Wikipedia Mcp Server

Overview

Wikipedia in any language edition. Search articles, read summaries or just the sections you need, list the footnotes an article cites, find geotagged pages near a coordinate, and list an article's language editions. Ships the wikipedia-guide agent skill (skill://wikipedia-guide/SKILL.md), which MCP clients read for the gates that span tools: switching editions, disambiguation pages, section numbers versus section_index, and what to check before reporting something absent.

Tools

Capability reference

wikipedia_search_articles _tool

• Free-text query (an empty one fails with empty_query), positive integer limit (default 10, clamped to 50), and offset paging, passing back enrichment nextOffset. Only a clamped call carries requestedLimit, appliedLimit, and a clamp notice. Wikipedia serves no result past the 10,000th: an offset at or beyond it fails with offset_too_large, and a page ending on that window carries enrichment truncated
• CirrusSearch operators include incategory:"1648 treaties", morelike:Peace of Westphalia, intitle:Python, hastemplate:Infobox, and insource:Python (support depends on the installation). sort defaults to relevance; alternatives are last_edit_desc, create_timestamp_desc, incoming_links_desc, and title_natural_asc. Enrichment echoes sort; keep it and query unchanged when paging
• Each result carries a plain-text snippet, page ID, word count, and — when the article has them — its short description and Wikidata QID (wikibase_item); enrichment suggestion carries Wikipedia's spelling correction (einstien → einstein)
• Invalid or over-complex insource:/…/ and intitle:/…/ regexes fail once with invalid_query, the upstreamCode, and guidance to correct or simplify the regex or remove the operator. A busy regex search remains retryable

wikipedia_get_summary _tool

• Takes a title; returns the REST summary extract — a fragment from the start of the lead (wikipedia_get_article with section_index: 0 reads the whole lead) — plus the short description, thumbnail URL, Wikidata QID (wikibase_item), and, for a geotagged article, latitude / longitude to pass straight to wikipedia_search_nearby
• page_type discriminates standard / disambiguation / no-extract (on disambiguation, re-query with wikipedia_search_articles); url, revision_id, and last_modified name the revision read, and ?oldid=<revision_id> is a permanent link to it
• An existing page without a summary extract succeeds with metadata, an empty extract, and guidance to open the canonical URL or try section_index: 0. Namespace pages may contain documentation rather than article prose. An empty description is absent; a missing title still fails with not_found

wikipedia_get_article _tool

• Without section_index: the full article with == Section == markers, or a section outline (truncated: true) when it exceeds WIKIPEDIA_ARTICLE_OVERFLOW_BYTES (default 80,000 bytes). With a section_index from wikipedia_get_sections (0 is the lead): that section plus its nested subsections
• Outline entries include verified section_index values where available, including 0 for the lead. Ambiguous or unindexed headings retain the section-list fallback. TOC verification gets one short attempt; unavailable verification is disclosed
• Redirects expose redirected_from and, for a section target, redirect_section. A full read remains full and includes targeted-read guidance; an explicit section_index, including 0, takes precedence
• Every read returns url and revision_id (?oldid=<revision_id> is a permanent link to the text read); a full read, outline included, also returns last_modified. With WIKIPEDIA_BASE_URL set, a section read omits url
• Section reads render phylogenetic clade trees as fenced, indented branches, pairing each row’s label with its leaf and preserving unnamed forks. An unparseable tree leaves [table omitted: unparseable clade]; trees have no table_index
• Section reads also carry data tables as pipe-delimited rows and infoboxes as label: value lines, which the full-article extract strips; a table over 40,000 rendered bytes leaves a [table omitted: N rows (table T)] marker. Pass table_index: T with the section to read that table in complete row pages, then follow next_cursor as cursor with the same inputs. Continuations read the first page’s revision; caption and header appear once. Each page's table text fits 40,000 UTF-8 bytes on both surfaces, except an indivisible oversized row or first-page caption/header returned alone with oversized_row: true

wikipedia_get_sections _tool

• Takes a title; returns section titles, heading levels, hierarchical numbering (e.g. "2.1"), and the section_index to pass to wikipedia_get_article — the first entry is the lead (index: 0, titled Introduction)
• Heading titles use the article renderer’s script, formula, whitespace and single-pass entity normalization, so CO₂ stays identical across the section list, article reads and verified outline entries
• Fails with no_sections on a stub or very short article — read it with wikipedia_get_article instead
• Redirect aliases expose redirected_from and optional redirect_section; redirect_section_index appears only when a unique anchor matches a real numeric TOC index

wikipedia_search_nearby _tool

• latitude −90…90 / longitude −180…180 (WGS 84), radius_meters 10–10,000 (default 1000), and positive integer limit (default 10, clamped to 500), with no pagination past limit; enrichment truncated flags more matches than limit allowed — at 500, Wikipedia's ceiling, sweep narrower radii instead. Only a clamped call carries requestedLimit, appliedLimit, and a clamp notice
• Returns articles sorted by distance_meters, with coordinates, short description, and Wikidata QID (wikibase_item). Each is placed by the article's own GeoData coordinate tag, not its Wikidata item's coordinate, so a mistagged article appears where the tag puts it

wikipedia_get_languages _tool

• Takes a title; returns each edition's language_code, edition_code, article title, and URL, plus source_title, the resolved title. Pass edition_code — not language_code — as language on other tools; the two can differ (gsw vs als)
• editions narrows the list to the codes asked for: codes with no article come back under missing, and total_languages stays the unfiltered count. Fails with no_other_languages when the article has no translations; a filter matching nothing is an empty list, not a failure

wikipedia_get_references _tool

• Takes a title; returns references[] in document order across every reference list, each with id, label (the in-text marker), group (a notes list such as lower-alpha, absent for the main list), text, urls (archive copies included), and dois (from doi.org links, percent-decoded). total_references counts the footnotes on all pages. wikipedia_get_article drops footnote markers and reference lists
• A short citation such as "Galinsky 2012, pp. 75–76" carries work_ids, and works on the same page holds the full citations they name, each once. Bibliography and further-reading entries no footnote points to are not listed — read those sections with wikipedia_get_article
• Each page holds whole footnotes within 40,000 UTF-8 bytes on both surfaces; follow next_cursor as cursor with the same title and language. Continuations read the first page's revision_id, and a cursor past the last footnote fails invalid_cursor. An unreadable pinned revision fails revision_unavailable. An article with no footnotes returns an empty list with a notice, not an error

Features

Built on @cyanheads/mcp-ts-core: stdio and Streamable HTTP transports, pluggable auth (none / jwt / oauth), swappable storage (in-memory, filesystem, Supabase, Cloudflare KV/R2/D1), structured logging with optional OpenTelemetry tracing.

Wikipedia-specific:

• Dual API integration — MediaWiki REST API (/api/rest_v1/) for summaries, Action API (/w/api.php) for search, full text, sections, geo search, and language links
• Retry and backoff on every required request, including the transient refusals (search too busy, rate-limited, read-only) the Action API returns inside an HTTP 200 body, with an upstream Retry-After honored and each request's retries bounded at 30 s. The best-effort description lookup on search results gets one short attempt; User-Agent header per Wikimedia API policy
• Every read path renders HTML through one renderer to the same plain-text shape — == Heading == markers, one list item per line, superscripts kept, code fenced: the full article from the Action API's HTML extract, a section from the parser's own HTML for that section, the summary from the REST extract_html. A section read additionally carries data tables, infoboxes, and the lists inside layout tables, none of which the extract carries. Footnote markers, reference lists, and the hidden Citation Style 1 maintenance messages (CS1 maint: …) are dropped on every path; wikipedia_get_references returns the footnotes
• Per-call language parameter on every tool (default en) — all Wikipedia language editions accessible in a single session; every title read follows redirects
• Language validation against a live edition registry built from the MediaWiki action=sitematrix endpoint (cached 24h) — catches structurally valid but nonexistent editions before they cause timeouts

Agent-friendly output:

• page_type on summaries discriminates standard / disambiguation / no-extract — no string parsing needed
• wikibase_item (Wikidata QID) on summaries and on search and nearby results enables direct cross-referencing with wikidata-mcp-server
• Article text, snippets, and titles are backslash-escaped on the way into the markdown content[] render, so an article that writes about markup or markdown syntax reads as itself instead of being interpreted by the client; structuredContent carries the same text unescaped. Fenced code blocks and table-row delimiters pass through unescaped, while table cells stay escaped
• section_index on table-of-contents entries links directly to the targeted-read parameter on wikipedia_get_article, index 0 included
• Titles MediaWiki cannot name a page with — < > [ ] { }, the | multi-title separator, percent escapes, magic tildes, relative paths — are refused before any request, with invalid_title; a trailing #fragment is accepted and resolves normally
• Recovery hints on every error type — callers get actionable next steps (e.g., "use wikipedia_search_articles to find the correct title")

Getting started

Connect to the hosted instance, no install:

claude mcp add --transport http wikipedia-mcp-server https://wikipedia.caseyjhand.com/mcp

Or run it locally:

claude mcp add wikipedia-mcp-server -- npx -y @cyanheads/wikipedia-mcp-server

For other clients, add this to the MCP config file:

{  "mcpServers": {    "wikipedia-mcp-server": {      "command": "npx",      "args": ["-y", "@cyanheads/wikipedia-mcp-server"]    }  }}

No API key is required — Wikipedia's API is public. For Docker or a source build, see Running the server.

Configuration

See .env.example for the full list of optional overrides.

Running the server

Local development

Requires Bun v1.4.0 or higher (or Node.js v24+).

• Clone and install:

  sh

  git clone https://github.com/cyanheads/wikipedia-mcp-server.gitcd wikipedia-mcp-server && bun installcp .env.example .env   # optional: customize WIKIPEDIA_USER_AGENT or logging

• Build and run:

  sh

  # One-time buildbun run rebuild
  # Run the built serverbun run start:stdio# orbun run start:http   # listens at http://localhost:3010/mcp (MCP_HTTP_PORT)

• Run checks and tests:

  sh

  bun run devcheck   # Lint, format, typecheck, securitybun run test       # Vitest test suitebun run lint:mcp   # Validate MCP definitions against spec

Docker

docker build -t wikipedia-mcp-server .docker run --rm -p 3010:3010 wikipedia-mcp-server

The Dockerfile defaults to HTTP transport, stateless session mode, and logs to /var/log/wikipedia-mcp-server. To run the published image over stdio from an MCP client, use "command": "docker" with "args": ["run", "-i", "--rm", "-e", "MCP_TRANSPORT_TYPE=stdio", "ghcr.io/cyanheads/wikipedia-mcp-server:latest"]. OpenTelemetry peer dependencies are installed by default — build with --build-arg OTEL_ENABLED=false to omit them.

Project structure

Development guide

See CLAUDE.md for development guidelines and architectural rules. The short version:

• Handlers throw, framework catches — no try/catch in tool logic
• Use ctx.log for request-scoped logging, ctx.state for tenant-scoped storage
• Register new tools in src/mcp-server/tools/definitions/index.ts
• Wrap external API calls: validate raw → normalize to domain type → return output schema; never fabricate missing fields

Contributing

Issues are welcome. Run checks and tests before submitting:

bun run devcheckbun run test

License

Apache-2.0 — see LICENSE for details.