~/nutrx-labs

$ cd ~/nutrx-labs/tools/rtfm

rtfmRetrieval Tool For Manualsv1.13 · MIT

Your agent reads the docs, so you don't paste them.

rtfm indexes the documentation you already have, from Confluence exports and PDFs to draw.io diagrams and SQL schemas, into a search index on your machine. Your coding agent asks it questions over MCP and gets back the relevant passages, with their sources.

  • cli
  • mcp server
View on GitHub ↗
alex@nutrx-labs: ~/rtfmzsh
# make AI agents read the manual
dotnet tool install -g Rtfm.Cli
dotnet tool install -g Rtfm.Mcp
rtfm init

// what it does

What it does

Exact terms and vague questions

Lexical search finds the endpoint name you typed verbatim. Local semantic embeddings find the page that explains the concept in other words. A local reranker then puts the best passages first.

Reads each format properly

Confluence .doc exports are unpacked from MHTML, PDF headings are inferred, draw.io diagrams become shapes and connections, SQL dumps become tables with foreign keys, and images are OCR'd. Not a dumb text dump.

Nothing leaves your machine

OpenSearch runs in a local container, the models run in-process and OCR ships inside the tool. No external APIs, no cloud accounts. Offline, search steps down to lexical-only instead of breaking.

Works with the agent you use

A standard stdio MCP server: Claude Code, Cursor, VS Code, Zed, Windsurf, Cline, Claude Desktop. rtfm mcp-config prints the right snippet for each client, or merges it into the config file for you.

Keeps up with the source

rtfm watch picks up edits, renames and deletes within seconds. Jira tickets and Confluence pages can be pulled straight from Atlassian Cloud, read-only, and re-indexed when they change.

Schema first, then the query

Point a .rtfmdb file at Postgres or SQL Server and the live schema is indexed. Opt in, and the agent can query rows too: read-only unless you allow writes, capped at 500 rows by default.

// how it works

How it works

  1. Index a folder of docs

    Point rtfm at a folder and name the project. Every supported file is converted to markdown, split into heading-aware chunks and indexed. The first run downloads the embedding and reranker models, about 90 MB each, once.

    rtfm index ~/docs/myproject --project myproject

  2. Check it from the terminal

    Search before you hand it to an agent. You get the top five hits with a score, the heading breadcrumb, the source file and its last-modified date.

    rtfm search "something you know is in the docs" --project myproject

  3. Wire up your agent

    Generate the MCP config for your client, or let rtfm merge it in place, with a .bak backup. The .mcp.json is the same in every repo except RTFM_PROJECT, so it is safe to commit and each repo searches its own docs.

    rtfm mcp-config --client claude-code --write

  4. Ask a question

    Ask your agent in plain words. It calls search_docs, pulls the surrounding section with get_document when a snippet is not enough, and answers from your docs instead of from memory.

  5. Keep the index fresh

    Run the watcher and it catches up on anything that changed while it was off, then follows edits as they happen across every folder you have indexed.

    rtfm watch --all

// good to know

Good to know

  • Needs .NET 10 and Docker. On native Linux, also set vm.max_map_count=262144 or OpenSearch will not start. Runs on Windows, macOS and Linux.
  • On macOS and Linux, ~/.dotnet/tools is often not on your PATH after install. dotnet tool install warns you; add it to your shell profile once.
  • Each running MCP server loads its own embedding model, roughly 100 to 200 MB of RAM per instance.
  • Jira and Confluence access is read-only and needs an Atlassian API token in an environment variable. rtfm stores the variable's name, never the token.
  • The database read guard stops accidental writes. It is not a security boundary. Point .rtfmdb files at local or test databases, and at production only on purpose.
  • PDFs come out flatter than Word exports, and their tables are extracted as text in reading order, not rebuilt as tables.

// the full manual

Everything else is in the README.

Configuration, every command, troubleshooting and the design notes.

Read the README ↗
© 2026 nutrx labs