BMAD-MCP

BMAD-MCP

Business-Minded Agile Development workflow orchestrator as an MCP (Model Context Protocol) server.

Complete agile development workflow: PO → Architect → SM → Dev → Review → QA

Interactive Requirements Gathering - Asks clarifying questions to ensure complete requirementsDynamic Engine Selection - Uses Claude by default, dual-engine when neededContent Reference System - Efficient token usage via file referencesHuman-Readable Task Names - Organizes by task name, not UUID

🎯 What is BMAD-MCP?

BMAD-MCP is a lightweight workflow orchestrator that manages the complete agile development process. It:

• Manages workflow state (which stage you're in, what's needed next)
• Dispatches role prompts (provides detailed prompts for each role)
• Saves artifacts (PRD, architecture, code, reports)
• Does NOT call LLMs (that's Claude Code's job)

🏗️ Architecture

User → Claude Code → bmad-mcp tool
                       ↓
            Returns: {
              stage: "po",
              role_prompt: "<complete PO prompt>",
              engines: ["claude", "codex"],
              context: {...}
            }
                       ↓
Claude Code executes:
  - Calls Claude (with role_prompt)
  - Calls Codex MCP (with role_prompt)
                       ↓
Claude Code submits results → bmad-mcp
                       ↓
bmad-mcp: merges, scores, saves, advances to next stage

📋 Workflow Stages

🚀 Quick Start

Installation (3 Steps)

# Step 1: Install globally from npm
npm install -g bmad-mcp

# Step 2: Add to Claude Code
claude mcp add-json --scope user bmad '{"type":"stdio","command":"bmad-mcp"}'

# Step 3: Verify installation
bmad-mcp
# Expected output: "BMAD MCP Server running on stdio"

That's it! Restart Claude Code and you're ready to use BMAD workflow.

Usage Example

Simply tell Claude Code to use BMAD:

User: Use bmad-task to create a user authentication system

Claude Code will:
1. Start BMAD workflow (PO stage)
2. Generate Product Requirements Document (with interactive Q&A)
3. Generate System Architecture (with interactive Q&A)
4. Create Sprint Plan
5. Implement code (using Codex)
6. Perform code review
7. Run quality assurance tests

All artifacts saved to: .claude/specs/user-authentication-system/

Configuration Location

MCP configuration is automatically added to ~/.claude/config.json:

{
  "mcpServers": {
    "bmad": {
      "type": "stdio",
      "command": "bmad-mcp"
    }
  }
}

🔧 Advanced Installation

Option 1: NPM Install (Recommended)

npm install -g bmad-mcp
claude mcp add-json --scope user bmad '{"type":"stdio","command":"bmad-mcp"}'

Option 2: Build from Source

git clone https://github.com/cexll/bmad-mcp-server
cd bmad-mcp-server
npm install
npm run build
npm link  # Makes bmad-mcp globally available

# Add to Claude Code
claude mcp add-json --scope user bmad '{"type":"stdio","command":"bmad-mcp"}'

Verify Installation

# Check if binary is available
which bmad-mcp
# Output: /usr/local/bin/bmad-mcp (or similar)

# Test the server directly
bmad-mcp
# Expected output: "BMAD MCP Server running on stdio"
# Press Ctrl+C to exit

# Restart Claude Code to load the configuration

Uninstall

# Remove from Claude Code
claude mcp remove bmad

# Uninstall npm package
npm uninstall -g bmad-mcp

📖 Usage

Basic Workflow

// 1. Start workflow
const startResult = await callTool("bmad-task", {
  action: "start",
  cwd: "/path/to/your/project",
  objective: "Implement user login system"
});

const { session_id, task_name, role_prompt, engines } = JSON.parse(startResult.content[0].text);

// 2. Execute with engines
if (engines.includes("claude")) {
  const claudeResult = await callClaude(role_prompt);
}
if (engines.includes("codex")) {
  const codexResult = await callCodexMCP(role_prompt);
}

// 3. Submit results
await callTool("bmad-task", {
  action: "submit",
  session_id: session_id,
  stage: "po",
  claude_result: claudeResult,
  codex_result: codexResult
});

// 4. Confirm and proceed (unified: saves + advances to next stage)
await callTool("bmad-task", {
  action: "confirm",
  session_id: session_id,
  confirmed: true
});

Actions

start - Start a new workflow

{
  "action": "start",
  "cwd": "/path/to/project",
  "objective": "Project description"
}

Returns:

{
  "session_id": "uuid",
  "task_name": "project-description",
  "stage": "po",
  "state": "generating",
  "stage_description": "Product Owner - Requirements Analysis",
  "requires_user_confirmation": true,
  "interaction_type": "awaiting_generation",
  "user_message": "📋 **BMAD 工作流已启动**...",
  "role_prompt": "<complete prompt>",
  "engines": ["claude"],
  "context": {...},
  "pending_user_actions": ["review_and_confirm_generation"]
}

submit - Submit stage results

{
  "action": "submit",
  "session_id": "uuid",
  "stage": "po",
  "claude_result": "...",
  "codex_result": "..."
}

Returns (if score >= 90):

{
  "session_id": "uuid",
  "stage": "po",
  "state": "awaiting_confirmation",
  "score": 92,
  "requires_user_confirmation": true,
  "interaction_type": "user_decision",
  "user_message": "✅ **PRD生成完成**\n质量评分：92/100...",
  "final_draft_summary": "...",
  "final_draft_file": ".bmad-task/temp/uuid/po_final_result_xxx.md",
  "pending_user_actions": ["confirm", "reject_and_refine"]
}

Returns (if score < 90 with clarification questions):

{
  "session_id": "uuid",
  "stage": "po",
  "state": "clarifying",
  "current_score": 75,
  "requires_user_confirmation": true,
  "interaction_type": "user_decision",
  "user_message": "⚠️ **需求澄清...**",
  "gaps": ["Target user group unclear", "..."],
  "questions": [{"id": "q1", "question": "...", "context": "..."}],
  "pending_user_actions": ["answer_questions"]
}

confirm - Confirm and save (unified action)

{
  "action": "confirm",
  "session_id": "uuid",
  "confirmed": true
}

Returns (saves artifact + advances to next stage):

{
  "session_id": "uuid",
  "stage": "architect",
  "state": "generating",
  "requires_user_confirmation": true,
  "interaction_type": "awaiting_generation",
  "user_message": "💾 **文档已保存，并已进入下一阶段**...",
  "role_prompt": "<architect prompt>",
  "engines": ["claude"],
  "previous_artifact": ".claude/specs/task-name/01-product-requirements.md",
  "pending_user_actions": ["review_and_confirm_generation"]
}

answer - Answer clarification questions

{
  "action": "answer",
  "session_id": "uuid",
  "answers": {
    "q1": "Target users are enterprise B2B customers",
    "q2": "Expected 10k concurrent users with <200ms response time"
  }
}

Returns:

{
  "session_id": "uuid",
  "stage": "po",
  "state": "refining",
  "requires_user_confirmation": true,
  "interaction_type": "awaiting_regeneration",
  "user_message": "📝 **已收到你的回答**...",
  "role_prompt": "<updated prompt with user answers>",
  "engines": ["claude"],
  "pending_user_actions": ["regenerate_with_answers"]
}

approve - Approve current stage (SM stage only)

{
  "action": "approve",
  "session_id": "uuid",
  "approved": true
}

Returns (when entering Dev stage):

{
  "session_id": "uuid",
  "stage": "dev",
  "state": "generating",
  "requires_user_confirmation": true,
  "interaction_type": "awaiting_generation",
  "user_message": "✅ **Sprint Plan 已批准**\n\n正在进入下一阶段：Developer - Implementation\n\nSprint Plan 包含 3 个 Sprint：\n1. Sprint 1: 基础架构\n2. Sprint 2: 核心功能\n3. Sprint 3: 优化和完善\n\n⚠️ 重要：请明确指示开发范围...",
  "role_prompt": "<dev prompt>",
  "engines": ["codex"],
  "pending_user_actions": ["specify_sprint_scope_then_generate"]
}

Important - Dev Stage Behavior:

• After approving Sprint Plan, workflow enters Dev stage but does NOT auto-start development
• User must explicitly specify development scope:
  • "开始开发" / "start development" → Implements ALL sprints (default)
  • "开发 Sprint 1" / "implement sprint 1" → Implements only Sprint 1
• This ensures users have full control over what gets implemented and when

status - Query workflow status

{
  "action": "status",
  "session_id": "uuid"
}

Returns:

{
  "session_id": "uuid",
  "current_stage": "dev",
  "current_state": "generating",
  "stages": {...},
  "artifacts": [...]
}

📁 File Structure

Your Project

your-project/
├── .bmad-task/
│   ├── session-abc-123.json          # Workflow state (with content references)
│   ├── task-mapping.json             # Maps session_id → task_name
│   └── temp/
│       └── abc-123/                  # Temporary content files
│           ├── po_claude_result_xxx.md
│           ├── po_codex_result_xxx.md
│           └── po_final_result_xxx.md
├── .claude/
│   └── specs/
│       └── implement-user-login/     # Task name (human-readable slug)
│           ├── 01-product-requirements.md
│           ├── 02-system-architecture.md
│           ├── 03-sprint-plan.md
│           ├── 04-dev-reviewed.md
│           └── 05-qa-report.md
└── src/

Session State File

{
  "session_id": "abc-123",
  "task_name": "implement-user-login",
  "cwd": "/path/to/project",
  "objective": "Implement user login",
  "current_stage": "dev",
  "current_state": "generating",
  "stages": {
    "po": {
      "status": "completed",
      "claude_result_ref": {
        "summary": "First 300 chars...",
        "file_path": ".bmad-task/temp/abc-123/po_claude_result_xxx.md",
        "size": 12450,
        "last_updated": "2025-01-15T10:30:00Z"
      },
      "final_result_ref": {...},
      "score": 92,
      "approved": true
    },
    ...
  },
  "artifacts": [".claude/specs/implement-user-login/01-product-requirements.md", ...]
}

🎨 Engine Configuration

PO & Architect Stages (Dynamic Engine Selection)

• Default: Only Claude (single engine)
• Dual Engine: Enabled when objective contains "codex" or "使用 codex"
• If dual engine enabled:
  • Calls both Claude and Codex
  • Each generates independent solution
  • BMAD-MCP merges results:
    • If both ≥ 90: choose higher score
    • If one ≥ 90: choose that one
    • If both < 90: choose higher score, refine
• Interactive Clarification:
  • First iteration: Identify gaps, generate 3-5 clarification questions
  • User answers questions
  • Regenerate based on answers
  • Iterate until score ≥ 90

SM Stage (Claude Only)

• Only calls Claude
• Scrum planning doesn't need Codex

Dev/Review/QA Stages (Codex Only)

• Only calls Codex MCP
• Uses GPT-5 for code tasks
• Important: Use model: "gpt-5" (NOT "gpt-5-codex")
• Parameters:
  • model: "gpt-5"
  • sandbox: "danger-full-access"
  • approval-policy: "on-failure"

🔄 Workflow Flow

graph TD
    A[Start] --> B[PO Stage: Generate]
    B --> C{Has Questions?}
    C -->|Yes| D[Clarifying: User Answers]
    D --> E[Refining: Regenerate]
    E --> F{Score >= 90?}
    C -->|No| F
    F -->|No| C
    F -->|Yes| G[Awaiting Confirmation]
    G -->|confirm| H[Saved + Architect Stage]
    H --> I{Has Questions?}
    I -->|Yes| J[Clarifying: User Answers]
    J --> K[Refining: Regenerate]
    K --> L{Score >= 90?}
    I -->|No| L
    L -->|No| I
    L -->|Yes| M[Awaiting Confirmation]
    M -->|confirm| N[Saved + SM Stage]
    N -->|approve| O[Dev Stage]
    O --> P[Review Stage]
    P --> Q[QA Stage]
    Q --> R[Complete]

🛠️ Development

Project Structure

bmad-mcp/
├── src/
│   ├── index.ts              # Main MCP server
│   └── master-prompt.ts      # All role prompts
├── dist/                     # Compiled output
├── package.json
├── tsconfig.json
└── README.md

Build

npm run build

Development Mode

npm run dev  # Watch mode

Test Locally

npm run build
node dist/index.js

📚 Master Orchestrator Design

All role prompts are embedded in a single master-prompt.ts file:

• Centralized management: All roles in one place
• Workflow definition: Clear stage sequence
• Engine configuration: Which engines for each stage
• Quality gates: Score thresholds and approval points

🤝 Integration with Codex MCP

When calling Codex for Dev/Review/QA stages:

// Claude Code calls Codex MCP
await callTool("codex", {
  prompt: role_prompt,  // From bmad-task
  model: "gpt-5",  // IMPORTANT: Use "gpt-5", NOT "gpt-5-codex"
  sandbox: "danger-full-access",
  "approval-policy": "on-failure"
});

⚙️ Configuration

Quality Thresholds

Defined in master-prompt.ts:

quality_gates: {
  po: { min_score: 90, approval_required: true },
  architect: { min_score: 90, approval_required: true },
  sm: { approval_required: true },
  dev: {},
  review: {},
  qa: {}
}

Artifact Filenames

artifacts: {
  po: "01-product-requirements.md",
  architect: "02-system-architecture.md",
  sm: "03-sprint-plan.md",
  dev: "code-implementation",
  review: "04-dev-reviewed.md",
  qa: "05-qa-report.md"
}

🔍 Troubleshooting

Server not starting

# Check installation
which bmad-mcp

# Test directly
bmad-mcp

Tool name error

• Important: The tool name is bmad-task, not bmad
• Use callTool("bmad-task", {...}) in your code
• Claude Code configuration should use bmad-task as the tool name

Session not found

• Ensure .bmad-task/ directory has write permissions
• Check session_id is correct
• Verify cwd path is absolute

Scores not detected

• Ensure generated content includes: Quality Score: X/100 or "quality_score": 92 in JSON
• Check score format matches pattern (0-100)
• Score ≥ 90 required for PO/Architect stages to advance

Clarification workflow issues

• If you see state: "clarifying", user must answer questions via answer action
• Do NOT auto-generate answers - wait for real user input
• Check requires_user_confirmation: true before proceeding

📝 License

MIT

🙋 Support

• Documentation: This README
• Issues: GitHub issues
• Reference: https://github.com/cexll/myclaude

Transform your development with BMAD - One workflow, complete agile process, quality assured.