MCP Tool Extractor

Protocol-level tool extraction from any MCP server config.
Operates at tools/list — extracts individual tools, not servers.

Setup

npm install

Usage

# Extract all servers from a config
node extract.mjs --config /path/to/all.json

# Extract specific servers only
node extract.mjs --config ./all.json --servers agentic-framework slack eme

# Custom output directory
node extract.mjs --config ./all.json --out ./output

# Folder of multiple configs
node extract.mjs --folder ./configs/

Output Structure

mcp-tools/
├── registry/
│   └── registry.json          # sha256 hash index — deduplication across runs
├── raw/
│   └── {server}/
│       └── {tool}.json        # Source of truth — all generation derives from here
├── packages/
│   └── {server}/
│       └── {tool}/
│           ├── index.ts       # Typed async function + interfaces
│           ├── schema.json    # Input schema
│           └── README.md      # Auto-generated docs with input table
├── tools/
│   └── {server}/
│       ├── standalone/
│       │   └── {tool}.js      # Zero-dep Node.js module
│       └── shared/
│           └── transport.js   # MCP call handler
└── extract-summary.json       # Run summary with per-server stats

Notes

• Registry dedup: On subsequent runs, tools with unchanged schemas are skipped automatically. Schema changes increment the version and re-extract.
• Timeouts: Each server gets 20 seconds to respond. Servers that fail are logged and skipped — extraction continues.
• Credentials: Servers that require env vars (Slack, Codacy, etc.) must have those vars set in your shell before running.
• stdio servers: The command must be installed and available on PATH (e.g., npx, uvx, node).
• streamable-http servers: The server must be running and reachable at the configured URL.