docpilot-mcp

A TypeScript Model Context Protocol (MCP) server that exposes 4 tools so any AI coding agent (Claude Code, Cursor, Codex, etc.) can fetch live, version-accurate documentation and code examples for any npm or PyPI package — and never again hallucinate a stale method signature.

The server runs locally on stdio, persists a vector index in ~/.cache/docpilot-mcp/, and never calls anything other than the package's own registry, GitHub repo, and docs site.

Requirements

• Node.js ≥ 20 (this project is tested on Node 24)

• npm ≥ 11 (any package manager that understands the package.json exports map)

• For JS-rendered doc sites (Stripe, Vercel, Cloudflare, etc.) the first call will need a one-time npx playwright install chromium download (~150 MB)

Related MCP server: Documentation Retrieval MCP Server (DOCRET)

Install

git clone <this repo> docpilot-mcp
cd docpilot-mcp
npm install
npm run build

npm install is fully deterministic against the locked package-lock.json. npm run build produces dist/ and exits 0.

Run

# Foreground stdio server (consumed by an MCP client)
npm start

# Dev mode with tsx hot reload
npm run dev

# Enumerate the 4 tools + their descriptions
npm run inspect

On startup the server writes a single line to stderr:

info docpilot-mcp ready (tools=4)

It then waits for JSON-RPC messages on stdin. When the parent MCP client closes stdin the server exits with code 0.

Pointing a client at the server

Once the package is published:

# Claude Code
claude mcp add docpilot -- npx -y docpilot-mcp

# Cursor / Codex (config file)
#   command:    npx
#   args:       -y
#   other args: docpilot-mcp

From a local checkout before publishing:

# Claude Code
claude mcp add docpilot -- node /abs/path/to/docpilot-mcp/dist/index.js

Tool examples

query_docs

{
  "package": "stripe",
  "version": "5.0.0",
  "question": "how do I paginate cursor results"
}

Returns a markdown answer (≤ 2000 tokens) plus a sources: [{ url, section, snippet, score }] array. The top snippet is always a literal phrase from the package's own docs.

get_changelog

{ "package": "stripe", "version": "latest" }

Returns the 10 most recent entries from the npm registry metadata, with a transparent fallback to CHANGELOG.md on the default branch.

search_examples

{
  "package": "stripe",
  "version": "5.0.0",
  "query": "create a customer"
}

Returns up to 10 { code, path, url, language } blocks. Every url is on github.com under the package's official repo.

resolve_method

{
  "package": "stripe",
  "version": "5.0.0",
  "method": "customers.create"
}

Returns { signature, params, returns, source: { url, path, line } }. The signature is parsed from the .d.ts file inside the npm tarball; the line is 1-indexed.

Caching

• Default cache root: ~/.cache/docpilot-mcp/

• Layout: index/<ecosystem>/<package>/<version>/ for the vector store, plus a raw/ tree for cached HTML and tarballs

• All disk writes are atomic (*.tmp → rename)

• See docs/cache-format.md (added in AC-7) for the full schema

Testing

npm test            # runs vitest run
npm run test:watch  # watch mode

The test suite uses recorded fixtures under test/fixtures/ so it stays fully offline and deterministic.

Project layout

src/
  index.ts              # entry point — wires transport + handlers
  server.ts             # MCP Server factory + tool registration table
  tools/                # the 4 tool handlers (one file per tool)
  sources/              # fetchPage, registry clients, GitHub resolver
  extractors/           # TS / Python / markdown chunkers
  index/                # embeddings + hnswlib store
  browser/              # playwright singleton pool
  net/                  # rate-limited, retrying HTTP client
  cache/                # content-addressed paths
  util/                 # logger, error helpers
test/                   # vitest specs + fixtures
docs/
  references/           # recorded raw materials (browser snapshots, registry JSON)
  cache-format.md
  verified-packages.md

License

MIT