The design leans hard on not doing damage: writes only appear once you authenticate, they vanish from the tool list when the wiki is read-only, and update-page takes a latestId so a concurrent edit blocks yours rather than silently overwriting it. parse-wikitext lets you preview a planned change against the live wiki before create-page or update-page ever touches it.
A server that lets an LLM client work with any MediaWiki wiki: 29 tools spanning page reads, page writes, files, wiki discovery and OAuth session management, with each call optionally naming which configured wiki to act on.
- Pages read singly or in clusters — get-page returns wikitext, rendered HTML or metadata only, and get-pages pulls up to 50 titles at once, reporting missing ones inline instead of failing
- get-page with metadata=true hands back the revision id used for edit-conflict detection, plus page size and section outline
- search-page runs full-text search over titles and content with snippets, and search-page-by-prefix returns up to 500 bare titles for autocomplete
- History and change tracking: get-page-history in segments of 20, get-recent-changes in segments of 50 filtered by type, namespace, user or tag, get-revision for one historical version, and compare-pages for a compact diff between two revisions, titles, or supplied wikitext
- parse-wikitext renders wikitext through the live wiki without saving, returning warnings, categories, wikilinks, templates and external URLs — a dry run before you commit an edit
- Writes gated behind auth: create-page, update-page (pass latestId for edit-conflict detection), move-page with or without leaving a redirect, and delete-page / undelete-page as a soft delete you can reverse
- Files handled four ways — upload-file and update-file from local disk, upload-file-from-url and update-file-from-url from a remote URL — plus get-file for metadata and get-file-data to send the image bytes inline to a model that cannot reach the wiki host
- get-category-members, get-links-here (wikilinks, transclusions or file usage), get-site-info and whoami cover navigation and the current identity edits will be attributed to
- Several wikis at once: list-wikis, add-wiki registering one as mcp://wikis/<servername>, remove-wiki, and oauth-status / oauth-logout managing stored tokens
Node with npx, launched as @professional-wiki/mediawiki-mcp-server with a CONFIG path pointing at your config.json. Reading is anonymous — no account needed — but write and file tools require authentication, and are hidden from the tool list entirely when the default wiki is set readOnly. The recommended sign-in is browser-based OAuth (oauth2ClientId and oauth2CallbackPort per wiki); a bot password is the fallback where Extension:OAuth is not installed. It runs over stdio by default, or over StreamableHTTP with MCP_TRANSPORT=http on PORT 3000.
One command plus a key — npx -y @professional-wiki/mediawiki-mcp-server@latest, then supply credentials