The canonical Python stdlib oracle for AI coding agents — always free, always MIT, token-frugal.
io.github.ayhammouda/python-docs-mcp-server MCP Server
python-docs-mcp-server is an MCP server for AI coding agents that acts as a “canonical Python stdlib oracle.” It targets Python documentation details such as exact symbols, exact sections, and exact versions, with offline availability. The project is always free, MIT-licensed, and described as token-frugal.
🛠️ Key Features
Python stdlib documentation lookup (“oracle”) for agents
Offline use
Exact symbols, sections, and versions
Always free; MIT licensing
Topics include MCP and documentation tooling
🚀 Use Cases
Query Python standard library documentation for accurate references
Support agent workflows that need version-specific, symbol-accurate context
Use MCP-based integration for documentation retrieval
Built for the moment your agent needs asyncio.TaskGroup signatures, pathlib.Path semantics, or what changed in 3.12 — not a web fetch, not a hosted API, not a vector store hallucinating section anchors. Just an indexed slice of docs.python.org, returned by symbol or by query, scoped to the version you actually ship on.
Why this exists
There is a difference between long context and usable context. Research like the
Lost in the Middle study (Liu et al., 2023)
found that models struggle to use relevant information buried in the middle of a
long input. Pasting an entire documentation page into the model invites that
failure. Returning the exact symbol, section, and version keeps the answer in a
short, high-signal context instead.
Generic docs retrieval is a rough fit for Python stdlib questions:
asyncio.TaskGroup should resolve to the actual symbol, not a fuzzy page hit
Python version matters (3.12 and 3.13 do not always say the same thing)
fetching a whole page burns tokens when one section answers the question
the official docs are canonical, but they do not ship as an MCP server
This server indexes the official docs locally and exposes a small set of MCP
tools for lookup and section retrieval.
Why use it
no API keys to manage
queries run against a local index, not a hosted docs API
results come from the official Python docs
the server is read-only at runtime
fewer dependencies to review in strict environments
What you get
symbol lookup through Python objects.inv
page and section retrieval with truncation and pagination
a local SQLite + FTS5 index; no runtime web scraping
Call get_docs(...) using the slug and anchor returned by the best hit
Result
The model gets the matching symbol and the relevant docs section, not a
full-page dump.
30-second demo
Ask your MCP client:
In Python 3.13, how should I use asyncio.TaskGroup and what changed from older asyncio patterns?
If setup is working, the client should use search_docs for the exact symbol,
then get_docs for the matching section. Instead of generic web results or an
entire docs page, it gets official stdlib text for the requested Python version,
trimmed to the section that matters.
Install
Run directly with uvx:
bash
uvx python-docs-mcp-server --version
Or install it once with pipx:
bash
pipx install python-docs-mcp-server
If uv is installed but the uv command is not on your PATH, reopen your
shell or use python -m uv ... as a fallback for local contributor commands.
Search result behavior
For kind="auto" with an explicit version, a nonempty search with spare result
slots may append one exact inventory target when the prompt contains one
known qualified identifier in that version. Original hits, snippets, scores,
order and notes are preserved. Full results, ambiguous identifiers, other kinds
and versionless queries are unchanged; the existing empty-result fallback is
unchanged. Canonical locations are deduplicated, including inventory aliases.
The append is admitted only when the complete proposed response serialized
by Pydantic's compact model_dump_json() is at most 8000 UTF-8 bytes. This is
not a global search cap: oversized original responses remain unchanged. It is
separate from get_docs(max_chars=8000), which limits content characters.
The first build downloads Python's objects.inv files, clones CPython docs
sources, runs sphinx-build -b json, and writes an SQLite index to your local
cache. It can take several minutes.
Python 3.15 documentation (opt-in)
The source-pinned documentation targets include Python 3.10–3.15. To index
3.15 alongside 3.14, explicitly run:
A build replaces the local index with the requested versions; include every
version you want to keep. The first-run examples above intentionally remain
unchanged. The builder still marks the highest requested version as the
index default; adding this target does not select 3.15 in an existing index.
get_docs without a version uses the configured/detected default, while
versionless search spans indexed versions. Pass version="3.15" for isolated
3.15 answers. list_versions reports the actual index, not all build targets;
an unindexed explicit version raises an error, never an older-docs fallback.
This is documentation coverage, not a claim that the server runs on Python
3.15 or that a named MCP client was tested with it. Queries remain offline and
read-only after a successful build. Until it is indexed, consult the
official 3.15 docs directly.
The 3.15 target is CPython tag v3.15.0, commit
d220409a13b77de7a6c30abe16ea90db01528edc, with Sphinx 8.2.3.
See integration checks
for explicit-version and comparison examples.
Configure your MCP client
Claude Code
This repository includes a project-scoped .mcp.json for clients that support
checked-in MCP server configuration. It points at the published package:
CI runs ruff, pyright, and pytest on macOS and Linux for Python 3.12
and 3.13
subprocess-based stdio and smoke tests cover the MCP protocol pipe
packaging tests check the wheel contents and CLI entry points
retrieval regression tests cover exact symbol hits, version behavior,
missing symbols, truncation, and local-version defaults
manual MCP QA lives in
.github/INTEGRATION-TEST.md, with MCP
Inspector for local checks and Claude/Cursor for real-client checks
Contributor commands and validation steps live in
CONTRIBUTING.md.
Tools
The server currently exposes seven MCP tools:
Tool
Description
search_docs
Search Python stdlib docs by query. Supports symbol lookup (asyncio.TaskGroup), module search (json), and free-text search. Returns ranked hits with BM25 scoring and snippet excerpts.
get_docs
Retrieve a specific documentation page or section by slug and optional anchor. Returns markdown content with budget-enforced truncation and pagination. Retrieved results are cached on disk by Python docs version and request identity.
lookup_package_docs
Look up official PyPI package metadata and return package-declared documentation/homepage/source URLs. This is a controlled PyPI metadata lookup, not generic web search.
list_versions
List all indexed Python versions with metadata.
detect_python_version
Report the Python version detected in the user's environment when the server started, and whether that version is indexed.
compare_versions
Diff a Python stdlib symbol between two indexed versions. Returns `change=added
whatsnew_for_version
Browse nonempty sections of the already-indexed official What's New page for one Python version. Optional kind filters by conservative heading-hierarchy category. start_index and max_sections paginate the filtered results; the structured response is capped at 20,000 characters, and truncated excerpts include a stable anchor and get_docs follow-up hint. Canonical titles and anchors are unchanged; if one section cannot fit its metadata plus nonempty indexed text and any required continuation hint, an explicit bounded error is returned. Page size may be reduced to reserve usable excerpts. The cap counts Unicode characters in the model JSON, excluding the MCP envelope. Querying is offline.
For search_docs with kind="auto" and an explicit indexed version, the
original lookup and full-text search run first. If both return no hits, the
server looks for dotted Python identifiers in the prompt. Exactly one distinct
identifier registered in that version enables a symbol-lookup fallback; repeated
mentions count once and unknown identifiers do not count. Multiple known
identifiers leave the result empty. A result note identifies fallback matches,
which are symbol documentation, not an answer to the complete prompt. Explicit
kinds, searches without a version, nonempty results, and result budgets are
unchanged.
Why not Context7 or generic docs retrieval?
Use this server when you want precise local Python docs retrieval rather than
broad web search:
official Python docs, not scraped mirrors or summaries
exact symbol resolution from objects.inv
version-aware results for Python 3.10 through 3.14
section retrieval instead of full-page dumps
PyPI-declared docs, homepage, and source links through lookup_package_docs
local read-only runtime with no API keys
Use Context7 or a generic docs fetcher for third-party libraries, arbitrary web
pages, or framework research. This server is not a universal docs search engine;
it is a focused stdlib retrieval tool for AI coding agents.
Retrieved docs cache
get_docs responses are cached across MCP client/server restarts in the
platform cache directory:
The cache stores completed get_docs results for the resolved Python docs
version plus request identity (slug, optional anchor, max_chars, and
start_index). Cache misses use the normal local index retrieval path and then
write the result.
Cache entries are also scoped to a fingerprint of the local index.db file
(path, size, and modification timestamp). If you rebuild or replace the local
docs index, older entries are ignored automatically. Deleting
retrieved-docs-cache.sqlite3 is safe; it removes cached retrieval results, not
the docs index.
Air-gapped install
The server needs no network at query time. To install it on a machine with no
network:
On a connected machine with the same OS, CPU architecture and Python minor
version, download the wheels and build the index. Elsewhere, add pip's
--platform, --python-version and --only-binary=:all: options to
pip download.
Put index.db at ~/Library/Caches/mcp-python-docs/index.db on macOS, or
at ~/.cache/mcp-python-docs/index.db on Linux
($XDG_CACHE_HOME/mcp-python-docs/index.db when XDG_CACHE_HOME is set).
Set PYTHON_DOCS_MCP_DISABLE_AUTO_INDEX=1, so a missing index never starts
a network build, and check the install:
bash
PYTHON_DOCS_MCP_DISABLE_AUTO_INDEX=1 python-docs-mcp-server doctor
lookup_package_docs needs PyPI. Offline, it returns a clear note instead of
package sources. There is no separate offline command.
PyPI package docs lookup
lookup_package_docs queries the official PyPI JSON API documented at
https://docs.pypi.org/api/json/ (GET /pypi/<project>/json) and returns only
sources declared in that package's PyPI metadata: the PyPI project URL,
docs_url, home_page, and allowlisted project_urls labels such as
Documentation, Homepage, Source, and Repository.
The tool makes the trust boundary explicit with
trust_boundary="pypi-declared-metadata". It does not crawl pages, perform web
search, or silently fall back to unofficial community mirrors.
Diagnostics
Check the local environment:
bash
uvx python-docs-mcp-server doctor
This checks the runtime Python version, SQLite FTS5, cache/index paths, the
index format, disk space, and the venv/ensurepip support needed by
build-index. If it reports FAIL: Index format, run the build-index
command it prints; it lists the versions your index already holds (or every
supported version, if it cannot read them). If it says your index is newer
than this release, upgrade python-docs-mcp-server instead.
Validate an existing index:
bash
uvx python-docs-mcp-server validate-corpus
Troubleshooting
FTS5 unavailable
The index needs SQLite's FTS5 extension. The python.org installers,
uv python install, and the Pythons from Linux distributions, Homebrew and
conda all include it. If doctor reports FAIL: SQLite FTS5, your Python was
built against a SQLite without FTS5. Install one that has it:
If doctor says build venv support is unavailable, install the venv package
for the same Python minor version that runs the server:
bash
sudo apt install python3.12-venv
Adjust 3.12 to match the version shown by doctor. Without this package,
build-index cannot create the disposable Sphinx environment it uses to build
JSON documentation content.
Migrating from the pre-rename CLI
Earlier development snapshots of this project used the PyPI name
mcp-server-python-docs. The published PyPI project is
python-docs-mcp-server. If your MCP client config still references
the old name via uvx, you will see a Package not found error,
because uvx resolves projects by PyPI name.
Change your config args from:
json
"args":["mcp-server-python-docs"]
to:
json
"args":["python-docs-mcp-server"]
The wheel still installs a legacy mcp-server-python-docs console
script for users who already have the package installed and invoke
the binary by name on $PATH. That script is an alias and will be
removed in a future release.
uvx cache stale
If uvx python-docs-mcp-server runs an old version:
bash
uvx --reinstall python-docs-mcp-server
Or clear the uv cache:
bash
uv cache clean python-docs-mcp-server
Claude Desktop on Windows (MSIX)
The MSIX-packaged version of Claude Desktop on Windows may have restricted PATH
access. If uvx is not found, specify the full path in your config:
Replace YOU with your Windows username. Find the exact path with where uvx.
Restart after rebuild
After running build-index, restart your MCP client so it picks up the new
database file. The server opens the index read-only on startup and does not
reload it while running.
On Windows, close the MCP client before rebuilding if the index file is locked.
Tested on macOS and Linux. Windows should work, but it is not verified on
every release.
The server requires Python 3.12+ to run. Source-pinned documentation targets
cover Python 3.10 through 3.15; Python 3.15 is opt-in. The versions available
for queries depend on the documentation versions selected when building the
index.