The most powerful plugin for Claude Code — 10 agents, 5-phase TDD workflow, self-healing memory, and multi-agent orchestration. One kernel. Zero untested code.
LLM OS architecture. Memory that heals itself. TDD that never skips.
You open Claude Code. You type a prompt. Claude writes code. You hope it works.
No structure. No tests. No quality gates. Every session starts from scratch. Every complex feature turns into prompt spaghetti.
You're the project manager, QA lead, and architect — all while trying to code.
Aura Frog treats Claude Code as an Operating System — Claude is the kernel, agents are processes, and the context window is managed RAM. You describe the feature. Aura Frog dispatches the right agent, enforces a 5-phase TDD workflow, and compresses context automatically so you never lose decisions.
You approve twice. Aura Frog handles the rest.
| ❌ Without Aura Frog | ✅ With Aura Frog |
|---|---|
|
|
Result: Production-ready code with tests, security review, and documentation — from a single prompt.
# In Claude Code (takes 30 seconds):
/plugin marketplace add nguyenthienthanh/aura-frog
/plugin install aura-frog@aurafrogStart your first workflow:
workflow:start "Your task here"Optional: Install the af CLI for health checks outside Claude Code
# In Claude Code:
setup:cli
# Or manually:
sudo ln -sf "$HOME/.claude/plugins/marketplaces/aurafrog/scripts/af" /usr/local/bin/afThen use anywhere: af doctor, af setup remote, af measure.
Works on other platforms (skills-only mode)
| Platform | Install | What Works |
|---|---|---|
| Claude Code | /plugin marketplace add nguyenthienthanh/aura-frog |
Everything |
| OpenAI Codex | cp -r aura-frog/skills/* ~/.codex/skills/ |
Skills + commands |
| Gemini CLI | cp -r aura-frog/skills/* ~/.gemini/skills/ |
Skills + commands |
| OpenCode | cp -r aura-frog/skills/* .opencode/skills/ |
Skills + commands |
Hooks, agent detection, and MCP servers are Claude Code exclusive.
No more "write code and hope." Every feature follows RED → GREEN → REFACTOR automatically:
✋ Phase 1: Understand + Design → You approve the plan
⚡ Phase 2: Test RED → Failing tests written
✋ Phase 3: Build GREEN → You approve the implementation
⚡ Phase 4: Refactor + Review → Auto quality + security check
⚡ Phase 5: Finalize → Docs + notifications
Two approvals. Five phases. Zero untested code.
10 specialized agents activate automatically — no configuration:
"Build a React dashboard" → frontend
"Optimize the SQL queries" → architect
"Set up CI/CD pipeline" → devops
"Fix the login screen crash" → mobile
"Run a security audit" → security
All 10 agents
| Agent | When it activates |
|---|---|
lead |
Coordinates workflows, team orchestration |
architect |
System design, databases, backend (Node, Python, Laravel, Go) |
frontend |
React, Vue, Angular, Next.js + design systems |
mobile |
React Native, Flutter, Expo, NativeWind |
strategist |
ROI evaluation, MVP scoping, scope creep detection |
security |
OWASP audits, vulnerability scanning, SAST |
tester |
Jest, Cypress, Playwright, Detox, coverage |
devops |
Docker, K8s, CI/CD, monitoring |
scanner |
Project detection, config, context |
router |
Agent + model selection |
For deep tasks, 4 agents independently analyze your plan — then challenge each other:
📐 Architect → "How to build it"
🔍 Tester → "How it can fail"
👤 Frontend → "How users experience it"
💼 Strategist → "Should we even build this?"
Plans survive 4 rounds of scrutiny before a single line of code. Catches scope creep and wasted effort before it happens.
Run project:init once. Every future session instantly understands your codebase — conventions, architecture, patterns, file relationships. 12 pattern detections. 7 context files generated.
No more re-explaining your project every session.
For complex work, Aura Frog spins up a real team working in parallel:
lead
├── architect → Designs the system
├── frontend → Builds the UI
├── tester → Writes tests
└── security → Reviews for vulnerabilities
All cross-reviewing each other's work.
Only activates when needed. Simple tasks stay single-agent (saves ~3x tokens).
6 bundled servers auto-invoke when Claude needs them:
"Build with MUI" → context7 fetches current MUI docs
"Test the login page" → playwright launches a browser
"Check test coverage" → vitest runs your suite
"Deploy to Firebase" → firebase manages the project
Plus Figma design fetching and Slack notifications.
More features
Detects your patterns, remembers corrections, creates rules that persist across sessions. Optional Supabase sync for teams.
Automatically matches effort to task size — typos get direct edits, features get full workflows, architecture gets collaborative planning. No configuration.
Workflow crashed? workflow:resume. Context full? Decisions preserved across /compact. Need to pause? workflow:handoff saves everything.
All cached context is treated as a hint — agents verify against actual files before acting. State only updates after confirmed success (Strict Write Discipline). No stale assumptions propagate.
MicroCompact (free, every 10 turns) → AutoCompact (one /compact call at 80%) → ManualCompact (full session snapshot). Context stays lean. Decisions survive.
3-tier rule loading (~75% less context), conditional hooks (~40% fewer executions), agent detection caching, session start caching (<1s repeat sessions).
| Component | Count | Why it matters |
|---|---|---|
| Agents | 10 | Right expert auto-selected per task |
| Skills | 44 | 8 auto-invoke on context, 36 on-demand |
| Commands | 90 | 5 bundled menus — discoverability built in |
| Rules | 45 | 3-tier loading — only what's needed per phase |
| Hooks | 35 | Conditional — skip processing for non-code files |
| MCP Servers | 6 | Zero-config, auto-invoked |
Full workflow target: ≤30K tokens across all 5 phases.
| Mode | Command | Best for |
|---|---|---|
| Full Workflow | workflow:start "task" |
New features, production code |
| Fast-Track | fasttrack: <specs> |
Pre-approved specs |
| Quick Fix | bugfix:quick "fix" |
Small bugs with TDD |
| Refactor | refactor "file" |
Code cleanup |
| All Documentation | docs/README.md |
| Getting Started | GET_STARTED.md |
| First Workflow Tutorial | FIRST_WORKFLOW_TUTORIAL.md |
| All Commands (90) | commands/README.md |
| All Skills (44) | skills/README.md |
| Agent Teams Guide | AGENT_TEAMS_GUIDE.md |
| MCP Setup | MCP_GUIDE.md |
| Hooks & Lifecycle | hooks/README.md |
| Troubleshooting | TROUBLESHOOTING.md |
| Changelog | CHANGELOG.md |
Claude = Kernel Context Window = RAM Project Files = Disk
Agents = Processes 5-Phase TDD = Scheduler MCP = Device Drivers
TOON = Compression Approval Gates = Interrupts Handoffs = IPC
aura-frog/
├── agents/ 10 processes (auto-dispatched per task)
├── skills/ 44 skills (8 auto-invoke + 36 on-demand)
├── commands/ 90 commands (5 bundled menus)
├── rules/ 45 rules (13 core + 15 agent + 17 workflow)
├── hooks/ 35 lifecycle hooks (conditional execution)
├── scripts/ 43 utility scripts
├── docs/ AI reference docs (phases, TOON refs)
└── .mcp.json 6 device drivers (MCP servers)
We welcome contributions — especially new MCP integrations, agents, skills, and bug fixes. See CONTRIBUTING.md or submit an issue.
Godot and SEO/GEO modules available as separate addons.
MIT — See LICENSE