语言 / Language: 简体中文 · English

DevCrew

Give AI a collaboration protocol. Let it work like a real team.

帮你做好 harness！

The Problem

When using AI (Copilot, Claude, Cursor…) for development:

| Problem | What happens | |---------|-------------| | No memory | Switch conversations, AI forgets everything | | No division of labor | AI plays PM + architect + dev + tester simultaneously | | Goes off track | Drifts from goals with no checkpoints to correct course | | Quality blind spots | No review process — bugs and tech debt accumulate silently | | No starting point | Don't know how to orchestrate AI collaboration |

Root cause: AI lacks a persistent collaboration protocol. DevCrew is that protocol.

Get Started in 30 Seconds

npm install -g @lordmos/dev-crew
cd your-project
crew init --platform copilot   # or cursor / claude

Three steps — INSTRUCTIONS.md is auto-placed where your AI platform reads it. The AI instantly gains multi-agent team orchestration capabilities.

| Platform | Command | Instruction File | |----------|---------|-----------------| | GitHub Copilot | crew init -p copilot | .github/copilot-instructions.md | | Cursor | crew init -p cursor | .cursorrules | | Claude Code | crew init -p claude | CLAUDE.md | | Generic | crew init | INSTRUCTIONS.md (reference manually) |

Multiple platforms at once: crew init -p copilot cursor

How It Works

You: I need to add auth middleware to the API

AI:  [PdM] Creating change add-api-auth, mode: Standard
     Plan — Requirements:
     - Goal: Add JWT auth to all /api/ routes
     - Acceptance: [ ] No token → 401  [ ] Expired token → 401
     Please confirm.

You: Confirmed

AI:  Design → Execute → Verify — All passed. Please confirm acceptance.

You: Confirmed

AI:  [OK] Change add-api-auth complete.

You only confirmed twice (requirements + results). Everything else was automatic.

What crew init Creates

your-project/
├── INSTRUCTIONS.md    ← AI behavior instructions (core file)
├── dev-crew.yaml       ← Project config (modes, specialists)
└── dev-crew/
    ├── specs/         ← Shared specifications
    └── memory/        ← Agent long-term memory (auto-accumulated)

AI reads INSTRUCTIONS.md and PjM orchestrates the team — agents collaborate in parallel following the PDEVI workflow.

Core Concepts

PDEVI Workflow

flowchart LR
    P["Plan"] --> D["Design"]
    D --> E["Execute"]
    E --> V["Verify"]
    V -->|Fail| I["Iterate"]
    I -->|Design issue| D
    I -->|Implementation issue| E
    V -->|Pass| Done["[OK] Done"]

Three modes for every scenario:

| Mode | Flow | Best for | |------|------|----------| | Standard | P → D → E → V → I | New features, refactoring | | Express | P → E → V | Bug fixes | | Prototype | P → D → E | Quick prototyping |

On-Demand Team Assembly

PjM creates agents on demand based on user needs. Common roles:

| Agent | Responsibility | |-------|---------------| | PjM Project Manager | Task decomposition, agent coordination, progress tracking | | PdM Product Manager | Requirements analysis, PRD import, acceptance criteria | | Architect | Tech decisions, task decomposition, dependency analysis | | Implementer | Code generation, refactoring, dependency management | | Tester | Test execution, acceptance checks, coverage | | Reviewer | Code review, security scanning, best practices |

Team size is flexible — PjM creates additional agents (e.g., DBA, Tech Writer, SRE) as needed, no manual assignment required.

Domain Specialists (29)

Beyond the core team, 29 domain specialists across 10 fields, activated on demand:

Game Dev (8) · UI/UX (3) · Security (1) · DevOps (3) · Testing (3) · Engineering (5) · Data (2) · AI/ML (1) · Web3 (1) · Spatial Computing (2)

# dev-crew.yaml
specialists:
  - game-designer
  - security-engineer

crew agents  # List all available specialists

See the full Specialist Directory

Skills

| Skill | CLI | MCP Tool | Purpose | |-------|-----|----------|---------| | init | crew init | crew_init | Initialize workspace + agent memory files | | plan | crew plan <name> | crew_plan | Create a change and start working | | status | crew status | crew_status | Check current progress | | checkpoint | crew checkpoint | crew_checkpoint | Phase audit + consistency check + memory sync | | release | crew release | crew_release | Archive changes + consolidate memory | | agents | crew agents | crew_agents | List available specialists |

Natural language works too — "run a checkpoint" triggers the checkpoint skill automatically

Use Cases

| Scenario | You say | DevCrew does | |----------|---------|-------------| | Greenfield | "I have an idea, build from scratch" | Init → guide requirements → Standard | | Existing PRD | "Here's the PRD, execute it" | Import PRD → refine → Standard | | Mid-project | "Code exists, help me continue" | Scan code → establish baseline → Standard | | Brainstorm | "Let's discuss the approach" | Explore mode (no code changes) | | Bug fix | "There's a bug, fix it fast" | Express mode | | Refactor | "This code needs refactoring" | Standard (full workflow) | | Prototype | "Build a quick prototype first" | Prototype mode | | Learn codebase | "Help me understand this code" | Explore mode (code analysis) |

Architecture

┌─────────────────────────────────────────────┐
│  Skills Layer                                │
│  ┌──────────────┐  ┌──────────────────────┐ │
│  │ CLI          │  │ MCP Server           │ │
│  │ crew init    │  │ crew_init            │ │
│  │ crew plan    │  │ crew_plan            │ │
│  │ crew status  │  │ crew_status          │ │
│  │ crew release │  │ crew_release         │ │
│  └──────────────┘  └──────────────────────┘ │
├─────────────────────────────────────────────┤
│  Protocol Layer (core, zero tool dependency) │
│  INSTRUCTIONS.md · PDEVI workflow ·          │
│  File conventions · Communication rules      │
└─────────────────────────────────────────────┘

Even without CLI/MCP, manually placing INSTRUCTIONS.md works. Skills make the workflow more efficient.

Documentation

| Doc | Description | |-----|-------------| | User Manual | Detailed guide for 8 scenarios | | Best Practices | Scenario walkthrough examples | | Specialists | 29 specialists · 10 domains |

Contributing

Contributions welcome! See CONTRIBUTING.md.

License

MIT

Credits

Domain specialists adapted from the agency-agents-zh project.