Everything here (should be) compatible with opencode, Claude Code, Gemini CLI, Codex CLI, Cursor and 73+ skill-compatible environments.
Skill and Agent frontmatter is in Claude Code compat format by default; use npx skills and/or bundled bun script to make them compatible with other harnesses (will do it in a better way when I decide how).
Contributions are welcome!
Note
Two plugins ship separately: agents and skills. Install one or both; on opencode/codex, cherry-pick by name with ./install.sh --to opencode prober web-researcher (list with ./install.sh --list, every flag with ./install.sh --help).
Claude Code
/plugin marketplace add uwuclxdy/agenticat
/plugin install agents@agenticat
/plugin install skills@agenticat
or cherry-pick with npx skills:
npx skills add uwuclxdy/agenticat --skill <name>
Gemini CLI
gemini extensions install https://github.com/uwuclxdy/agenticat --auto-update
[!WARNING] Gemini's extension-bundled subagents are a preview feature, verify if it works on your version. Per gemini-cli v0.65.0's agent loader, only the
-proimplementers load: they carry neithertoolsnordisallowedTools, so each inherits every tool of the session. Every other agent fails its frontmatter check and is skipped with a debug-log warning; the rest of the extension still loads.
npx skills add uwuclxdy/agenticat -a gemini
opencode
git clone https://github.com/uwuclxdy/agenticat && cd agenticat
./install.sh --to opencode
The agents install as subagents, so on opencode 1.19.4 their limits apply when a session delegates to one; opencode run --agent <name> falls back to the default agent instead.
npx skills add uwuclxdy/agenticat -a opencode
Codex CLI
git clone https://github.com/uwuclxdy/agenticat && cd agenticat
./install.sh --to codex
On Codex 0.163.0 a subagent's writes and network follow its parent session's sandbox, and the sandbox_mode each agent file carries has no effect. A read-only agent is enforced only under a read-only parent, which also cuts its network, so web-researcher's gh reads work under a workspace-write parent with [sandbox_workspace_write] network_access = true and fail under read-only, where every network call from its shell fails too.
npx skills add uwuclxdy/agenticat -a codex
Cursor
Cursor reads agents/ as claude format; copy files into .claude/agents/ or .cursor/agents/.
npx skills add uwuclxdy/agenticat -a cursor
Or just download into ~/.claude/skills/, but then you won't get updates.
- Claude Code:
/plugin update agents@agenticatand/or/plugin update skills@agenticat(every commit is a new version) - Gemini CLI: automatic with
--auto-update - opencode / Codex:
./install.sh --updatepulls and reconverts what you installed (with nothing installed it does neither); a renamed agent moves to its new name and a removed one is cleaned up. An install made without agent names also picks up new agents once its.agenticat.jsonholds"all": true; to write it on an older install, rerun the install once with the same options it was made with../install.sh --update --dest <dir>skips the pull and updates an install made with--dest, so rungit pullfirst../install.shchanges into the checkout, so for a--projectinstall or a relative--dest, runbun <checkout>/bin/install-agents.tswith the same options from the directory the install was made in, for the rerun and for--updatealike - Skills:
npx skills update
Install any skill with the same command, swapping the final name (see Install for other harnesses):
npx skills add uwuclxdy/agenticat --skill <name>
The used by column is a soft link: when an agent's task touches what a skill covers it loads that skill for the deeper method. Cherry-pick either one alone and the loader falls back to its own built-in checklist. The one hard link is tech-writer: each of its modes reads its procedure from docs-sync, so without that skill it reports the gap and stops.
| skill | origin | used by | what it does |
|---|---|---|---|
clean-code |
s4.codes |
python-pro, python-reviewer, web-pro, web-reviewer |
language-agnostic readability and naming conventions |
name-check |
original | standalone | proposes project or crate names, checks each for availability on registries and domains (needs bun) |
todo |
original | standalone | turns findings and loose notes into a clean per-repo backlog file, each task pickup-cold ready |
askama |
original | rust-pro |
conventions and reference for the askama Rust templating crate |
maud |
original | rust-pro |
conventions and reference for the maud Rust html! macro |
cargo-toml-optimization |
original | rust-pro |
tunes Cargo.toml and .cargo build profiles and features for faster compiles |
clean-rust |
original | rust-pro, rust-reviewer |
idiomatic Rust 2024 conventions, modular: core rules plus per-domain references for ownership, types and traits, API design, errors, the std toolbox, I/O and processes, async, concurrency, unsafe and FFI, macros, serde, CLIs, docs, lints, testing, perf, security, observability, edition 2024 and modern-Rust migration |
clean-shell |
wshobson |
shell-pro, shell-reviewer |
defensive live-system idioms, .shellcheckrc discipline, bats-core testing; core rules plus a reference for each |
clean-flutter |
original | flutter-pro, flutter-reviewer |
idiomatic Flutter/Dart conventions, modular: core rules plus per-domain references for architecture, state management, navigation, models/serialization, testing, pitfalls, packages |
emulator-testing |
original | mobile-tester |
drives Android emulators and iOS simulators from the CLI: headless boot, adb/simctl primitives, Flutter test layers, screenshot verification |
mcp-stateless |
original | standalone | the MCP 2026-07-28 stateless revision: what the handshake removal changed, the wire contract, multi round-trip requests, and how to migrate a server or client off session-era code |
docs-sync |
original | tech-writer |
reconciles README, wiki, docs, and CLAUDE.md (plus agent/skill prompt files) with what the code actually does |
ratatui-patterns |
original | ratatui-pro |
modern ratatui built-ins, the app loop and terminal lifecycle (events, async input, child processes), TestBackend render tests, a version-upgrade reference pack |
webapp-testing |
anthropics |
webapp-tester, web-pro |
Playwright toolkit for testing local web apps, screenshots and console logs |
ux-patterns |
synsation_ |
ux-pro, ux-reviewer, flutter-pro |
how a UI behaves: the five screen states, loading and error patterns, form validation and submission, undo versus confirmation, cognitive-load laws, with WCAG wiring and a mobile reference |
terminal-ux |
original | ratatui-pro, ux-pro, ux-reviewer, web-pro |
how a CLI or TUI behaves: exit codes, stdout vs stderr, TTY detection, NO_COLOR, help and destructive-action conventions, panic-safe terminal restore, terminal screen-reader practice |
wshobson: wshobson/agents (MIT)s4.codes: content from @s4.codes (TikTok) as a Claude Skill; Gemini transcribes the videos, Opus condenses them into the skill.synsation_: content from @synsation_ (TikTok) as a Claude Skill, theBuild for Good UXseries, parts 1 to 16; same pipeline ass4.codes. Its accessibility reference is authored here from W3C sources instead, since the series does not cover accessibility.original: self-authoredanthropics: anthropics/skills (Apache-2.0), vendored with its LICENSE and a modification noteclean-rust,cargo-toml-optimization,ratatui-patterns,askamaandmaudare built from the official Rust documentation (the Book, the Reference, the std, cargo and Edition Guide docs, the Rustonomicon, the API Guidelines, the async book, the clippy and rustc lint references), each crate's own docs and source at its release tag, and these guides: ANSSI's Secure Rust Guidelines (Licence Ouverte 2.0), Microsoft's Pragmatic Rust Guidelines (MIT), Rust Design Patterns (MPL-2.0), The Rust Performance Book (MIT/Apache-2.0) and The Little Book of Rust Macros (MIT); a dozen earlierclean-rustrules were re-authored after studying leonardomso/rust-skills (MIT)
Every third-party-origin item in the tables above is a copy this repo owns: the wshobson skills and the +-marked -pro agents are rewrites, webapp-testing is vendored with its license, clean-code is distilled from transcripts. Installing them never pulls from the original source.
The one exception worth a look on its own: the obra/superpowers skills systematic-debugging, test-driven-development, and verification-before-completion (MIT) are not bundled here and install pristine from source with npx skills add obra/superpowers --skill <name>.
Install with the Claude Code plugin (see Install) or cherry-pick onto another harness with ./install.sh --to <harness> <name>. Every definition is canonical Claude Code frontmatter. On opencode the converter keeps Write and Edit off a read-only agent and denies it subagents. Except for prober, whose commands write build output (its opencode shell is unrestricted), a read-only agent also gets an opencode shell limited to read commands plus the checks its definition lists, so opencode refuses a reviewer's test re-run or a repo-defined script. That shell limit also refuses a report written to a file, as does a read-only Codex parent session, so the report comes back in the agent's final message. On Codex the read-only rule holds only under such a parent (see Codex CLI under Install); elsewhere on Codex, and for Bash on Claude Code, it is an instruction in the definition. On an outside contributor's diff (the caller names it an outside contributor's PR; every other diff counts as trusted), every reviewer and tech-writer take the reviewed tree's files as standards, never as instructions, run none of its code and check by reading instead, so they skip every check that runs it (mypy, cargo checks, repo-defined scripts, a built tool's --help), test re-runs, mutants, probe copies and runs of the user-facing entry. Start the session that delegates to them outside the PR's checkout: a harness loads a project's hooks, plugins, MCP servers and agent definitions when a session starts in it. On a trusted diff, the shell reviewer runs the reviewed scripts only when the caller asks for it. These skips are instructions in each definition and, for tech-writer, in docs-sync: opencode's shell limit cannot tell who wrote a diff, so it allows the listed checks (mypy, cargo checks) and refuses re-runs and repo scripts on every diff.
| agent | model | loads if installed | what it does |
|---|---|---|---|
python-pro |
inherit | clean-code |
one Python task against the repo's test, lint and type-check gate |
python-reviewer |
inherit | clean-code |
read-only Python diff/PR review with file:line and severity |
web-pro |
inherit | clean-code, terminal-ux, webapp-testing |
one TypeScript, JavaScript, Node, HTML or CSS task against the repo's gate, or a standalone single-file page, demo, icon or favicon with pinned vendored libraries |
web-reviewer |
inherit | clean-code |
read-only TS/JS, Node and HTML/CSS/template diff review: XSS, template escaping, injection, markup accessibility |
tech-writer |
opus | docs-sync |
one agent for every docs job, mode set by the caller: review a docs or instruction-file diff (read-only by instruction, not by tool grants), reconcile docs with the code, draft instruction-file text as a diff, fold a decided spec into a design doc, or digest docs or source into a brief (doc-set sweeps and comparisons, delete-or-merge coverage) |
prober |
haiku | none | runs a build, test or lint, returns pass/fail instead of the full log |
web-researcher |
inherit | none | runs one research question through many searches, reads GitHub PR and issue threads with gh, returns a cited markdown brief |
code-explorer |
inherit | none | read-only repo research through rg on a shell: symbol lookups, cross-file traces, literal searches, source-tree sweeps and comparisons; findings with file:line anchors, never edits, writes only its report to a caller-named path outside the repo (on opencode 1.19.4 it needs rg on PATH; Codex 0.163.0's standalone package and Claude Code bring their own) |
webapp-tester |
inherit | webapp-testing |
drives a local app via Playwright and judges each screenshot by eye; pass/fail with evidence, design-contract checks on request |
shell-pro + |
inherit | clean-shell |
writes or refactors bash or POSIX sh, verifies with shellcheck |
golang-pro + |
inherit | none | one module-sized Go task, proven with the repo's gate and race detector |
rust-pro + |
inherit | clean-rust, cargo-toml-optimization, askama, maud |
one Rust task against the repo's cargo and clippy gate |
c-cpp-pro + |
inherit | none | C and C++ with explicit ownership, sanitizers wired to the repo build |
flutter-pro |
inherit | clean-flutter, ux-patterns |
one Flutter/Dart task against the repo's analyze/format/test gate |
flutter-reviewer |
opus | clean-flutter |
read-only Flutter/Dart diff/PR review with file:line and severity |
mobile-tester |
sonnet | emulator-testing |
drives a local Android AVD or iOS simulator, reports pass/fail with screenshots |
rust-reviewer |
opus | clean-rust |
read-only Rust diff/PR review: correctness, safety, async, invariants |
shell-reviewer |
sonnet | clean-shell |
read-only shell/bash review: quoting, error handling, injection, portability |
root-cause-investigator |
opus | none | evidence-first root-cause hunt for disputed regressions, returns a hypothesis ledger |
tui-tester |
sonnet | none | drives a TUI/CLI in a real tmux pty like a user, reports pass/fail with captured screens |
ratatui-pro |
sonnet | ratatui-patterns, terminal-ux |
one Rust TUI task with current ratatui built-ins plus TestBackend tests |
ux-pro |
sonnet | ux-patterns, terminal-ux |
builds a screen's loading, success, error, empty and partial states with the repo's own components |
ux-reviewer |
opus | ux-patterns, terminal-ux |
read-only audit of what a UI does: missing states, silent failures, error copy, validation gaps |
loads if installedis soft, has the same fallback rule as the skills table and the same exception:tech-writerneedsdocs-sync.+began as wshobson/agents personas (MIT), rewritten and debloated.
Repos evaluated as skill or agent sources. Adopted ones are marked; the rest were passed over for the reason given.
show the list
wshobson/agents: adopted. source of the shell skills and the-proagents, all rewritten and debloated. a re-check folded theship-mateplugin's rule of naming each E2E check by its acceptance criterion intowebapp-tester.obra/superpowers: linked external, see External skills.anthropics/skills:webapp-testingvendored (Apache-2.0). re-checked; the other skills stay upstream.vercel-labs/skills: adopted as thenpx skillsinstaller, not a content source.VoltAgent/awesome-claude-code-subagents: 100+ subagent index, generic, nothing worth vendoring.0xfurai/claude-code-subagents:playwright-experttoo thin, wrote ownwebapp-tester.vijaythecoder/awesome-claude-agents:frontend-developercompared, no fit.yusuftayman/playwright-cli-agents: CLI-focused, replaced by ownwebapp-tester.rshah515/claude-code-subagents: browsed, generic bulk.contains-studio/agents:frontend-developercompared, redundant.lodetomasi/agents-claude-code:playwright-profetched, redundant.darcyegb/ClaudeCodeAgents: itskarencompletion checker partly overlapsproberand the testers.leonardomso/rust-skills: vendored once then delisted. one fatSKILL.md, not modular enough. its standout rules were later re-authored into the first-partyclean-rust.actionbook/rust-skills: competing rust system, passed over.vercel-labs/open-plugin-spec: spec only, nothing implements it.affaan-m/ecc: "67 agents, 271 skills" claims unverified.numman-ali/openskills: skills-only, no agent definitions.anthropics/claude-code: read for plugin-marketplace mechanics (issue #20301), not content.google-gemini/gemini-cli: read for extension and subagent mechanics.charmbracelet/crush: subagent format unshipped at research time.cursor/plugins: spec has no agent-definition concept.mohitagw15856/pm-claude-skills:--agentpicks a destination harness, not agent defs.dmicheneau/opencode-template-agent: template, not an installer.antfu/skills-npm: competing install-from-npm concept.VintLin/skill-flow: competing cross-agent skill manager.hesreallyhim/awesome-claude-code: index only, used for discovery.sickn33/antigravity-awesome-skills: bulk dump, no curation.google-labs-code/stitch-skills: unrelated, cited only for an issue.gemini-cli-extensions/security: gemini-specific.Piebald-AI/claude-code-system-prompts: system-prompt mirror, reference only.agentskills/agentskills: spec site, not content.ruvnet/ruflo: orchestration platform (swarms, memory store, background workers); its agent defs are short persona stubs, nothing adopted.davepoon/buildwithclaude: plugin marketplace, sampled; persona agents, nothing adopted.hesreallyhim/a-list-of-claude-code-agents: four persona defs, no license.stretchcloud/claude-code-unified-agents: 53 persona defs padded with invented sample code.Dicklesworthstone/claude_code_agent_farm: MIT with a rider that grants no rights to OpenAI, Anthropic, their affiliates or anyone acting for them, not adoptable.zhsama/claude-sub-agent: persona pipeline with invented 0-100 gate scores; no license file.disler/infinite-agentic-loop: generation loop that runs until context runs out, with no review step.agent-sh/agentsys: workflow plugins; its skill-invocation check already exists inbin/lint-agents.ts.Dimillian/Claude: persona defs, no license.aniketkarne/ClaudeNightsWatch: unattended daemon that skips permission prompts.NicholasSpisak/claude-code-subagents: persona defs, no license.FradSer/dotclaude: plugin collection, much of it vendored from other repos; nothing adopted.vizra-ai/claude-code-agents: 59 thin role defs with no method or gate.ayush-that/sub-agents.directory: directory of 166 persona defs.bijutharakan/multi-agent-squad: persona agents and agile-process scaffolding.gtrusler/claude-code-heavy: parallel-research shell script, machinery only.blas0/UnseveredMemory: memory hooks and templates, mostly machinery; nothing adopted.multica-ai/andrej-karpathy-skills: one guidelines skill restating model defaults.mattpocock/skills: adopted in part. its unique-tag rule for debug prints folded intoroot-cause-investigator; the rest stays upstream.7onez/cti-expert: threat-intelligence skill, outside this catalog's scope.scriptbyai.comClaude Code resource list: link index, used for discovery.
About 30 more were surveyed in a separate trending/growth analysis, not as adoption candidates, and are omitted here.
MIT.
Skills with a third-party origin keep their upstream licenses (see origin).