Native pi extension port of the RTFD MCP server — feed your coding agent up-to-date documentation from PyPI, npm, crates.io, GoDocs, DockerHub, GitHub, GCP, Zig, and LogScale, without a cloud service or API key.

RTFD runs entirely on your machine. Queries go straight to the source, and the documentation you fetch never leaves your system. This extension brings that same functionality to the pi coding agent as a native extension — no MCP server process, no mcp.json entry, just a symlink and pi auto-discovers it.

Why?

LLMs hallucinate APIs and ship outdated code because their training data is months or years old. RTFD gives agents access to the actual current documentation for a library, image, or service, so generated code uses correct, version-specific APIs instead of guessed ones.

• Accurate: fetch live READMEs, metadata, and version info straight from the source.
• Private: runs locally; your queries go directly to the upstream registry, nothing in the middle.
• No API keys: uses public endpoints (optional GITHUB_TOKEN / gh for higher rate limits).
• 10 providers, 30+ tools: one provider per ecosystem, plus cross-ecosystem aggregation and chunked streaming.

Providers & tools

Provider	Example tools
PyPI	pypi_metadata, fetch_pypi_docs
npm	npm_metadata, fetch_npm_docs
crates.io	search_crates, crates_metadata
GoDocs	godocs_metadata, fetch_godocs_docs
Zig	zig_docs
DockerHub	search_docker_images, docker_image_metadata, fetch_docker_image_docs, fetch_dockerfile
GitHub	github_repo_search, github_code_search, fetch_github_readme, get_repo_tree, list_repo_contents, get_file_content, get_commit_diff, list_github_packages, get_package_versions
GCP	search_gcp_services, fetch_gcp_service_docs
LogScale	search_logscale_docs, list_logscale_functions, logscale_syntax, logscale_function
Aggregator	search_library_docs (PyPI + npm + crates + GoDocs + GitHub at once)
Server	get_cache_info, get_cache_entries, get_next_chunk

Large responses are chunked with a continuation token — call get_next_chunk to retrieve the next slice.

Install

Prerequisites

• The pi coding agent (peer dependency).
• Node.js with npm.

From npm (recommended)

pi install npm:rtfd-pi

Restart pi (or run /reload) and the tools are available — no -e flag or mcp.json entry needed. Pi installs the package under ~/.pi/agent/npm/ and auto-discovers the extension via the pi.extensions manifest.

From git

pi install git:github.com/aserper/rtfd-pi
# or
pi install https://github.com/aserper/rtfd-pi

As a local extension (developer setup)

git clone https://github.com/aserper/rtfd-pi.git ~/projects/rtfd-pi
cd ~/projects/rtfd-pi
npm install

# Link into pi's extensions directory so pi auto-discovers it
ln -s ~/projects/rtfd-pi ~/.pi/agent/extensions/rtfd-pi

Restart pi and the tools are available — no -e flag or mcp.json entry needed.

Configuration

All configuration is via environment variables (read by the pi process). Defaults match the original RTFD MCP server.

Variable	Default	Purpose
RTFD_FETCH	true	Set false to disable all content fetching (metadata-only mode).
RTFD_CACHE_ENABLED	true	Enable the in-memory response cache.
RTFD_CACHE_TTL	604800	Cache TTL in seconds (1 week).
RTFD_CHUNK_TOKENS	provider-specific	Max tokens per chunked response.
GITHUB_TOKEN	—	GitHub token for higher rate limits.
GITHUB_AUTH	token	token (use GITHUB_TOKEN), cli (use gh auth token), auto (token then gh), or disabled.

GitHub auth tip

For the GitHub provider's tools to work with higher rate limits, set:

export GITHUB_AUTH=auto        # falls back to `gh auth token` if GITHUB_TOKEN is unset
# or, explicitly:
export GITHUB_TOKEN=gho_...    # a personal access token

If gh is installed and authenticated, GITHUB_AUTH=auto will pick up its token automatically.

Development

npm test              # run the vitest suite
npm run test:watch    # watch mode
npm run test:unit     # shared/ unit tests only
npm run typecheck     # tsc --noEmit

The test suite includes parity tests against the original Python RTFD server's output shape, plus golden-fixture tests for the providers with recorded cassettes (PyPI, npm, crates).

Project layout

src/
  providers/      # one file per ecosystem (pypi, npm, crates, github, ...)
  shared/         # http, cache, chunking, env, tokens, content extraction
  tools/          # server-level tools: search_library_docs, cache admin, get_next_chunk
  index.ts        # extension factory — registers all provider + server tools
tests/
  parity/         # parity tests vs. the python RTFD server
fixtures/         # recorded HTTP cassettes (pypi, npm, crates)
docs/             # PORT-SPEC, PLAN, AUDIT-FINDINGS

How it works

Each provider exposes its upstream API as a set of pi tools. The shared HTTP client and cache are singletons, so all providers share one cache and one chunking manager. search_library_docs fans out across PyPI, npm, crates.io, GoDocs, and GitHub in a single call. Large responses are token-counted and chunked; get_next_chunk retrieves successive slices via a shared continuation-token store.

Security note

This extension grants agents access to unverified content from external sources (GitHub, PyPI, etc.). That introduces risks including indirect prompt injection. Use RTFD_FETCH=false to restrict to metadata-only lookups, and review fetched content before acting on it in autonomous modes.

Releasing

Releases are automated through GitHub Actions. To publish a new version to npm:

npm version patch   # or minor / major — bumps package.json, commits, tags vX.Y.Z
git push --follow-tags

Pushing the v* tag triggers the Release workflow, which runs typecheck + tests, then publishes to npm with provenance. GitHub and npm versions stay in lockstep because the tag matches package.json.

The NPM_TOKEN repo secret must be set (see GitHub Actions secrets) with a granular npm access token that has publish rights to rtfd-pi.

License

MIT — see LICENSE.