durable Objects, Agents SDK, Sandbox, Cloudflare One
Bundles the main Cloudflare remote MCP server for API and docs access
Weaknesses
Only useful for projects that run on Cloudflare
Bundles only the main cloudflare MCP server, not product-specific ones
Some skills target previews, e.g. sandbox-next is a 1.0 preview
No tagged releases
no GPU
Needs Cloudflare account, Cloudflare remote MCP server
Models: Claude Code, OpenAI Codex, GitHub Copilot, Cursor, OpenCode
Agent Skills for auditing web performance, Core Web Vitals, accessibility and SEO
3k starsMITlast commit Aug 2026
Six SKILL.md packages (web-quality-audit, performance, core-web-vitals, accessibility, seo, and a security and code-quality skill) that guide a coding agent through Lighthouse, WCAG 2.2 and Core Web Vitals checks. They separate CrUX field data, RUM, lab traces and static inspection as evidence types. Chrome DevTools MCP is optional; without it the skills fall back to Lighthouse CLI, PageSpeed Insights/CrUX and manual checks.
Strengths
Separates field, RUM, lab and static evidence instead of mixing them in one score
Installs via npx skills, Claude Code plugin, Codex marketplace or Gemini CLI extension
Stack-agnostic; covers React, Vue, Svelte, Astro and plain HTML
Falls back to Lighthouse CLI and PageSpeed Insights when DevTools MCP is absent
Weaknesses
Unofficial; not maintained by the Lighthouse or Chrome teams
Live audits need Chrome DevTools MCP, which is a separate install
No tagged releases
Instruction files only; results depend on the host agent following them
Skills that let one coding agent delegate tasks to other coding CLIs
2.5k starsMITlast commit Oct 2026
A set of agent skills, installed with the Skills CLI, that let an orchestrating agent hand work to another coding CLI such as Codex, Claude Code, Cursor Agent, OpenCode or Aider. Each skill ships a Node relay that runs the CLI, writes a structured result.json and never commits. A setup skill discovers installed CLIs and maps them to named lanes like feature, tests and ui.
Strengths
Covers 18 implementer CLIs behind one result.v1 contract, so you can swap them freely
Relays use Node built-ins only, with no dependencies, telemetry or network calls of their own
Relays never commit; you review the diff and land the commit
Project lane config is bound to explicit approval and fails closed if edited or cloned
Weaknesses
Several CLIs have no enforced read-only mode or sandbox, per the README footnotes
Verification is uneven: Aider was tested only against a stub endpoint, not on macOS or Linux
Needs each implementer CLI installed and authenticated separately
No tagged release, so pinning to a version tag is not yet possible
no GPU
Needs Node 18+, git, Skills CLI (npx skills), at least one implementer CLI (Codex, Claude Code, Cursor Agent, OpenCode, Aider, etc.)
Models: Aider with any OpenAI-compatible endpoint, Models via the installed CLIs (Codex, Claude Code, OpenCode, Cursor, etc.)
28 Agent Skills that give coding agents structured reasoning procedures
1.7k starsMITlast commit Aug 2026
A catalog of 28 Agent Skills, each a SKILL.md procedure for a reasoning method such as pre-mortem, theory of constraints, scientific-method debugging, or reversibility checks. A router skill picks none, one, or up to three skills for a problem. Installs through the skills CLI, as a Claude Code plugin, or by copying directories into .agents/skills or .claude/skills.
Strengths
Router skill can return NONE, one skill, or up to three complementary skills
Three install paths: npx skills CLI, Claude Code plugin, or plain directory copy
Works with Claude Code, GitHub Copilot, Codex, Cursor and other Agent Skills tools
Includes validation script and structural, routing and outcome evals in the repo
Weaknesses
Own audit reports only a provisional +4.0 point gain, below its +5 point margin
All 28 skills are manual-only; none trigger automatically
Routing eval requires the authenticated droid CLI
Skills are prompt procedures, not tooling; benefit depends on the model following them
A collection of SKILL.md files, each teaching an AI assistant one professional task such as writing a PRD, decoding a lease or running a postmortem. Each skill bundles a framework, output template and quality checks. An npx installer copies chosen skills into the config folder of the tool in use, and a Claude Code plugin and MCP server are also provided.
Strengths
Plain markdown files; installing copies text and uninstalling deletes it, no runtime or account
Installs into Cursor, Codex, Windsurf, Cline and Zed via npx, plus Claude plugin, MCP and REST access
153 bundles grouped by job, with 178 translations in 10 languages
CI checks structure, near-duplicates and a 15-document prompt-injection suite on every change
Weaknesses
Mostly non-coding content (PM, job search, document decoding); little aimed at software engineering
Only 409 of 1,305 skills have eval cases; 399 have blind-judged scores
0 of 161 high-stakes skills have a current review by a qualified professional
ChatGPT and Gemini use copied exports rather than a native install
Claude Code plugin that runs Codex reviews and delegates tasks
34.5k starsApache-2.0last commit Jul 2026
A Claude Code plugin that adds slash commands for running Codex code reviews and handing tasks to Codex without leaving Claude Code. It wraps the local Codex CLI and app server, so it reuses your existing Codex login, config.toml and base URL settings. Background jobs can be tracked and cancelled, and sessions can be resumed in Codex.
Strengths
Reuses local Codex CLI auth and config.toml; no separate runtime
Read-only /codex:review and steerable /codex:adversarial-review with --base <ref>
Background jobs via /codex:status, /codex:result and /codex:cancel
Sessions can be continued in Codex with codex resume <session-id>
Weaknesses
Requires Claude Code plus a ChatGPT subscription or OpenAI API key
Usage counts against Codex limits; optional review gate can drain them quickly
Multi-file reviews and rescue tasks can run long
/codex:transfer needs a Codex version that supports session import
no GPU
Needs Claude Code, Codex CLI (@openai/codex), Node.js 18.18+, ChatGPT account or OpenAI API key
Installable agent skills for visual plans, PR recaps and review automation
4.6k starsMITlast commit Oct 2026
A collection of agent skills installed with `npx @agent-native/skills@latest add` into Claude Code, Codex, Cursor, Copilot and similar agents. It includes visual-plan and visual-recap (MDX plans and diff recaps, with an optional PR GitHub Action), webmcp, visual-edit, agent-watchdog and usage-limit skills, plus an experimental Factory set that turns feedback and errors into policy-gated PRs. Also published as a Claude Code and Cowork plugin marketplace.
Strengths
Single npx installer with a picker, user or project scope, and a per-skill --skill flag
Works across Claude Code, Codex, Cursor, OpenCode and Copilot via the shared .agents path
Factory config separates fix, reply, review, approve and merge policies in YAML
Local-files mode serves visual plans over localhost instead of uploading them
Weaknesses
Visual plans and recaps are viewed in the Agent-Native plans app, hosted by default
Factory skills are experimental and need Slack, Sentry or GitHub connections you set up
/rewind needs macOS and the separate Clips Desktop app
Skills are prompt instructions, so results depend on the host agent and model
Agent skill for building scroll-driven websites with a design standard
4k starsMITlast commit Sep 2026
scroll-craft is a skill for Codex, Claude Code or other coding agents that builds scroll-driven sites with video scrubbing, pinned sections and pointer effects. It supplies a design workflow, eight page grammars, a fingerprint gate against repeating earlier builds, and a headless-browser pass that checks dead scroll, contrast and stuck video. It also ships as a Claude Code plugin.
Strengths
Headless browser checks every scroll position for dead scroll, contrast and undecoded clips
Fingerprint gate forces each new build to differ on at least 4 of 6 dimensions
Works with your own photos and footage; the generation API key is optional
Installs as a Claude Code plugin or as a plain folder for other agents
Weaknesses
Author states it has only been run on Windows; no macOS build done
Needs a full ffmpeg build and playwright-core with Chrome for encoding and verification
Generated assets need a KIE_AI_API_KEY and real spend for video-heavy builds
Opinionated: refuses common layouts and palettes, and the engine is not edited per project
Claude Code and Codex plugin that makes the AI ask for your design first
3.9k starsMITlast commit Oct 2026
VibeWise is a plugin for Claude Code and Codex that adds a learning mode. The AI asks for your approach, discusses tradeoffs, and writes code only after you confirm a design and an implementation checkpoint. Preferences, notes and a project map are stored in .vibe-wise/ in your project and shared between both tools.
Strengths
Works in both Claude Code and Codex, sharing the same .vibe-wise/ notes
Explicit Build, Design and Implementation checkpoints gate when code is written
Adjustable experience level and checkpoint frequency (Light, Normal, Frequent)
No extra account, backend or telemetry; Python 3 only, no extra packages
Weaknesses
Slower than direct code generation by design, since every step waits on your input
Requires Claude Code or Codex; not usable standalone
In Codex, the learn command must be re-run at the start of each session
Saved notes enter the AI context, so they follow your tool's data settings
Claude Code skill that generates HTML slide decks or converts PowerPoint files
31.2k starsMITlast commit Jun 2026
A skill for Claude Code, packaged as a plugin, that builds single-file HTML presentations with inline CSS and JS. It shows three visual style previews to pick from, then writes the full deck, or extracts text, images and notes from a .pptx and rebuilds it for the web. Other coding agents can follow the same SKILL.md if they have filesystem and shell access.
Strengths
Output is one HTML file with inline CSS/JS, no npm or build step
Converts .pptx files, extracting text, images and speaker notes
Ships 12 included styles plus 34 optional bold templates loaded on demand
Installs as a Claude Code plugin or by copying files into ~/.claude/skills
Weaknesses
Plugin install flow and slash command are specific to Claude Code
Other agents must read SKILL.md manually; no install path is documented for them
Output is fixed 16:9 HTML, not editable .pptx
Style previews depend on the agent's generation, so results vary per run
no GPU
Needs Claude Code or another coding agent, Python (scripts/extract-pptx.py)
Models: Claude Code, Codex, Kimi Code, OpenCode, Gemini CLI
Agent skill set applying Mao-era methodology to AI problem-solving
3.8k starsMITlast commit Oct 2026
A set of nine methodology skills plus a core "seek truth from facts" kernel for coding agents, drawn from Mao Zedong's writings (contradiction analysis, investigation first, protracted strategy and others). A hook injects a roughly 50-line kernel at session start, and the other skills load only when relevant. Installs with `npx qiushi-skill` into Claude Code, Cursor, Codex, OpenCode and other hosts, or through the Claude Code marketplace.
Strengths
One npx command installs to Claude Code, Cursor, Codex, OpenCode, OpenClaw, Hermes and nanobot
Only a ~50-line kernel stays resident; method skills load on demand
Includes slash commands, two subagents (investigator, self-critic) and a validate script
Each skill has an original-texts.md with source quotations
Weaknesses
Main README is in Chinese; the English version is a separate file
No benchmarks or measured effect on agent output quality in the README
Content is built on a specific political text corpus, which may not suit every team
Agent skills library for Azure SDKs and Microsoft Foundry coding agents
3.1k starsMITlast commit Oct 2026
A collection of roughly 175 SKILL.md skills for Azure SDKs and Microsoft Foundry, split by language (Python, .NET, TypeScript, Java, Rust) plus core and language-agnostic Foundry skills. It also ships custom agents, an AGENTS.md template, MCP configs and plugin packages. Install with `npx skills add microsoft/skills`, which copies chosen skills into the agent's directory, such as .github/skills/ for GitHub Copilot.
Strengths
Skills cover Azure SDKs across Python, .NET, TypeScript, Java and Rust
Installer wizard picks skills per agent and symlinks across multiple agents
Includes custom agents, an AGENTS.md template and MCP configs alongside skills
Has CI test harness and Copilot SDK evaluation workflows for skills
Weaknesses
Azure and Foundry specific; little use outside the Microsoft ecosystem
README warns that loading all skills causes context rot, so selection is manual
Marked work in progress; several Foundry skills are preview
README skill counts are inconsistent (174, 175, 28 vs 29 .NET)
no GPU
Needs Node.js (npx), GitHub Copilot or another coding agent
AWS plugins, skills and managed MCP server for coding agents
3k starsApache-2.0last commit Oct 2026
Official AWS toolkit that gives coding agents (Claude Code, Codex, Cursor, Kiro, fx) a managed AWS MCP Server plus task-specific skills. The MCP server exposes 300+ AWS services through one authenticated endpoint, runs sandboxed Python scripts and searches AWS documentation. IAM condition keys, CloudWatch metrics and CloudTrail logging separate agent actions from human ones.
Strengths
IAM condition keys let policies target agent actions separately from human actions
CloudWatch metrics and CloudTrail audit logging for every MCP request
Plugins for Claude Code, Codex and Cursor; skills and MCP config for Kiro, fx and others
Documentation search works without AWS credentials
Weaknesses
Core MCP server is a managed AWS endpoint, not self-hosted
API calls and script execution need an AWS account with local credentials
Plugins exist only for Claude Code, Codex and Cursor; other agents need manual setup
Requires uv installed to run the MCP proxy
no GPU
Needs AWS account and credentials, uv, AWS MCP Server (managed endpoint), Node.js (npx skills)
Claude Code and Codex skill that turns a topic into a narrated explainer video
2.7k starscustom licenselast commit Sep 2026
A skill for Claude Code and Codex that guides an agent through research, narration, TTS voiceover, storyboarding, parallel shot building and QC. The result is a 1280x720 motion-graphics MP4 in Chinese or English, with every frame written as a Remotion (React/TypeScript) component. It ships a Remotion template, style and motion specs, QC scripts and one reference film. It is not a CLI.
Strengths
Four user checkpoints: length and language, narration, voice, and a 30-second look check
Deterministic renders: frames are pure functions of frame number with seeded randomness
Local kokoro-82m, piper and kokoro-onnx, cloud edge-tts, or bring your own audio
Includes a reference film paper trail and a frame-metrics QC script
Weaknesses
PolyForm Noncommercial license; commercial use needs the author's authorization
One visual style with two backdrops; other changes mean editing style guide and ui.tsx
Landscape 1280x720 only, no 9:16; Chinese and English only
Runs take about 1 to 3 hours and 2 to 3 GB disk; verified on macOS, Windows untested
no GPU
Needs Claude Code or Codex, Node >=18, Remotion 4.0.507, ffmpeg, Python 3 with edge-tts, zsh, kokoro-82m (English narration), Chromium (Linux ARM)
Agent skills for Node.js, Fastify, TypeScript, OAuth and GitHub workflows
2k starsMITlast commit Oct 2026
A collection of eleven skills for AI coding agents, written by Matteo Collina. They cover Node.js and Node core internals, Fastify, TypeScript types, OAuth 2.0/2.1, neostandard with ESLint 9, Diátaxis documentation, gh CLI usage, and AGENTS.md generation. The repo also has a skill-optimizer skill and docs on cross-model benchmarking.
Strengths
Focused on the Node.js ecosystem: Fastify, Node core, neostandard, OAuth and TypeScript
Includes a skill-optimizer skill and a documented cross-model benchmarking workflow
Includes an init skill that creates and maintains AGENTS.md files
MIT licensed, with each skill covering one named topic
Weaknesses
README gives no install command or list of supported agents
Most skills are Node/JavaScript-specific, with little use outside that stack
No tagged releases
README gives only one-line skill descriptions, with no example output
Shared team memory and auto-generated skills for coding agents
1.6k starsApache-2.0last commit Oct 2026
Hivemind installs hooks or plugins into Claude Code, OpenClaw, Codex, Cursor, Hermes, pi and Claude Cowork (alpha). It captures prompts, tool calls and responses as traces in a Deeplake workspace, and mines them into SKILL.md files shared across the team. Agents recall past work through hybrid lexical and semantic search over those traces.
Strengths
One installer wires hooks for seven agents; `hivemind status` shows what is connected
Mines captured traces into SKILL.md files scoped to a workspace
Capture can be disabled by env var or a per-directory `.hivemind` file
README reports LoCoMo results: 25% lower cost, 31% fewer turns versus no memory
Weaknesses
Requires a Deeplake account and cloud API; no fully local mode is described
All workspace users can read every captured prompt and tool output
Claude Cowork capture is alpha and covers Local Agent Mode sessions only
Background summary and skill-mining runs invoke `claude -p`, adding model cost
Claude Code output style that makes Claude write clear, grammatical Korean
1.5k starsMITlast commit Aug 2026
A Claude Code plugin with two output styles that instruct Claude to keep Korean particles, endings and common vocabulary instead of terse, telegraphic text. One variant keeps Claude Code's coding instructions and the other removes them. The guideline text can also be pasted into other AI tools as a plain writing instruction.
Strengths
Two variants: fluent-korean keeps coding instructions, fluent-korean-not-coding drops them
Installs with two /plugin commands in Claude Code
Optional add-on blocks for honorifics, beginner tone, or forcing Korean reasoning
Style files are plain markdown, reusable in CLAUDE.md or other AIs
Weaknesses
README says results may fall short in long sessions or with many instructions
Uses somewhat more tokens because omitted sentence parts are restored
Subagent compliance is not guaranteed and must be checked by the user
README is mostly Korean, with only a short English summary
Harness that iterates agent tasks against evaluation and keeps lessons as playbooks
1.3k starsApache-2.0last commit Oct 2026
autocontext runs a task from a plain-language goal through repeated generations scored by evaluation. It keeps useful lessons, discards dead ends, and writes traces, reports, playbooks and datasets to disk for later runs. It ships Python and TypeScript CLIs (autoctx), an MCP server, a Pi extension, and optional local-model distillation.
Strengths
Filesystem-first run output (trace.jsonl, playbook.md, score.json) that can be diffed and replayed
Works with Anthropic, OpenAI-compatible, OpenRouter, Claude CLI, Codex, Pi, and vLLM or Ollama
Context changes are stored as candidates and served only after matched trials confirm them
Exposes its tools to Claude Code or Cursor via autoctx serve mcp
Weaknesses
Large surface area: many commands, scenario families and opt-in features to learn
Python and TypeScript packages differ; some TS features lack Python parity
No Docker image or compose file; install through uv or bun
Reuse studies are described as infrastructure evidence only, not production promotion
GPU optional
Needs uv or bun, Node.js 22.19+ (npm CLI), LLM provider or local agent such as Pi
Models: Anthropic, OpenAI-compatible, OpenRouter, Claude CLI, Codex
Claude Code slash commands and GitHub Actions for PR, security and design review
3.9k starsMITlast commit Oct 2026
A repository of three review workflows for Claude Code: code review, security review and design review. Each combines slash commands with GitHub Actions or subagents; the design review drives a browser through Playwright MCP. It also ships a free 16-page PDF guide on structuring context, tools, workflows and verification around coding agents.
Strengths
Covers three review types: PR code review, OWASP-based security scan, and browser-based UI review
Claude Code plugin that lets Claude watch videos via ffmpeg frames and transcripts
1.4k starsMITlast commit Oct 2026
An MCP server and Claude Code plugin that extracts video frames with ffmpeg and sends them to Claude as base64 images with timestamps, alongside an audio transcription. Audio comes from the Gemini API, local Whisper (whisper.cpp or Python openai-whisper), or the OpenAI Whisper API. It also accepts YouTube URLs via yt-dlp, using manual or automatic captions before falling back to the audio backend.
Strengths
Three audio backends, including fully offline whisper.cpp with automatic model download
YouTube URLs use manual or auto captions first, with transcription_source labeled
Six MCP tools, including video_analyze and video_detail for drilling into specific moments
Frame fps, time range, resolution and format (jpeg, png, webp) are adjustable
Weaknesses
Requires Node.js 24+ and ffmpeg installed locally
YouTube support needs yt-dlp installed separately
Only a perception layer: no interpretation, and frames count against Claude's context
Windows users may need a global npm install workaround if npx fails
GPU optional
Needs Node.js 24+, ffmpeg, yt-dlp (optional, YouTube), whisper.cpp or openai-whisper (local backend), Gemini API key or OpenAI API key (cloud backends)
Models: Gemini API, whisper.cpp, openai-whisper, OpenAI Whisper API
Agent skills that generate HTML wireframes, prototypes, plans and diagrams
3.7k starsMITlast commit Oct 2026
Six skills for coding agents that produce self-contained HTML artifacts: wireframes, working prototypes, plans, diagrams, and general reports or landing pages. They install with `npx skills add`, as a Claude Code plugin, or as a Codex plugin, and can also be read as plain reference. Each skill works on its own, and the broad `html` skill routes requests to the specialist ones.
Strengths
Six skills, each installable individually via `npx skills add --skill`
Packaged for Claude Code and Codex plugin marketplaces as well as the skills CLI
Usable as a reference without installing anything
Skills require artifacts to be responsive, accessible, self-contained and browser-verified
Weaknesses
Output quality depends on the agent and model; no benchmarks or examples of results in README
No release published, so no versioned installs
README does not say how the browser verification step is run
README promotes other tools (Plannotator, tot) alongside the skills
no GPU
Needs Node.js (npx skills CLI), Claude Code or Codex
A set of 172 agent skills grouped into 16 department plugins (CEO, Technology, Security, IT Operations, Product, Marketing, Demand Generation and others), each installable on its own. Skills load when a request matches and can be called as `department:skill`. Manifests for Claude Code and ChatGPT/Codex are generated from the same tree, and each department has a subagent charter in `.claude/agents/`.
Strengths
Each department installs separately, so a project loads only the skills it needs
One tree generates both Claude Code and ChatGPT/Codex manifests
Namespaced names like security:threat-modeling avoid skill name collisions
Includes 11 worked cross-department scenarios and an agent charter per department
Weaknesses
Most skills cover business, marketing and IT; only part is engineering-focused
No tagged release, so versioning is unknown
README gives no benchmarks of skill quality
Install steps are shown in detail only for Claude Code
Python MCP server (stdio, SSE, Streamable HTTP) that stores agent memories in SQLite or PostgreSQL as a URI-routed graph such as core://agent/identity. The agent gets 7 tools to create, update, alias, search and trigger-link memories, and a React dashboard lets the owner browse, diff and roll back every change. A system://boot entry loads configured core memories at session start.
Strengths
Works with any MCP client over stdio, SSE or Streamable HTTP, not tied to one LLM
Every AI write creates a snapshot; dashboard offers diff, accept and rollback
Runs on SQLite locally or PostgreSQL; namespaces isolate multiple agents in one database
Public read-only demo server and dashboard available without installing
Weaknesses
Search is full-text keyword only, not semantic
Needs Python 3.10+ and Node.js; frontend builds on first start
README is mostly Chinese and leans on persona and roleplay examples
Docker setup starts from an empty PostgreSQL with no example data
no GPU
Compose
Needs Python 3.10+, Node.js, SQLite or PostgreSQL, Docker Compose (optional)
86 agent skills for iOS 26, Swift 6.3 and SwiftUI development
1.2k starscustom licenselast commit Jul 2026
A set of 86 self-contained Agent Skills covering SwiftUI, Swift concurrency, SwiftData, StoreKit 2, Foundation Models, HealthKit and other Apple frameworks, written against iOS 26+ APIs. Install through the skills CLI, a Claude Code plugin marketplace with themed bundles, Codex's skill installer, or by uploading zipped skills to Claude or ChatGPT. Each skill stands alone, so you can install only the ones you need.
Strengths
86 skills, each self-contained with no cross-dependencies, so partial installs work
Targets iOS 26+ APIs such as Liquid Glass and Foundation Models, avoiding deprecated patterns
Several install paths: skills CLI, Claude Code marketplace, Codex installer, manual zip upload
Themed Claude Code bundles limit how many skills load into the context window
Weaknesses
License is PolyForm Perimeter 1.0.0, which is not an OSI open-source license
Installing all 86 skills can load a lot into the context window
Apple platforms only; no Android or cross-platform guidance
README gives no evaluation of how much the skills improve agent output
no GPU
Needs Node.js (for the npx skills CLI), A compatible coding agent such as Claude Code, Codex, Cursor or Copilot
Models: Claude Code, OpenAI Codex, Cursor, GitHub Copilot
Bash loop that reruns Amp or Claude Code until every PRD story passes
22.1k starsMITlast commit Feb 2026
Ralph is a bash script (ralph.sh) that repeatedly launches Amp or Claude Code with a fresh context, each time picking the highest-priority story in prd.json where passes is false. It implements one story, runs typecheck and tests, commits on success, and records learnings in progress.txt and AGENTS.md. It also ships prd and ralph skills for writing a PRD and converting it to prd.json.
Strengths
Fresh context each iteration; state persists only in git, progress.txt and prd.json
Supports both Amp and Claude Code via a --tool flag
Small footprint: one shell script, prompt templates and two skills, plus jq
Installable as a Claude Code plugin marketplace entry
Weaknesses
Needs working typecheck and tests; broken checks compound across iterations
Stories must fit one context window; large ones must be split by hand
Requires an authenticated Amp or Claude Code install, which are separate tools
Agent skill that audits a codebase and writes plans for cheaper agents to run
9.3k starsMITlast commit Sep 2026
An Agent Skills-format skill, invoked as /improve, that maps a repo, runs parallel audits across nine categories, re-checks each finding against the cited file and line, and writes self-contained markdown plans into plans/. It never edits source itself. An optional execute mode dispatches a cheaper model in an isolated git worktree and reviews the diff against the plan.
Strengths
Advisor re-reads every cited file:line before showing findings, dropping false positives
Plans inline code excerpts, verification commands and STOP conditions for small executor models
Writes only to plans/; executors work in disposable worktrees and merging stays manual
Scoped modes: quick, deep, branch, security, plus reconcile to refresh the backlog
Weaknesses
Needs an agent host that supports the Agent Skills format
Full audits fan out parallel subagents on a capable model; cost is not documented
No release tagged and no stated model or version requirements
Plan quality depends on the advisor model; no benchmarks given
no GPU
Needs Agent Skills-compatible coding agent, Node.js (npx) for install, git, GitHub (optional, for --issues)
Agent skill that turns PRDs into GitHub issues and parallel agent work
8.4k starsMITlast commit Oct 2026
CCPM is an Agent Skill that guides a coding agent from brainstorming and a PRD to a technical epic, task files, GitHub issues and a dedicated git worktree. It then launches parallel agents on independent work streams for an issue. Project state lives as markdown under `.claude/`, and status, standup and search run as bash scripts without LLM calls.
Strengths
Follows the agentskills.io standard; works with Claude Code, Codex, OpenCode, Factory, Amp, Cursor
Project state is plain markdown in the repo, with no external database
Status, standup, search and validate run as bash scripts with no token cost
Task files carry depends_on, parallel and conflicts_with metadata for parallel agents
Weaknesses
Requires an authenticated gh CLI and a GitHub repository
Parent-child sub-issues need the gh-sub-issue extension; otherwise falls back to task lists
Benchmark and team-result figures are self-reported with no method described
Original /pm:* slash commands only exist on the v1 branch
Agent skill that writes, scores and renders product launch videos as HTML
1.6k starsMITlast commit Oct 2026
An Agent Skills package plus an `om` CLI that lets Claude Code, Codex and similar agents produce short launch films. Each film is a single index.html (GSAP, Canvas, Three.js, SVG) rendered frame by frame through local Chrome and FFmpeg, with a score written in ABC notation. A browser editor lets you click a frame and leave a comment, and the agent edits only that spot.
Strengths
Output is one plain HTML file; no video-generation model, each frame depends only on code and time
Click-to-comment editor sends timestamp, scene and source location to the agent
Music and 23 sound effects are generated locally with no paid API
Exports MP4, transparent MOV, WebM, GIF or a share page; 12 starter templates
Weaknesses
Only tested on macOS, per the README
Needs Node 22+, Chrome and FFmpeg; voiceover adds uv, Python 3.12 and whisper-cli
Quality depends on the model; author mainly uses Opus 5.5
Cannot generate live-action footage or hear the audio; README and docs are mainly in Chinese
Library of agent skills for spec-driven, verifiable autonomous coding loops
1.4k starsMITlast commit Oct 2026
A collection of SKILL.md folders for coding agents, covering exploring unknowns, writing specs, implementing them in slices, reviewing, and visual checks. Install with `npx skills add dzhng/skills` or copy a skill folder into `.claude/skills/`. It targets Claude Code, Codex, opencode, Cursor and other harnesses that support skills.
Strengths
Skills are standalone; each can be installed or copied alone via `--list` or `--only`
Covers the whole loop: explore-unknowns, write-spec, implement-spec, review, close-spec
Visual review skills compare screenshots against a baseline, with a diff script included
Works with any harness that reads skills; MIT licensed
Weaknesses
Author's personal library; workflow assumes his spec and choices.md conventions
Full loop relies on a harness `/goal` mode, as in Claude Code or Codex
Long unattended runs of days; the only evidence is one 3d 18m run
Some skills are niche, such as WebGPU renderer rules and Xiaohei illustrations
Reference library of Claude Code skills, hooks and agents with auto-activation
10k starsMITlast commit Jul 2026
A copy-into-your-project .claude/ directory holding 4 skills, 9 hooks, 8 agents and 4 slash commands, extracted from a TypeScript microservices codebase. A UserPromptSubmit hook reads skill-rules.json and prompts Claude to load matching skills, using regex by default or an optional AI classifier (Gemini, OpenAI, Anthropic, Ollama). A setup wizard (setup.ts) installs it and runs 8 health checks; Codex CLI is also wired up.
Strengths
Regex activation mode needs no API key and makes no network calls
Setup wizard copies files and runs 8 verification checks, plus /verify-setup
Same hooks and skills work under Codex CLI via an adapter shim
Skills split into SKILL.md plus resource files for progressive loading
Weaknesses
Not a working app; components must be copied and customized per project
Bundled skills target Express, Prisma, React and MUI v7 only
Hooks are bash scripts; Windows needs WSL2
Stop hooks like tsc-check need heavy customization; Codex lacks a Skill tool event
no GPU
Needs Node.js 18+, npm, jq (optional Stop hooks), bash (macOS, Linux or WSL2), Gemini/OpenAI/Anthropic API key or Ollama (optional AI mode)
Spec-driven development plugin for Antigravity and Claude Code agents
3.8k starsApache-2.0last commit Sep 2026
Conductor is a plugin that makes a coding agent follow a Context, Spec and Plan, Implement protocol. It writes product, tech-stack and workflow files plus per-track spec.md and plan.md into a conductor/ directory in the repo. Slash commands cover setup, new-track, implement, status, revert and review.
Strengths
Keeps project context as versioned markdown in the repo (product.md, tech-stack.md, workflow.md)
Git-aware revert works on tracks, phases and tasks instead of commit hashes
Plan is reviewed and approved before the agent writes code
Installs as a plugin on Antigravity and Claude Code; handles new and existing projects
Weaknesses
README warns the spec-driven approach increases token consumption, especially on larger projects
Supports only Antigravity and Claude Code; other agents are not documented
Protocol is defined in SKILL.md prompts, so adherence depends on the host agent
Interactive GUI dialogs only in hosts that support them; plain terminals get text menus
Plugin marketplace of Claude Code agents, commands, hooks and skills
3.7k starsMITlast commit Oct 2026
A Claude Code plugin marketplace hosting 117 agents, 175 slash commands, 28 hooks, 26 skills and 51 bundled plugins, installed with `/plugin marketplace add davepoon/buildwithclaude`. A web UI at buildwithclaude.com also indexes community plugins, MCP servers and other marketplaces. Files can alternatively be copied by hand into ~/.claude/agents and ~/.claude/commands.
Strengths
Installs through the native /plugin marketplace flow, per plugin or as all-in-one bundles
Claude Code plugin that blocks edits that violate test-driven development
2.4k starsMITlast commit Oct 2026
TDD Guard installs as a Claude Code plugin with hooks that check each edit. It blocks implementation written without a failing test, or code beyond what the current test requires, and tells the agent what to do instead. The README says it has grown into Probity, and new projects should start there. TDD Guard is still maintained.
Strengths
Supports Vitest, Jest, Storybook, pytest, PHPUnit, Go, Rust, RSpec and Minitest
Validation rules can be customized, and lint rules can drive refactoring
Can be toggled on and off mid-session
Installs through Claude Code plugin commands and a /tdd-guard:setup skill
Weaknesses
Superseded by Probity; README tells new projects to start there
Works only with Claude Code
Needs a test reporter configured for your framework
Hooks run with your user permissions
no GPU
Needs Node.js 22+, Claude Code, A supported test framework
Agent skill that extracts a design's tokens and style into JSON, then rebuilds it
2k starsMITlast commit Aug 2026
An Agent Skills package that analyzes screenshots, images or URLs and writes a JSON "Design DNA" profile covering design tokens, qualitative style and visual effects. A second phase generates self-contained HTML/CSS/JS from that JSON plus your content. Optional Node scripts run k-means color measurement and score the output against the reference by ΔE.
Strengths
Optional measure-colors and verify scripts give exact hexes and PASS/FAIL ΔE checks
DNA JSON can be committed, shared and reused across projects and agents
One npx command installs it to Cursor, Claude Code, Codex, Copilot and others
Structure, analyze and generate phases work alone or chained
Weaknesses
Color measurement scripts apply to image references, not live URLs
Default output is self-contained HTML/CSS/JS; other frameworks are not documented
Output quality depends on the host agent; README suggests a manual polish pass
No tagged releases
no GPU
Needs Node.js, npx, skills CLI, A host agent such as Cursor or Claude Code
gptcommit is a Rust CLI that installs a git prepare-commit-msg hook and fills in the commit message from the staged diff using an OpenAI model. It summarizes each changed file with up to four concurrent requests, then generates a title and summary, with optional conventional-commit prefixes. Settings are layered from a user TOML file, a per-repo file and GPTCOMMIT__* environment variables.
Strengths
Runs inside normal `git commit` via a prepare-commit-msg hook; generated text stays editable
Layered config (user, repo, env vars) with `gptcommit config` subcommands
Git-style `file_ignore` patterns skip lockfiles and minified files before anything is sent
Supports custom API base URL, proxy and model override; output in en, zh-cn, zh-tw, ja
Weaknesses
Only OpenAI-style APIs are documented; no local model support is described
Sends diff contents to the API, and costs grow with N+2 prompts per commit
Hook skips commits that use a template or amend unless `allow-amend` is set
A failed file summary stops generation instead of being skipped
no GPU
Needs OpenAI API key, git, Rust toolchain (cargo) or Homebrew to install
Agent Skill with reference docs for designing and auditing agent harnesses
2.4k starsMITlast commit Oct 2026
A provider-neutral Agent Skill, installed into Codex or Claude Code skill directories, that activates on agent-architecture questions. It supplies a SKILL.md and a references/ set covering MVP blueprints, tool permissions, context and compaction, evals, security, durable runtimes and voice agents, with worked example dialogues. It is documentation and guidance, not a runtime or library.
Strengths
Installs via `npx skills add` or a plain git clone into the skills directory
A Codex skill that turns a text prompt or reference image into a layered holographic card. A pipeline script builds an editable Blender scene (card.blend) and a responsive Three.js viewer with parallax, rainbow foil and sparkle effects. A second mode makes lenticular two-image cards that switch the whole face by viewing angle.
Strengths
Delivers an editable card.blend, four-layer PNG assets, card-config.json and a web viewer
Fetches a portable Blender into the project and checks its SHA-256
One command runs the pipeline: run_pipeline.py with holographic or lenticular mode
Rarity, number and text live in card-config.json; rerun the pipeline to update
Weaknesses
Designed to be driven by Codex; no standalone image generation described
Web shader output differs from Blender renders in color, glow and card back
README is mainly Chinese; English version is in a separate file
License not identified; companion gallery repo is private
Official GSAP skills that teach coding agents correct animation API usage
17k starsMITlast commit Apr 2026
Eight Agent Skills-format skill folders covering GSAP core tweens, timelines, ScrollTrigger, plugins, gsap.utils, React (useGSAP), performance, and Vue/Svelte lifecycle handling. Installs with `npx skills add` or the Claude Code plugin marketplace, and works with Cursor, Codex, Windsurf and others. Copilot gets a separate instructions file to copy by hand.
Strengths
Maintained by the GSAP authors (greensock), so API guidance tracks the library
Approval-gated agents and context files for OpenCode that follow your coding patterns
4.9k starsMITlast commit Oct 2026
OpenAgents Control adds markdown-defined agents (OpenAgent, OpenCoder, SystemBuilder) and subagents to the OpenCode CLI. A ContextScout subagent loads your project's coding standards from `.opencode/context/` before code is generated, and every file write or command requires approval. A BETA Claude Code plugin is also offered. The README says a rewrite is coming that will replace much of this version.
Strengths
Agents are plain markdown files you edit, with no compile step
Team patterns live in `.opencode/context/` and can be committed to the repo
Approval required before file edits, bash commands and subagent delegation
Installs with one curl script; also ships a Claude Code plugin (beta)
Weaknesses
Targets OpenCode v1; README warns a major rewrite will replace much of it
Sequential approval flow is slower than autonomous parallel agents
The 80% token reduction claim is given without a methodology
Requires OpenCode CLI, Bash 3.2+ and Git; Claude Code plugin is beta
no GPU
Needs OpenCode CLI, Bash 3.2+, Git
Models: Claude, GPT, Gemini, MiniMax, Local models
Claude skills that reverse-engineer, rebuild and launch a clone of an app
5.1k starsMITlast commit Oct 2026
Eleven Claude skills run in sequence: recon, architecture, design, build, backend, test, diff, entrepreneur, brand, launch and deploy. Each skill writes into a replica/ folder that the next one reads. Six standard-library Python tools handle image diffing, parity scoring, review ranking, contrast checks, a rebrand sweep and store listing limits.
Strengths
Covers the path from recon to deploy, each skill reading the previous one's output
Python tools use only the standard library and make no network calls
Review tool drops unlinked reviews, so printed quotes trace to a source URL
Sweep script blocks deploy while the original's name, domain or colors remain in code
Weaknesses
Results depend on app complexity; README says a spreadsheet engine is not a weeks-long job
Recon uses public pages, screenshots and your own account, not source or private APIs
Built for Claude; other chats get single SKILL.md modes without the Python tools
Trademark and legal risk stays with you; README says it is not legal advice
Agent skills for building native-feeling Expo and React Native apps
3k starsMITlast commit Sep 2026
Two agent skills installed via `npx skills`: one researches top-grossing mobile apps through the Appllama MCP, the other builds Expo / React Native screens to an Apple HIG bar. The design skill covers navigation, motion rules and a simulator loop that records whole flows. It works with Claude Code, Cursor, Codex and other agents that load skills.
Strengths
Design skill installs alone and works without the MCP connected
Concrete build rules: push vs replace, sheets, springs with velocity, no JS-thread animation
Simulator loop records and reviews full flows, not single screenshots
Plain skill directories; can be copied manually into .claude/skills/
Weaknesses
Research skill needs the Appllama MCP, which is part of the paid Pro plan
Each MCP call spends one credit; credits reset monthly
Build guidance targets Expo / React Native and iOS conventions only
No tagged release; the MCP backend is a hosted service
Claude Code plugin with four skills for a classify, verify, report workflow
2.3k starsMITlast commit Oct 2026
Packages a stepwise agent workflow as four skills: fable-method (classify the ask, define done, gather evidence, decide, act, verify, report), fable-loop (orchestrated subagent run), fable-judge (adversarial check of finished work) and fable-domain (generates domain adapters). Installs as a Claude Code plugin or Codex skills, with AGENTS.md for other harnesses. The repo includes eval logs from fifteen rounds and 260+ runs, self-described as smoke-test grade.
Strengths
Every rule is tied to a committed eval transcript and per-scenario case study
fable-judge re-runs claimed checks and diffs changes to catch false completion reports
Installs as a Claude Code plugin, Codex skills, or AGENTS.md for other agents
Eight domain adapters load lazily; eval nulls and failures are logged beside the wins
Weaknesses
Evals are 1-4 runs per cell with LLM judges; the README calls them smoke-test grade
No measured lift on ordinary small tasks with capable models
Haiku surfaced the skipped-deploy decision in only 1 of 12 runs
Seven of the eight domain adapters lack sourced references; medical has none
Agent skills documenting 228 Kotlin and Compose Multiplatform pitfalls, with a scanner
1.6k starsGPL-3.0last commit Oct 2026
Ten area skills in the agent-skills markdown format index 228 trap files on Kotlin, Compose Multiplatform and desktop JVM, mined from the author's SimpMusic app. An eleventh skill, footgun-scan, runs a Python script that matches 32 patterns against added lines in a diff and can fail a CI step. Installs as a Claude Code plugin or as editable files via the skills CLI.
Strengths
Area skills load an index first, so only 11 short descriptions sit in context each turn
Each trap pairs a failure mode with verification commands to run on your own tree
footgun-scan runs standalone with --base and --fail, usable as a CI step
Installs for Claude Code, Cursor, Codex, Copilot, Windsurf, Gemini via skills CLI
Weaknesses
Lessons come from one codebase (SimpMusic), so coverage skews to media, Room and Compose
Scanner has only 32 patterns; look hits need manual review of the trap
GPL-3.0 license applies to the whole repository
Eval results are self-reported by the author; no independent benchmark
no GPU
Needs Python 3, Claude Code or another agent supporting agent skills, git (for scanner --base)
Models: Claude Code, Cursor, Codex, Copilot, Windsurf
Agent skill that generates Lottie animations from text prompts
5.6k starsMITlast commit Jul 2026
Text-to-lottie is a skill installed with `npx skills add diffusionstudio/lottie` that lets Claude Code, Codex or another skills-capable coding agent write Lottie JSON animations. The agent sets up a workspace with a bundled player that loads scenes from `public/projects/<project>/<scene>/lottie.json` and live-updates as files change. The README includes a prompt guide and playback snippets for web, React Native, iOS, Android and Flutter.
Strengths
Works with any coding agent that supports skills, not tied to one vendor
Bundled player live-reloads and scrubs scenes while the agent edits the JSON
Output is plain Lottie JSON; README says it can be imported into After Effects
README gives playback examples for web, React Native, iOS, Android and Flutter
Weaknesses
Needs a separate coding agent and its model access; no model is bundled
Output quality depends on prompts; README advises supplying SVGs or screenshots
By default only a background color control is exposed unless requested
No hosted demo or separate docs site listed in the README
no GPU
Needs Claude Code, Codex or other skills-capable coding agent, Node.js (npx)
Agent skill collection for thesis, patents, diagrams, PPT and Office files
3k starsApache-2.0last commit Sep 2026
A set of SKILL.md folders you copy into the skills directory of Cursor, Claude Code, Codex, OpenCode or similar tools. Skills cover thesis and patent writing, drawio and Excalidraw diagrams, drawio-to-pptx slides, WeChat article writing, project docs, PDF and Office handling, and a five-step dev workflow. Docs and example prompts are in Chinese.
Strengths
Plain SKILL.md folders; agents trigger them from the description field without manual selection
CLI that fetches and caches package source for coding agents to read
3k starsApache-2.0last commit Sep 2026
opensrc is a Rust CLI that downloads the source code of a package from npm, PyPI, crates.io or GitHub and caches it locally. `opensrc path <package>` fetches on first use and returns the cached directory, so agents can run rg, cat or find against dependency source. The repo is a Turborepo monorepo that also contains a Next.js docs site.
Strengths
Returns a plain filesystem path, so it works with rg, cat, find and any agent
Covers npm, PyPI, crates.io and GitHub through one command
Caches after first fetch, so repeat lookups return the path instantly
Apache-2.0 licensed, with a CLI written in Rust
Weaknesses
Not an MCP server; agents must run shell commands themselves
README does not describe cache location, size limits or eviction
Registries beyond npm, PyPI, crates.io and GitHub are not listed
Labeled a Vercel Labs experiment, so stability guarantees are unknown
no GPU
Needs Node.js 24+ and pnpm 11 (development only), npm (install)
Terraform and OpenTofu guidance skill for AI coding agents
2.4k starscustom licenselast commit Jul 2026
An Agent Skills package that gives coding agents guidance on Terraform and OpenTofu: module structure, native tests vs Terratest, remote state backends, CI/CD workflows, and Trivy/Checkov scanning. Examples default to AWS, with Azure and GCP covered by the same guidance. It installs into Claude Code, Cursor, Copilot, Gemini CLI, OpenCode, Codex and Kiro, and can optionally use terraform-ls.
Strengths
Installs on many agent hosts: Claude Code, Cursor, Copilot, Gemini CLI, Codex, Kiro, OpenCode
Covers testing, modules, state, CI/CD and security scanning in one skill
Decision matrices and DO vs DON'T patterns, with version notes (use_lockfile 1.10+, write_only 1.11+)
Works without terraform-ls or the MCP server; both are optional
Weaknesses
Guidance only; it does not run or validate Terraform itself
AWS is the default in examples; Azure and GCP are mapped on request
Two marketplace names clash, so install must go through antonbabenko/agent-plugins
Install steps differ per host, and some are manual git clones
no GPU
Needs Terraform 1.0+ or OpenTofu 1.6+, AI agent with skill support, terraform-ls (optional), Terraform MCP server (optional)
Design skills and UI prompts that AI coding tools load via MCP
2.2k starscustom licenselast commit Oct 2026
TypeUI is a library of design skills and UI prompts, published as DESIGN.md and SKILL.md files, that you feed to AI coding tools to get more consistent interfaces. It is delivered through a hosted TypeUI MCP server and plugins, with setup guides for about 25 tools including Codex, Claude, Cursor, VS Code Copilot, Windsurf, Zed, Lovable, Replit and v0. Any model works if it accepts markdown or general prompting.
Strengths
Skills ship as plain DESIGN.md and SKILL.md files, usable with any markdown-aware tool
Per-tool install guides for about 25 hosts, including Cursor, Codex, Claude and Zed
Large catalog of named visual styles, such as Bento, Brutalism and Glassmorphism, with previews
Weaknesses
MCP setup points to a hosted TypeUI server; self-hosting is not described in the README
License is reported as NOASSERTION, so reuse terms are unclear
Output quality depends on the host model; the README gives no benchmarks
README is truncated here, so skill counts and pricing are unknown
no GPU
Needs TypeUI hosted MCP server
Models: Any provider that supports markdown files or prompting
Agent skill that teaches coding agents to build with the TypeSafe API
2.8k starsMITlast commit Oct 2026
A single skill, typesafe-ai, that guides coding agents in designing TypeSafe workflows, finding current docs and cookbooks, and composing typed judgments in code. It installs as a Claude Code plugin or through skills.sh for other agents. It is a client-side add-on for the TypeSafe service, which returns typed decisions and probabilities from System One models.
Strengths
Two install paths: Claude Code plugin marketplace and npx skills for other agents
Installs project-local by default, with -g for global install
MIT license, with the skill readable as plain SKILL.md
Weaknesses
Only useful with the TypeSafe service; pricing and API terms are not in the README
Contains one skill only
README does not list supported agents beyond Claude Code
No releases published
no GPU
Needs TypeSafe service, Claude Code or skills.sh-compatible agent, Node.js (npx)
Collection of agent skills for indie hackers: SEO, Reddit/X research, logos, domains
1.9k starsApache-2.0last commit Oct 2026
A set of ten SKILL.md-based skills for coding agents such as Claude Code, Cursor, Codex and Windsurf. They cover SEO/GEO optimization, Reddit, Twitter/X and Product Hunt search, domain price comparison, logo and banner creation, image generation with Gemini 3 Pro Image, and archiving session learnings. Install via the Claude Code plugin marketplace or `npx skills add`.
Strengths
Installs through Claude Code marketplace or `npx skills add` for 16+ agents
Each skill is a self-contained folder with SKILL.md, following the Agent Skills standard
Includes a template directory for writing and contributing new skills
Claude Code plugin with skills, subagents and provider settings templates
1.7k starsMITlast commit Sep 2026
A Claude Code plugin marketplace repo bundling 12 skills (deep-research, github-fix-issue, github-review-pr, image generation via Gemini and OpenAI, codex-skill, handoff and others) and six subagents. It also ships settings.json templates for providers such as GitHub Copilot, LiteLLM, DeepSeek, Qwen, Vertex AI, Azure and OpenRouter. The README is written in Chinese.
Strengths
Installs through the Claude Code plugin marketplace with two documented command styles
Settings templates for about 11 providers, including Copilot, LiteLLM, Azure and OpenRouter
Skills can be installed individually with npx skills
MIT licensed, with a script to update installed plugins
Weaknesses
README and skill descriptions are in Chinese only
settings.json is not installed by the plugin and must be set up by hand
npx skills installs may lag behind the repo, per the README
Manual install overwrites ~/.claude; image and Codex skills need API keys or CLIs
no GPU
Needs Claude Code, Node.js (npx), Python 3, GEMINI_API_KEY, OPENAI_API_KEY, Codex CLI, yt-dlp, copilot-gateway or LiteLLM (optional)
CLI that maps a codebase into compact context files for AI coding assistants
1.4k starsMITlast commit Jul 2026
codesight scans a project and writes a compact map of routes, database models, components, env vars and import-graph hot files to .codesight/. It also generates CLAUDE.md, .cursorrules, codex.md and AGENTS.md, a per-topic wiki, and can run as an MCP server with 14 tools. TypeScript gets AST parsing; other languages use regex detection.
Strengths
Runs with one npx command, no config, no API keys, zero dependencies
Detects 30+ frameworks and 14 ORM parsers across about 14 languages
Blast-radius command traces affected files, routes and models through the import graph
Generated wiki and map are committed to git and need no LLM calls
Weaknesses
Non-TypeScript languages use regex detection, so recall is lower than AST
Misses routes defined dynamically; README reports 38 of 43 on one project
Token-savings figures are estimates from formulas, not measured assistant sessions
Agent skill that replaces fear-based prompts with trust-based debugging guidance
1.4k starsMITlast commit Sep 2026
NoPUA is a prompt skill for coding agents. It activates after repeated failures or give-up phrases, or on /nopua. It replaces threat-style escalation with a staged method: switch perspective, zoom out, reset assumptions, then hand off with full context. The authors report a 9-scenario benchmark on Claude Sonnet 4.6, a three-way comparison against a fear-driven skill, and an arXiv paper.
Strengths
Auto-triggers on 2+ failures, give-up phrases and frustration, or manually via /nopua
Publishes benchmark methodology, raw data and an arXiv paper (2603.14373)
Lists support for Claude Code, Codex CLI, Cursor, Kiro, OpenCode and others
Loopy is an installable skill for Codex, Cursor and Claude Code that finds, audits, crafts, runs, debriefs, saves and publishes "loops": prompts with a verification check and a stopping rule. It works against the Loop Library website, a Cloudflare Worker-backed catalog of published loops that agents can also read as JSON and plain text. Saved loops go into a project-level LOOPS.md.
Strengths
Installs with one npx command for Codex, Cursor or Claude Code
Catalog is readable without the skill via catalog.json, catalog.txt and llms.txt
Run mode is bounded and returns a receipt with evidence and stopping reason
Saving and publishing require explicit approval of the exact preview
Weaknesses
Loop records are not in the repo; the catalog lives in a private Cloudflare database
Needs Node.js and npx for install
Recommendations and publishing depend on the hosted signals.forwardfuture.com site
Claude Code template for generating and executing feature specs called PRPs
13.9k starsMITlast commit Mar 2026
A file template for Claude Code that adds two custom slash commands, /generate-prp and /execute-prp. The first turns a feature request in INITIAL.md into a Product Requirements Prompt after researching the codebase and docs. The second implements it with test and lint validation loops. It also ships a CLAUDE.md rules file, a base PRP template and an examples/ folder convention.
Strengths
Small file-based setup: two command files, CLAUDE.md and templates, easy to read and adapt
PRPs include validation gates so the agent runs tests and linting and iterates
Includes a complete example PRP and a sample INITIAL.md
MIT licensed
Weaknesses
Written for Claude Code; other assistants would need the commands ported
README states it does not cover RAG or tool use
No releases and no automated checks of generated PRPs; quality depends on the model
The examples/ folder is empty by default; results depend on examples you add
Agent skill that redesigns GitHub READMEs with SVG heroes and Markdown
1.9k starsMITlast commit Oct 2026
An installable skill for coding agents that reads a repository, then restructures its README or produces only visual assets such as SVG heroes, section headers, workflow diagrams and optional GIFs. Visuals live in SVG while explanations, commands and links stay in Markdown. It has whole-README and asset-only modes plus a read-only audit, and does not commit or push without approval.
Strengths
Asset-only mode leaves README copy, order and links untouched
Keeps text in Markdown so the page stays searchable and copyable
Hybrid SVG plus AI-generated cutouts, with written production guides
Read-only audit mode; commits, pushes and PRs need explicit approval
Weaknesses
Output quality depends on the host agent; no evaluation or benchmarks given
PNG/WebP or AI image generation needs separate rendering or image tools
CLI that installs SKILL.md skills and lists them in AGENTS.md
10.8k starscustom licenselast commit Jan 2026
OpenSkills installs Anthropic-format SKILL.md skills from GitHub, private git repos or local paths into .claude/skills or .agent/skills. It writes the same <available_skills> XML that Claude Code uses into AGENTS.md, and agents load a skill on demand with `npx openskills read <name>`. This lets Cursor, Windsurf, Aider, Codex and other agents that read AGENTS.md use the same skills.
Strengths
Same SKILL.md format and folder layout as Claude Code, so skills are portable
Skills load on demand through a shell command, keeping context small
Installs from GitHub, private git URLs or local paths; project-local by default
No server or MCP support needed; runs via npx
Weaknesses
Agents must run shell commands to load skills; no native Skill tool
Requires Node.js 20.6+ and Git
Skill sourcing is limited to git repos and local paths; no separate registry
Detected license is NOASSERTION, though the README states Apache 2.0
An agent skill that has the model generate a target screenshot with image generation, build a game or scene toward it, then have a separate critic agent compare the live screenshot to the target. The loop repeats until the critic is satisfied, and can optionally generate a new target. The README says it is tested only with GPT-6 Astra in Codex.
Strengths
Closed loop with a separate critic agent comparing screenshots to a target image
Installs with `npx skills add` or by cloning into a skills directory
Works with built-in or API image generation (e.g. a Gemini key)
Live demo and example prompt show a Three.js isometric scene
Weaknesses
Tested only with GPT-6 Astra in Codex; other models are untested
Needs an agent with image generation and vision input
Subagents strongly preferred, so results may degrade without them
Blender needed for custom 3D modeling; no resource requirements documented
Needs Image generation (built-in or API, e.g. Gemini), Vision-capable agent, Blender (optional)
205 copy-paste SOUL.md agent templates for OpenClaw
4k starsMITlast commit May 2026
A curated repository of SOUL.md persona files for OpenClaw agents, grouped into 24 categories such as productivity, development, marketing, finance and security. A template is registered with `openclaw agents add`, and a small Node quickstart (`bot.js`) can run one. The README also lists use cases, MCP servers, model configs and a machine-readable agents.json, and links to the paid CrewClaw service for deploy packages.
Strengths
205 templates across 24 categories, each a plain SOUL.md file
Machine-readable agents.json index alongside the templates
MIT license; Node quickstart runs a template with npm install and node bot.js
Extra material: model configs, use cases, MCP server list, troubleshooting
Weaknesses
Only useful with the OpenClaw runtime; templates are not standalone agents
README promotes the commercial CrewClaw deploy service throughout
README counts conflict: says 205 agents but also 'All 100 agents'
Templates are prompt files only; no tests or per-template versioning shown
A catalog of 67 named visual styles (Brutalism, Glassmorphism, Material, Neobrutalism and others), each packaged as a folder with SKILL.md for agent instructions and DESIGN.md for human-readable rationale. A skill is pulled into a project with `npx typeui.sh pull <name>`, and previews are hosted on typeui.sh. Targets Claude Code, Cursor, Codex and similar tools.
Strengths
One-command install per skill via npx typeui.sh pull
Each skill pairs agent rules (SKILL.md) with human-readable DESIGN.md
SKILL.md covers tokens, component rules, accessibility constraints and quality gates
MIT license; visual previews for every style on typeui.sh
Weaknesses
Installation depends on the typeui.sh CLI and registry, not plain git
Only 67 predefined styles; customization process is unknown from the README
No stated evaluation of output quality across agents
README shown is truncated, so the full style list is unverified
Claude Code skill that delegates tasks to the Codex CLI
1.5k starsMITlast commit Oct 2026
A Claude Code plugin and skill that lets Claude run `codex exec` and resume Codex sessions for code analysis, refactoring and editing. Claude asks for the model and reasoning effort, picks a sandbox mode (read-only for analysis), and summarizes the output. Stderr thinking tokens are suppressed with `2>/dev/null` by default to save context.
Strengths
Installs as a Claude Code plugin via marketplace, or as a standalone skill copy
Defaults to read-only sandbox for analysis tasks
Hides Codex thinking tokens by default to keep context small
Asks for model and reasoning effort unless the prompt already sets them
Weaknesses
Requires the separate `codex` CLI installed with valid credentials
Legacy models such as gpt-5.5 are rejected for ChatGPT-account sign-ins
Operational details live in SKILL.md, not the README
Agent skill that checks iOS/macOS projects for App Store rejection causes
1.4k starsMITlast commit Oct 2026
An agent skill, installed with `npx skills add`, that loads an app-type checklist, pulls App Store metadata with the `asc` CLI, and scans the Xcode project against markdown rejection rules. Rules cover review notes, subscription terms, privacy manifests, third-party AI consent and unused entitlements. Findings are reported with severity, affected files and fixes.
Strengths
Rules are plain markdown files, each mapped to an Apple guideline number
Ten app-type checklists: subscriptions, kids, health, games, AI, VPN, crypto and others
Reports severity and affected files, then applies fixes and re-runs checks
Reads metadata through the asc CLI, so it checks what is in App Store Connect
Weaknesses
Depends on the asc CLI and its canonical JSON metadata layout; fastlane layouts need adapting
Only 14 rules documented, so coverage of Apple guidelines is partial
Checks are driven by the agent reading instructions, so results may vary between runs
Plugin that runs parallel optimization experiments on a codebase via coding agents
1.5k starsApache-2.0last commit Oct 2026
evo installs as a plugin for coding-agent hosts such as Claude Code, Codex and Cursor. Its discover skill finds a benchmark and gates, then optimize runs a tree search of experiments, each in its own git worktree, keeping changes that improve the score. A local dashboard (default port 8080) tracks experiments, and runs can go to Modal, E2B, Daytona, AWS, Azure or SSH.
Strengths
Gates discard experiments that beat the score but fail tests or a held-out floor
Tree search with five frontier strategies, including argmax, top_k and pareto_per_task
Supports eight agent hosts and eight execution backends, local and remote
Dashboard for monitoring experiments, frontier strategy and backend selection
Weaknesses
Sends anonymous telemetry by default; opt out with evo telemetry off
CLI, skills and hooks must stay version-matched; drift caused silent failures earlier
Needs a separate host agent CLI and its model access; cost is not documented
Remote backends require extra installs and provider accounts; setup steps are not covered
no GPU
Needs uv, git, Claude Code, Codex, Cursor, Kimi, OpenClaw, Hermes, Opencode or Pi, Modal, E2B, Daytona, AWS, Azure or SSH (optional, remote runs)
Hook-enforced agent protocol pack for Cursor, Claude Code and other coding assistants
2.3k starsMITlast commit Jun 2026
Harmonist installs 193 markdown agent definitions plus IDE hooks into a project. The stop hook blocks a code-changing turn until a reviewer such as qa-verifier has run and session-handoff.md is updated, with a retry cap of 3. Memory writes go through a schema-validated, secret-scanning CLI, and shipped files are checked against a SHA-256 manifest before install.
Strengths
Stop hook gates turns on reviewer runs and memory updates, not on prompt wording
Pure Python 3.9+ stdlib; no npm, Docker or database; runs on native Windows
MANIFEST.sha256 check refuses tampered agent files during upgrade and extras install
Memory entries are schema-validated and scanned for about 30 secret patterns
Weaknesses
Primary integration targets Cursor; other assistants go through adapters of unstated maturity
PROTOCOL-SKIP lets the model bypass the gate; only skip-rate warnings check it
Needs an assistant that supports subagent dispatch and hooks
Integration is done by an AI following a prompt, so results depend on the model
no GPU
Needs Python 3.9+, Git, Bash 3.2+ (optional), AI coding assistant with subagent support
Models: Cursor, Claude Code, Copilot, Windsurf, Aider
Proxy that lets the Claude Code CLI use OpenAI-compatible APIs
2.8k starsMITlast commit Mar 2026
A Python server that exposes the Anthropic /v1/messages endpoint and translates requests into OpenAI-style API calls, so the Claude Code CLI can run against OpenAI, Azure OpenAI, Ollama or any compatible provider. Requests for haiku, sonnet and opus models are mapped to configurable SMALL_MODEL, MIDDLE_MODEL and BIG_MODEL values. Tool calls, SSE streaming and base64 images are converted.
Strengths
Maps haiku, sonnet and opus requests to separate backend models via env vars
Converts tool use, SSE streaming and base64 image input
Works with OpenAI, Azure OpenAI, Ollama or any OpenAI-compatible base URL
Optional ANTHROPIC_API_KEY setting restricts which clients can use the proxy
Weaknesses
MAX_TOKENS_LIMIT defaults to 4096, which may cut long Claude Code outputs
No tagged releases
Only one test script is mentioned; coverage is unknown
README describes MIDDLE_MODEL inconsistently (opus in one place, sonnet in another)
no GPU
Docker + Compose
Needs OpenAI-compatible API provider, Claude Code CLI
Claude Code slash-command workflows that delegate coding to Codex, Gemini or OpenCode
2.8k starsAGPL-3.0last commit May 2026
Installs Claude Code modules (do, omo, bmad, requirements, essentials, sparv) as slash commands, skills, agents and hooks under ~/.claude. Claude Code orchestrates, while a Go binary, codeagent-wrapper, hands code editing and test runs to the Codex, Claude, Gemini or OpenCode CLI. Installed with an npx installer that has --list and --update options.
Strengths
Routes execution to Codex, Claude, Gemini or OpenCode through one wrapper binary
Modules can be enabled or disabled individually in config.json
npx installer supports --list, --update and a custom --install-dir
Includes a 5-phase /do workflow plus 11 basic commands such as /code and /debug
Weaknesses
Requires the chosen backend CLIs installed and supporting specific flags
AGPL-3.0 license; commercial use without AGPL terms needs a separate agreement
Several overlapping workflows (do, omo, bmad, requirements) make the choice unclear
Known issues: Gemini cannot read .gitignore files, Codex may need approval_policy set
CLI that moves a coding session between 16 AI coding tools
1.6k starsMITlast commit May 2026
continues reads local session files from 16 AI coding CLIs and editors, including Claude Code, Codex, Copilot CLI, Gemini CLI, Cursor and Cline. It builds a handoff document with recent messages, file changes and tool activity, then launches a different tool with it. It also does native same-tool resume, session listing, inspection and bulk export.
Strengths
Any-to-any handoff across 16 tools, 240 source and destination pairs
Read-only access to session files; index cached at ~/.continues/sessions.jsonl
Four verbosity presets plus .continues.yml for per-project defaults
Scriptable: list with JSON/JSONL output, inspect, and dump to markdown or JSON
Weaknesses
Requires Node.js 22.5+ for built-in node:sqlite
Newer parsers were adapted from another project's adapters; fidelity unverified
Handoff is a summary document, not full history; detail depends on preset
No tagged releases
no GPU
Needs Node.js 22.5+, At least one supported AI coding tool installed
Models: Claude Code, Codex, GitHub Copilot CLI, Gemini CLI, Cursor
Plan-Apply-Unify workflow commands for structured development in Claude Code
1.3k starsMITlast commit Aug 2026
PAUL installs 26 slash commands into Claude Code that enforce a plan, apply, unify loop. Plans carry Given/When/Then acceptance criteria, each task is checked against the spec after execution, and project state lives in a .paul/ directory (STATE.md, ROADMAP.md, per-phase PLAN and SUMMARY files). Implementation stays in-session; subagents are used only for research and discovery.
Strengths
State persists in plain markdown files under .paul/, so sessions resume with /paul:resume
Each task gets an Execute/Qualify check against its acceptance criteria, with four status codes
Installs with one npx command, globally or per project, on Mac, Windows and Linux
Scope-adaptive plans (quick-fix, standard, complex) with a coherence check before approval
Weaknesses
Works only with Claude Code; no other coding agent is mentioned
Graph, cost attribution and rule injection need BASE v2, linked from a community site
Mandatory UNIFY step and per-task qualification add process overhead to small changes
Claims such as subagent output being ~70% quality are stated without benchmarks
Claude Code skills, sub-agents and slash commands, installable as plugins
2.6k starsMITlast commit Mar 2026
A marketplace repo of about 40 skills in the Agent Skills format, plus six sub-agents and seven slash commands. Skills cover diagrams and docs, React and MUI, database schema design, git commits, Jira, Datadog and writing style. Each item installs separately through the Claude Code plugin marketplace, or all at once with `npx skills add`.
Strengths
Each skill, agent and command installs individually as its own plugin
Skills also install in Codex, Cursor and AdaL via `npx skills add`
Each skill ships a SKILL.md for the agent and a README with examples
Skills can be copied by hand into ~/.claude/skills
Weaknesses
Agents and slash commands work only in Claude Code CLI
Opinionated selection from a single maintainer; no tests or quality checks described
Some skills need outside tools or services (Codex, Gemini, Perplexity, Jira, Datadog)
Package manager that installs and translates agent skills across 46 coding agents
1.6k starsApache-2.0last commit Jun 2026
SkillKit is a TypeScript CLI that installs skills from GitHub, GitLab, gists or local paths and translates them between agent formats such as SKILL.md and Cursor .mdc. It recommends skills from your repo's stack, scans skills for security issues, and can serve skill search over REST or MCP. Extras include session memory, AI skill generation and a terminal UI.
Strengths
Translates one skill into formats for 46 agents, including Claude Code, Cursor, Codex and Copilot
Installs from GitHub, GitLab, gists and local paths; runs via npx with no setup
Slim install with --omit=optional: 118 packages versus 418 for the full install
Exposes skill search over REST on port 3737, an MCP server and a Python client
Weaknesses
Broad scope: memory, mesh, messaging, issue planner and workflows all bundled in one tool
Skill quality depends on third-party sources; the 400K+ skills count comes from the README
AI skill generation needs an external model provider such as Claude, GPT-4, Gemini or Ollama
Install demo detects 32 agents while the README claims 46 supported
Review-pass skills that check AI-written code, tests and docs before merge
1.3k starsMITlast commit Jul 2026
A package of five Markdown skills for coding agents, installed with the Skills CLI (`npx skills add`). Each guard is run on the agent's diff as a second pass: clean-code-guard, test-guard, docs-guard, plus wp-guard and woo-guard for WordPress and WooCommerce code. It works with Claude Code, Codex, Cursor, OpenCode and other agents the CLI supports.
Strengths
Content is Markdown plus YAML metadata; README states no scripts, network calls or credentials
Installs per skill or as a whole package, per agent or globally
Deeper references load only when needed, keeping SKILL.md small
Covers WordPress and WooCommerce specifics such as HPOS, nonces and escaping
Weaknesses
Skills install as copies and need `npx skills update` to pick up changes
Works as a reactive pass; the agent must be told to invoke each guard
Two of five guards apply only to WordPress or WooCommerce code
No tagged release; no catch-rate measurements in the README
Claude Code add-on for session handoffs, memory, and a library of skills and agents
3.9k starsMITlast commit Jan 2026
Installs 109 skills, 32 agents and 30 hooks into Claude Code's ~/.claude config. Hooks write YAML handoffs before compaction and recall learnings from a PostgreSQL+pgvector store, so sessions resume with state preserved. A TLDR analysis layer (AST, call graph, CFG, DFG, slicing) with a FAISS index is meant to cut tokens spent reading code.
Strengths
Hooks write YAML handoffs before compaction and reload ledger, handoff and memory at session start
Memory in PostgreSQL with pgvector; remote Postgres supported via CONTINUOUS_CLAUDE_DB_URL
Hook-driven skill suggestions for natural-language prompts, plus workflows like /fix, /build, /tdd
Uninstall mode archives ~/.claude and restores the earlier backup
Weaknesses
Requires Docker for PostgreSQL, plus Python 3.11+ and uv; 12-step setup wizard
Tied to Claude Code; no other coding agent is supported
Large footprint of 109 skills and 30 hooks that modifies ~/.claude
No tagged releases; the 95% token-savings figure is the README's own claim
Compose
Needs Claude Code CLI, Python 3.11+, uv, Docker, PostgreSQL with pgvector
Claude Code commands, skills and auditor subagents for planning and debugging
2k starsMITlast commit Apr 2026
A Claude Code plugin bundling 27 slash commands, 9 skills and 3 auditor subagents. It covers meta-prompting (/create-prompt, /run-prompt), hierarchical project planning (create-plans), todo capture, context handoff and a debugging methodology. Other skills generate new skills, hooks, subagents and MCP servers. It installs through the Claude Code plugin marketplace or by copying files into ~/.claude.
Strengths
Installs as a Claude Code plugin with two commands, or by copying files into ~/.claude
Auditor subagents review skills, commands and subagent configs for structure and compliance
create-plans skill turns a brief into ROADMAP and PLAN.md files that Claude executes
Includes MCP server generation for Python and TypeScript, and a Ralph loop setup skill
Weaknesses
Works only with Claude Code; the OpenCode port is a separate community repo
No release tags; versioning is by commit only
Thinking-model commands are prompt templates; no tests or benchmarks are shown
Domain expertise skills are not included; you must create them yourself
MCP client plugin for Neovim that connects MCP servers to chat plugins
1.8k starsMITlast commit Jan 2026
mcphub.nvim manages MCP servers from inside Neovim and exposes their tools, resources and prompts to Avante.nvim, CodeCompanion.nvim and CopilotChat.nvim. It supports STDIO, Streamable-HTTP and SSE transports, OAuth with PKCE, and a marketplace for installing servers. Servers can also be written natively in Lua.
Strengths
Works with Avante, CodeCompanion and CopilotChat; MCP prompts become slash commands
Supports STDIO, Streamable-HTTP and SSE transports, plus OAuth with PKCE and header auth
Reads VS Code style config, JSON5, and project-local configs merged with global ones
Multiple Neovim instances stay in sync; Lua-native MCP servers and a dev mode with hot reload
Weaknesses
MCP roots and sampling are not supported
Tests are listed as unchecked in the README TODO
#variables support in Avante is not implemented
Last release 2025-07-31; the README gives no install steps, only a docs site
no GPU
Needs Neovim, Avante.nvim, CodeCompanion.nvim or CopilotChat.nvim
Claude Code skills that build Obsidian study vaults and quiz you on them
1.4k starsMITlast commit Feb 2026
Two Claude Code skills. /tutor-setup converts PDF, TXT, MD, HTML and EPUB files, or a source code project, into an Obsidian StudyVault with linked notes, dashboards and practice questions. /tutor then runs quiz rounds of 4 multiple-choice questions and tracks proficiency per concept, updating the vault files after each round.
Strengths
Auto-detects document or codebase mode from project markers like package.json or go.mod
Codebase mode generates onboarding notes, flow traces and exercises for new developers
Per-concept tracking stores attempts, correct count and error notes to drive drill sessions
Installs with one npx skills command or a manual install.sh
Weaknesses
Requires the Claude Code CLI; other agents are only implied via npx skills
Vault format and dashboards are built around Obsidian, though it is only recommended
No release published; no tests or output quality benchmarks described
Quiz format is fixed at 4 questions per round with 4 options each
no GPU
Needs Claude Code CLI, Obsidian (recommended), npx skills (optional installer)
Security skill that steers AI coding agents away from common web vulnerabilities
1.4k starsApache-2.0last commit Feb 2026
A skill folder for coding agents that steers generated code away from common web vulnerabilities. It covers access control, client-side, server-side, authentication and API issues such as IDOR, XSS, SSRF, SQL injection and JWT handling, with checklists per class. Install by cloning the repo into the skills directory of Claude Code, Cursor, Codex, GitHub Copilot or Antigravity.
Strengths
Installs by cloning into a skills directory; no runtime or services needed
Documents install paths for five agents: Claude Code, Cursor, Codex, Copilot, Antigravity
CLI that runs coding agents in sandboxes with secret redaction
1.9k starsMITlast commit Nov 2025
VibeKit is an npm CLI that wraps coding agents such as Claude Code, Gemini CLI, Grok CLI, Codex CLI and OpenCode, running their output in isolated Docker containers. It redacts secrets and API keys from completions and records logs, traces and metrics of agent operations. Separate packages, @vibe-kit/sdk and @vibe-kit/auth, are provided for embedding it in applications.
Strengths
Wraps several agent CLIs (Claude Code, Gemini, Codex, Grok, OpenCode) behind one command
Local Docker sandboxing; README says no cloud dependency is required
Redacts secrets and API keys from agent output
Separate SDK package for running agents in sandboxes from application code
Weaknesses
Needs Docker locally for the sandbox; the README does not list setup steps
Redaction rules and detection coverage are not documented in the README
Last release was 2025-08-04, while the last commit is 2025-11-10
Where observability data is stored or viewed is not described in the README
no GPU
Needs Docker, Node.js
Models: Claude Code, Gemini CLI, Grok CLI, Codex CLI, OpenCode
Pseudocode language for writing structured prompts that LLMs interpret directly
1.4k starsMITlast commit Dec 2025
SudoLang is a pseudocode language for prompting LLMs, with interfaces, constraints, /commands, pattern matching and a |> function composition operator. Programs are plain text that models such as ChatGPT, Claude, Gemini and Llama interpret without a pasted spec. The repo holds the language documentation, examples and a VS Code/Cursor extension (.vsix) for highlighting .sudo, .sudo.md and .mdc files.
Strengths
Works across ChatGPT, Claude, Gemini and Llama without pasting a spec
Includes a VS Code/Cursor extension with highlighting for .sudo, .sudo.md and .mdc
Docs, examples folder and several articles and talks for learning
MIT licensed; README claims 20-30% fewer tokens than equivalent natural-language prompts
Weaknesses
No runtime or interpreter; behavior depends on how each model reads the prompt
Token savings and reasoning gains are claims in the README, not benchmarked here
Extension installs from a local .vsix after cloning, not from a marketplace
No tagged releases; last commit 2025-12-23
no GPU
Needs VS Code or Cursor (optional, for the extension)
Models: ChatGPT, Bing Chat, Anthropic Claude, Google Gemini/Gemma, Meta Llama