MCP server to read Verso-generated documentation sites (e.g., the Lean Language Reference).
Runs on your machine from its package; the steps are below.
At most: only reads or looks up information.
Odel charges nothing per call. The publisher may charge for its own service.
Runs locally over stdio
This server isn't hosted — your MCP client launches it from a package registry. Use one of the commands below, or drop the config into your client (e.g. Claude Desktop).
Run
uvx verso-mcp
MCP client config
{
"mcpServers": {
"verso": {
"command": "uvx",
"args": [
"verso-mcp"
]
}
}
}An MCP (Model Context Protocol) server that lets an AI agent search and read documentation built with Verso, Lean's documentation authoring tool.
Verso powers most of the Lean ecosystem's reference docs and books — the
Lean Language Reference,
Functional Programming in Lean,
Theorem Proving in Lean 4,
and more. Verso Manual-genre sites publish a machine-readable cross-reference
index (xref.json); this server consumes that index and the rendered HTML.
Warning
The format of the xref.json files that this server depends on is
Verso-internal and may change at any time. Such changes could render this
server non-functional. I'll try to keep up with any such changes, but can't
make any promises.
Point it at one or more Verso sites and an agent gets four read-only tools:
list_sites, list_kinds, search, and fetch_page.
| Tool | Description |
|---|---|
list_sites | Enumerate the configured Verso sites and their aliases. |
list_kinds | List a site's entry kinds (tactics, terms, sections, options, …) with counts. |
search | Name-ranked search over a site's cross-reference index, with kind filtering and pagination. |
fetch_page | Fetch a page — or a single #anchor entry — from a site and return it as Markdown. |
list_kinds, search, and fetch_page take an optional site argument (an
alias from list_sites); omit it to use the default site. All tools are
read-only and accept a response_format of markdown (default) or json.
Entry "kinds" are derived dynamically from each site's xref.json, so
project-specific domains (Lake commands, error explanations, …) are picked up
automatically — nothing about a particular site is hard-coded.
Set the VERSO_MCP_SITES environment variable to a comma-separated list of
alias=url pairs. A bare URL (no alias=) gets an alias derived from its path.
VERSO_MCP_SITES="lean-reference=https://lean-lang.org/doc/reference/latest/,
fpil=https://lean-lang.org/functional_programming_in_lean/,
tpil=https://lean-lang.org/theorem_proving_in_lean4/"
The first site listed is the default. If VERSO_MCP_SITES is unset, the server
defaults to a single site, the Lean Language Reference.
A site URL must be the root directory that contains xref.json (Verso writes it
there for Manual- and Tutorial-genre sites). The configured site roots also
serve as the network allowlist — see Safety.
uv. uv fetches the package and its dependencies
automatically on first launch — no manual virtualenv or pip install step.
Add an entry to your MCP configuration (.mcp.json, ~/.claude.json, or
claude_desktop_config.json):
{
"mcpServers": {
"verso": {
"type": "stdio",
"command": "uvx",
"args": ["verso-mcp"],
"env": {
"VERSO_MCP_SITES": "lean-reference=https://lean-lang.org/doc/reference/latest/, fpil=https://lean-lang.org/functional_programming_in_lean/"
}
}
}
}
All optional:
| Variable | Default | Purpose |
|---|---|---|
VERSO_MCP_SITES | Lean Language Reference | Comma-separated alias=url site list (see above). |
VERSO_MCP_CACHE | ~/.cache/verso-mcp | Cache directory (each site cached in its own subdirectory). |
VERSO_MCP_RATE_PER_SEC | 2 | Sustained outbound request rate (requests/second). |
VERSO_MCP_RATE_BURST | 5 | Token-bucket burst capacity. |
VERSO_MCP_RATE_MAX_WAIT | 3 | Max seconds to wait for a token before refusing. |
Warning
MCP servers have a lot of
risks.
As far as MCP servers go, this one (verso-mcp) should be relatively
innocuous: it is read-only, runs no shell commands, and only reaches the
documentation sites you configure (or just the Lean Language Reference, if
left unconfigured). The main caveat is that a malicious or compromised
documentation site could try to steer the model via indirect prompt injection.
verso-mcp cannot prevent that, so don't let an agent that uses it take
consequential actions without your review.
The server is built to be a well-behaved client of documentation sites:
.., %2e%2e, backslash variants) is
rejected.robots.txt — each host's robots.txt is fetched and respected
for this server's User-Agent; a host can target it specifically with a
User-agent: verso-mcp group.xref.json and pages are cached on disk per
site (24 h TTL) with ETag/If-None-Match conditional revalidation, so a
repeated lookup costs at most a 304 Not Modified.User-Agent on every request.evaluation.xml is a 10-question evaluation suite in the format used by
Anthropic's mcp-builder skill. The questions target the default site (the
Lean Language Reference); each is read-only, independent, and has a single
stable, verifiable answer.
xref.json).
Blog-genre sites have no xref.json. Tutorial-genre sites also emit one and
should work, but are untested.search is name-based — it matches entry names and titles in the
cross-reference index, the same granularity as Verso's own on-site search. It
does not do full-text search of page bodies.xref.json schema is an undocumented Verso internal; it may shift between
Verso releases, which could break this MCP server if it doesn't keep up.This project was written almost entirely by Claude Opus 4.7 (1M), an AI assistant. It was built for my own personal use and is shared here only in case it is useful to others. It comes with no warranty whatsoever. Use it at your own risk.