rtfm

rtfm — Read the F- Full Manual.

A local MCP server with thin Claude Code and OpenCode fronts that makes large, can't-read-by-hand technical documentation searchable — PDF spec sets, documentation repositories (and, later, public doc sites) — and answers questions strictly from that material, with grounded citations. The server is published as the rtfm-mcp Python package.

The eternal advice "read the manual" falls apart when the manual is a 400-page protocol spec or a vendor doc-set no human can hold in their head. rtfm ingests the full manual so the answer is one cited query away.

Limitations

rtfm indexes the extractable text a document already carries. It does not OCR images and does not execute JavaScript, so image-only/scanned PDFs and image-only or JS-rendered web pages are out of scope — they index as empty. (See ADR 0005.)

Install

Claude Code

# in Claude Code:
/plugin marketplace add canlin-zhang/rtfm
/plugin install rtfm@rtfm

OpenCode

opencode plugin add 'github:canlin-zhang/rtfm#main::path:plugins/opencode'

The plugin registers the rtfm MCP server and the read-the-manual skill.

Any MCP client (PyPI)

The server is also published as rtfm-mcp:

uvx --from rtfm-mcp rtfm-server      # or: pipx install rtfm-mcp && rtfm-server

Point the client's MCP configuration at that command:

{
  "mcpServers": {
    "rtfm": {
      "command": "uvx",
      "args": ["--from", "rtfm-mcp", "rtfm-server"]
    }
  }
}

Prerequisites: the plugin builds a private venv under $CLAUDE_PLUGIN_DATA/venv (Claude Code) or ~/.rtfm/fronts/<front>/venv (otherwise), installs the pinned rtfm-mcp there once, and runs it from there. It uses uv when present (uv venv + uv pip install); without uv it falls back to python3 -m venv + pip, so python3 with a working venv/pip module is the hard requirement (Debian/Ubuntu: python3-venv). Later launches do not touch the network. See ADR 0019. The manual uvx/pipx commands above are for a one-off MCP client config, not the plugin. Optional: install poppler-utils (pdftotext) for best-quality PDF extraction.

Platform support: Linux and macOS. Windows is not officially supported.

Two ways to use it

Quick-and-dirty (the default drop-dir). Drop PDFs or Markdown/text into ~/.rtfm/default and just search — the default dir re-indexes itself on search, so dropped files show up with no extra step. Keep it small; it's a scratch space.

Organized corpus (pointed-to folders). For a large, structured doc set (e.g. a vendor's documentation tree), add a dir source in ~/.rtfm/manifest.toml pointing at the folder, then build it once:

reindex("vendor-docs")

Searches are instant afterward. dir sources auto-reindex inline when the delta is small (RTFM_AUTO_REINDEX_MAX, default 10); a larger delta returns a warning telling you to run reindex once. git_repo sources are the exception to the budget: a stale or never-indexed source always auto-reindexes inline on search — no budget gate, warns loudly on failure — including the initial clone for managed sources, so the first search over a fresh corpus can take a while. Byte-identical files that repeat across version subfolders are extracted once; find_duplicates lists every path a given content lives at.

Use

Drop PDFs / .md / .mdx / .rst / .txt into ~/.rtfm/default/, then ask your agent to search them, or add more sources by editing ~/.rtfm/manifest.toml (see manifest.example.toml).

Or invoke the bundled read-the-manual skill — /rtfm:read-the-manual in Claude Code, @read-the-manual in OpenCode — to have a question answered strictly from the corpus: verbatim quotes with page/line citations, no model-knowledge fallback.

Status

Early but working. Today rtfm gives you zero-config search over a default drop-folder: PDF, Markdown, reStructuredText, and text files are indexed into one content-addressed full-text index with page/line citations, exposed through the search / read / reindex / find_duplicates / list_sources / health_check tools and the read-the-manual skill. The core is published as rtfm-mcp; the Claude Code and OpenCode fronts are thin. HTML with anchor citations, web sources, file-curation tools, and OCR are planned — see the roadmap.

Architecture decisions live in docs/adr/; domain language in CONTEXT.md.

License

MIT © Canlin Zhang