Leahy & Co’s ownAn AI assistant that can’t wreck the wiki

A connector that lets an AI assistant search, read and edit a company wiki, with guard rails that stop it damaging the pages people rely on. In daily use.

The problem

A wiki is only useful if it’s kept current, and keeping it current is the job nobody gets to. An AI assistant can do a lot of that work, but only if it can be trusted with the pages.

The wiki’s own interface made that harder than it looked. Its update call quietly unpublished a page or wiped fields whenever it was given less than everything, and its search didn’t look inside page text at all. An assistant using it as it came could break pages without anyone noticing.

What I did

I built a connector on the Model Context Protocol, the open standard AI assistants use to work with other systems, with the safety designed in from the start.

  • Every change goes through one path that re-sends the page’s full state, so an edit can never unpublish a page or blank its title by accident.
  • The assistant edits one section at a time, and the rest of the page stays exactly as it was, byte for byte.
  • A dry run shows the exact change before it’s made. An edit that would cut a page by more than half is refused, and so is any edit to a page that changed after the assistant read it.
  • Deletes and restores need a confirmation, and the whole connector can run read-only.
  • Search covers the text of every page, and one query finds broken links and pages nothing links to.
  • It was tested with automated checks and against the real wiki, reading and writing, before it was trusted with live pages.

What changed

  • The wiki gets maintained as part of everyday work, by an assistant that can’t quietly damage it.
  • Broken links and forgotten pages turn up in one query instead of by accident.
  • It’s in daily use.

Under the bonnet

  • Node.js, the official MCP SDK and the Wiki.js GraphQL API
  • 30 tools for searching, reading, editing, links and page history
  • Shrink, conflict and confirmation guards; dry-run diffs; a read-only mode
  • A local full-text index of the whole wiki, built in about a quarter of a second
  • Unit tests, plus live read and write checks against the real wiki

Services this touched

Got something like this?

A free chat about the problem, then a written quote before any work starts.