Labsco
MCP SERVER

Librarian

by mlziade ↗

Wikipedia lookups an LLM can fact-check with — search articles, read a page, pull one section, in any language.

Encyclopedic & Cultural Reference
Summary
Section-level retrieval is what makes it cheap enough to use constantly.

Listing a page's sections and then fetching only the one that matters is the difference between a fact check costing a paragraph and costing an entire article — which decides whether verification happens on every claim or only when someone remembers to ask. The migration notes are worth reading before you wire it up: this has moved to the stateless 2026-07-28 spec, the session header and WebSocket transport are gone, and the stdio entry point was deleted, so it is HTTP-only now.

What it is

An MCP server that gives a model Wikipedia access shaped for verification rather than browsing. Five tools cover the path from a vague query to a specific passage, and a built-in prompt instructs a client to check claims proactively rather than waiting to be asked.

What you get
  • `search_wikipedia_pages` — searches articles on a topic and returns the top 5 results with the information needed to pick between them
  • `get_wikipedia_page_info` — comprehensive detail for a page, including content, summary, hyperlinked words and categories
  • `get_wikipedia_page_summary` — a quick summary with a customisable sentence length
  • `get_wikipedia_page_sections` — the list of sections on a page, which is how you avoid pulling a very long article whole
  • `get_wikipedia_page_sections_info` — the content of specific sections, addressed by title or index
  • Multi-language queries on every tool through a language parameter, defaulting to `en`
  • A `fact_checking_instructions` prompt that turns proactive verification on for the conversation
Requirements

No account and no key. Python 3.13 or higher with uv: clone, `uv sync`, then `uv run python librarian_server.py`, which starts on `http://0.0.0.0:8000` with the MCP endpoint at `/mcp`. Register it with `claude mcp add librarian -t http http://localhost:8000/mcp`. Production runs behind gunicorn with the uvicorn worker class. Communication is Streamable HTTP only, against the MCP 2026-07-28 spec — the earlier WebSocket and `/sse` endpoints have been removed, so a client configured for those will not connect. The project is `librarian` (0.1.0 in pyproject).

Setup effort

One command — npx -y mcp-remote https://your-server-host/mcp