🐦 AgentCrow

Claude Code Agent Teams is powerful. But it doesn't know which agents to use.
AgentCrow does. 9 builtin agents + external agents. Auto-dispatch. npx agentcrow init

Problem • Install • How It Works • Agents • Commands • 한국어 • 日本語 • 中文 • Español • Português • Deutsch • Français • Русский • हिन्दी • Türkçe • Tiếng Việt

---

🆚 Agent Teams alone vs. Agent Teams + AgentCrow

	Agent Teams alone	+ AgentCrow
Spawn subagents	✅	✅
Know which agents to use	❌ you decide	✅ auto-matched
9 builtin + external agent roles	❌ you write prompts	✅ ready to go
Auto-decompose prompts	❌ you split manually	✅ one prompt in
Agent identity & rules	❌ blank subagents	✅ personality, MUST/MUST NOT
Works without config	❌ needs --agents JSON	✅ npx agentcrow init
Multiple divisions (eng, game, design...)	❌	✅

Agent Teams gives you the engine. AgentCrow gives it a brain.

---

  You:    "Build a SaaS dashboard with Stripe billing, auth, and docs"

  🐦 AgentCrow auto-dispatches 5 agents:

    🖥️  frontend_developer  → React dashboard UI, charts, responsive layout
    🏗️  backend_architect   → Auth system, REST API, database schema
    💳  backend_architect   → Stripe integration, webhook handlers
    🧪  qa_engineer         → E2E billing flow tests, auth edge cases
    📝  technical_writer    → API reference, onboarding guide

  You didn't pick them. AgentCrow did.

⬇️ One line. That's it.

npx agentcrow init

Then just run claude as usual. AgentCrow handles the rest.
macOS · Linux · Windows

---

👀 Before / After

❌ Without AgentCrow You: Build a dashboard with API, tests, and docs Claude: (one agent does everything) - reads all files - writes all code - runs all tests - writes all docs = one context window = forgets early work = 10+ minutes

✅ With AgentCrow You: same prompt AgentCrow auto-dispatches: @ui_designer → layout @frontend_dev → React code @backend_arch → API @qa_engineer → tests @tech_writer → docs = parallel agents = each focused = better results

---

⚡ Install

npx agentcrow init

That's it. This does two things:

First run only — downloads agents to ~/.agentcrow/ (global, shared across all projects)

Every run — merges AgentCrow section into .claude/CLAUDE.md (your existing rules are preserved)

Note

Agents are stored globally at ~/.agentcrow/. Second project onward = instant, no download.

Tip

Already have a CLAUDE.md? AgentCrow appends its section — your existing rules stay untouched.

⚙️ How It Works

  ┌─────────────────────────────────────┐
  │  Your prompt                        │
  │           ↓                         │
  │  ┌────────────────────────────┐     │
  │  │ CLAUDE.md reads agent list │     │
  │  │ Claude decomposes prompt   │     │
  │  │ Dispatches Agent tool      │     │
  │  │ Each agent works in scope  │     │
  │  └────────────────────────────┘     │
  │           ↓                         │
  │  Files created, tests written,      │
  │  docs generated — by specialists    │
  └─────────────────────────────────────┘

1. You run claude in a project with AgentCrow initialized
2. You type a prompt — anything complex
3. Claude reads CLAUDE.md — sees the agent roster and dispatch rules
4. Claude decomposes — splits your prompt into focused tasks
5. Claude dispatches — uses the Agent tool to spawn subagents
6. Each agent works — with its own expertise

No API key. No server. Just Claude Code + CLAUDE.md.

🤖 9 Builtin Agents + External Agents

Division	Examples
Engineering	frontend_developer, backend_architect, ai_engineer, sre
Game Dev	game_designer, level_designer, unreal, unity, godot
Marketing	content_strategist, seo_specialist, social_media
Testing	test_automation, performance_tester
Design	ui_designer, ux_researcher, brand_guardian
Builtin	qa_engineer, korean_tech_writer, security_auditor
+ more	sales, support, product, strategy, spatial-computing...

🔧 Commands

npx agentcrow init              # Set up agents + CLAUDE.md (English default)
npx agentcrow init --lang ko    # Korean template
npx agentcrow status            # Check if active
npx agentcrow off               # Disable temporarily
npx agentcrow on                # Re-enable
npx agentcrow agents            # List all agents
npx agentcrow agents search ai  # Search by keyword
npx agentcrow compose "prompt"  # Preview decomposition (dry run)

💡 Example Prompts

Build an AI-powered news aggregator with crawling and email alerts
→ ai_engineer + data_pipeline_engineer + frontend_developer + devops_automator

E-commerce site with Stripe, inventory management, and admin panel
→ frontend_developer + backend_architect + ui_designer + qa_engineer

Open source CLI tool with tests, docs, and CI/CD pipeline
→ frontend_developer + qa_engineer + technical_writer + devops_automator

Simple prompts run normally. AgentCrow only activates for multi-task requests.

🛡️ Zero Overhead

🟢 Complex prompts	Auto-decomposed into agents
🔵 Simple prompts	Runs normally, no agents
🔴 agentcrow off	Completely disabled

Important

AgentCrow only touches .claude/CLAUDE.md and .claude/agents/. No project dependencies, no background processes. agentcrow off backs up and removes both cleanly.

🤝 Contributing

git clone --recursive https://github.com/jee599/agentcrow.git
cd agentcrow && npm install && npm test  # 60 tests

📜 License

MIT — External agents from agency-agents.

---

🐦 One prompt. Many agents. Zero config.