AI coding agents now read 50–66% of developer docs. How are you keeping your API docs updated?

Came across a wild stat recently

Mintlify reported that AI coding agents like Claude Code and Cursor made up 66% of traffic across their hosted docs in July, up from 15.2% in January. GitBook saw a similar jump (51.8% of page reads).

So almost every dev tool team I talk to is shipping `/llms.txt` files, .md page mirrors, or a docs MCP server.

The catch we kept hitting: an MCP server or llms.txt is only as accurate as the Markdown/MDX page behind it. Say a backend PR quietly drops a pagination limit from 100 to 50, or adds a webhook enum, and nobody touches the docs on merge. Your MCP server now feeds wrong instructions straight into your users’ Cursor and Claude Code sessions.

The usual fixes don’t hold up:

1. “Did you update the docs?” checkboxes get ticked out of habit.

2. OpenAPI generators update the reference table but leave the hand-written MDX guides, webhook pages and curl snippets stale.

3. When docs live in a separate repo, reviewers never even see the page a merge just broke.

That’s the problem I’m building DocFit for (docfit.dev, launching on Product Hunt Nov 3rd). It treats docs like CI: every merged PR gets checked against the docs, and the docs PR comes to you instead of waiting for a pre-release cleanup.

For the technical founders and devs here:

1. Do your docs live in the same repo as your backend, or a separate one?

2. How do you currently keep your API/developer docs updated?