mcp-server-rfc
An MCP server that reads the RFC corpus with 40x fewer tokens and flags
obsoleted specs.
No clone, no virtualenv, no path. That is the whole installation. Your client
spawns the server on your machine and talks to it over stdin and stdout.
A client showing the server as connected has not necessarily used it, and a
model will happily answer an RFC question from memory instead. On Claude Desktop
the per-server log records only the handshake, so it cannot tell you either way
— docs/desktop-verification.md says where the evidence actually is.
Tools
With the section index, a full-text query is words, and each result is one
section ranked by BM25, carrying the section number to pass to get_rfc and the
line that best matches. Without it, full-text search falls back to rg or
grep over a synced mirror, where the query is a literal string and results are
whole RFCs; regex=true opts into pattern matching there and in titles. The
payload names the backend in tool. Title results put current RFCs before
superseded ones, so a page cut short at limit keeps the document you want
rather than the one it replaced.
get_rfc refuses any read of more than 1500 lines that is not a named section,
answering with its size and a pointer to list_sections rather than filling the
context with a specification the model had one question about; a start_line of
1 does not get around it. full=true overrides it. A section you asked for by
name is never refused: list_sections reports how long
each one is, and max_lines caps any read, so the choice is informed rather
than blocked.
Every response carries a banner with the RFC's status and, when it applies, a
warning that it has been superseded:
That warning is the reason to use this rather than let a model fetch
rfc-editor.org itself. Models cite dead specifications with great confidence.
Full-text search
Works with no setup, fetching documents on demand. Full-text search additionally
needs the section index, which the user downloads from a shell:
About 270 MB from this project's GitHub releases, checked against its published
SHA-256. It ranks sections by BM25 and serves every read from disk, so the
server stops fetching from the RFC Editor. uvx mcp-server-rfc sync instead
copies the corpus as 512 MB of text files from the RFC Editor's volunteer-run
rsync mirror, which is what rg and regex=true search.
The same command that runs the server runs its CLI: sync downloads, status
reports what is present, and uvx mcp-server-rfc <command> --help covers the
rest. It is also the command the server names in the error you get when
full-text search is asked for without either.
Both are optional. The server deliberately does not expose a sync tool: hundreds
of megabytes on someone's disk should be a person's decision, not a model's.
If an index or a mirror already exists — synced for the companion skill, say — this server
picks it up automatically. It looks under the platform's data directory:
$XDG_DATA_HOME/rfc-ai-tooling, or %LOCALAPPDATA%\rfc-ai-tooling on Windows,
falling back to ~/.local/share/rfc-ai-tooling. Set $RFC_MIRROR to point at a
non-default location — in a container with no home directory, that or --mirror
is what to pass.
In a container, uvx on the host would write to the wrong disk. Keep the mirror
in a named volume, fill it through the image, and set $RFC_CLI so the server's
errors name that command instead of uvx mcp-server-rfc:
HTTP mode
The config above is the stdio transport, which needs a client that can spawn a
process on your machine. A session running in the cloud — claude.ai on the web,
Cowork — cannot, and no configuration fixes that. For those, the server has to
be deployed somewhere they can reach:
RFC_TRANSPORT, HOST and PORT are read from the environment too. The
repository carries a Dockerfile and a fly.toml that do this. No public
instance of this server exists and none is planned — if you want one, you host
it, and you should put authentication in front of it, since the server has none
and would otherwise be an open proxy to a volunteer-run mirror.
Full-text search works hosted only where the section index is on the
server's disk (sync --index there). Without it, a hosted instance returns an
error saying so instead of silently answering a title search. The obsolescence
banner, section reads and the unscoped-read guard work either way, because they
need only the 2 MB RFC index.
If your client has a shell
Use the skill instead — Claude Code, Codex CLI, OpenClaw, Cursor and Zed all
qualify. It does the same things with fewer moving parts and gives the agent
ripgrep over the corpus:
This server is for clients that give the model no shell of its own. Claude
Desktop is the clearest case: its chat cannot run a command on your machine, but
it will spawn this server there.
Source, issues and the skill: https://github.com/shbernal/rfc-ai-tooling
MIT.
