mcp-to-skill

Convert MCP servers to Claude Skills. Save ~90% context tokens.

Requirements

• Node.js 18+
• npm/npx

Usage

1. Create a MCP config to consume

Stdio MCP (npx-based):

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {"GITHUB_TOKEN": "your-token"}
    }
  }
}

HTTP MCP

{
  "mcpServers": {
    "my-server": {
      "type": "http",
      "url": "http://localhost:8080/mcp"
    }
  }
}

Direct format (without mcpServers wrapper)

{
  "name": "github",
  "command": "npx",
  "args": ["-y", "@modelcontextprotocol/server-github"],
  "env": {"GITHUB_TOKEN": "your-token"}
}

2. Generate the skill

npx -y mcp-to-skill generate \
  --mcp-config my-mcp.json \
  --output-dir ./skills/my-skill

3. Copy to Claude skills directory

cp -r ./skills/my-skill ~/.claude/skills/

Claude discovers it automatically. The generated SKILL.md instructs Claude to call tools via:

npx -y mcp-to-skill exec \
  --config /path/to/mcp-config.json \
  --call '{"tool": "tool_name", "arguments": {...}}'

Output

skills/my-skill/
├── SKILL.md         # Instructions for Claude (~100 tokens)
└── mcp-config.json  # MCP server config

Fine-tuning

In my experience it is best to fine tune the skill.md a little by enhancing the description to be more concise about the tool (e.g. give the agent context on the purpose of the mcp/tool). Additionally some MCP's tool usage descriptions are too verbose and can be edited down for further savings.

Development

git clone https://github.com/larkinwc/ts-mcp-to-skill.git
cd ts-mcp-to-skill
npm install
npm run build

# Run locally
npx tsx src/index.ts generate --mcp-config config.json --output-dir ./skills/name

# Test executor
npx tsx src/index.ts exec --config mcp-config.json --list
npx tsx src/index.ts exec --config mcp-config.json --describe tool_name
npx tsx src/index.ts exec --config mcp-config.json --call '{"tool": "...", "arguments": {}}'

License

MIT