日本語

An MCP server that exposes the Agent Skill creation guide (9 phases) as MCP resources. AI agents retrieve only the phases they need on demand and follow the process to build SKILL.md files.

Quick Start

Claude Code:

claude mcp add skill-forge-mcp -- npx skill-forge-mcp

Gemini CLI:

gemini mcp add skill-forge-mcp -- npx skill-forge-mcp

VS Code (GitHub Copilot) — .vscode/mcp.json:

{
  "servers": {
    "skill-forge-mcp": {
      "command": "npx",
      "args": ["skill-forge-mcp"]
    }
  }
}

Cursor:

{
  "skill-forge-mcp": {
    "command": "npx",
    "args": ["skill-forge-mcp"]
  }
}

Claude Desktop

{
  "mcpServers": {
    "skill-forge-mcp": {
      "command": "npx",
      "args": ["skill-forge-mcp"]
    }
  }
}

Usage

Ask your agent:

"I want to create a skill for React component design. Follow the SkillForge MCP process."

The agent will automatically:

1. Fetch the process structure from process://manifest
2. Read Phase 1 (process://phase/1) for scoping, run baseline measurements
3. Record progress with mark_progress as it advances through each phase
4. Generate the final SKILL.md following Phase 6 guidelines

Use search_process for keyword lookups across phases.

The 9 Phases

Phase	Name	Purpose
0	Skill Specification	SKILL.md structure and frontmatter
1	Scoping & Baseline	Measure failure patterns; define research scope
2	Domain Research	Establish quality criteria and theoretical foundations
3	Gap Analysis	Verify whether research alone enables the agent to act
4	Deep Implementation Research	Fill gaps with code examples, anti-patterns, validation
5	Structuring & Completeness	Confirm coverage across all categories
6	Distillation into SKILL.md	Condense into ≤500 lines; maximize token efficiency
7	Deploy & Validate	Place, verify spec compliance, security review
8	Evaluate & Iterate	Compare against baseline, improve iteratively

Features

• Staged access — retrieve content at phase or section granularity
• Cross-phase search — keyword search across all 9 phases
• Progress tracking — record and query per-phase completion status
• Prompt templates — create_skill and resume_skill prompts for guided workflows
• Structured output — outputSchema + structuredContent on all tools for programmatic consumption
• State persistence — optionally retain progress across sessions
• Low overhead — ~1,500 token fixed cost to the context window

API

Resources

URI	Description
process://manifest	Full index (JSON)
process://phase/0 – process://phase/8	Phase 0–8 content

Resource Templates

Template	Description
process://phase/{phaseId}/section/{sectionName}	Retrieve a single section
process://phases/{+phaseIds}	Batch retrieval (e.g. 1,2,3)

Tools

Tool	Description	Input
search_process	Keyword search across all phases	{ "query": "frontmatter", "maxResults": 5 }
mark_progress	Record phase progress	{ "phaseId": 1, "status": "in-progress" }
get_status	Progress summary for all phases	{}

status: "not-started" · "in-progress" · "completed"

Prompts

Prompt	Description
create_skill	Full guided workflow (Phase 0→8). Accepts a topic argument.
resume_skill	Resume from current progress. Checks get_status and continues.

Configuration

Set SKILL_FORGE_PERSIST=true to persist progress to ~/.skill-forge-mcp/state.json:

{
  "mcpServers": {
    "skill-forge-mcp": {
      "command": "npx",
      "args": ["skill-forge-mcp"],
      "env": { "SKILL_FORGE_PERSIST": "true" }
    }
  }
}

Development

git clone https://github.com/popyson1648/skill-forge-mcp.git
cd skill-forge-mcp
npm install
npm run build
npm test          # 52 tests

Project structure

src/
├── index.ts        # Entry point
├── content.ts      # Content loading & section extraction
├── search.ts       # Cross-phase search
├── state.ts        # State management & persistence
├── status.ts       # Status table formatter
├── content/        # English content (served)
└── content-ja/     # Japanese translations (developer reference only)
tests/
├── content.test.ts
├── search.test.ts
├── state.test.ts
├── resources.test.ts
└── tools.test.ts

Requirements: Node.js >= 18

Contributing

Contributions are welcome! Feel free to open an Issue or submit a Pull Request.

License

MIT