Find out what your docs fail to answer.
AI assistants are reading your documentation on behalf of your users, and you cannot see any of it. Turn your docs site into an MCP server and every question, every retrieved passage, and every gap becomes visible.
Free tier: 3 servers, 50 requests a month. No credit card.
Your biggest reader leaves no analytics trail
A growing share of your documentation is consumed by an AI assistant on a developer's behalf. That reader never loads your page, never fires your analytics, and never files a ticket when a page fails to answer the question. The failure surfaces later as a support request that never mentions the docs.
What that looks like
- Page views no longer describe how docs are actually used
- No signal on which questions your content fails to answer
- No way to prove a rewrite improved anything
- Support tickets that trace back to a gap you never saw
How it works
Three steps, start to finish
- 1
Add your documentation as a source
Paste your docs site URL into the wizard. If the site is large, point at a section or connect the docs GitHub repo for fuller coverage. Add your changelog and API reference as separate sources so you can measure them independently.
- 2
Publish the endpoint to your developers
Deploy and share the generated config with your users so they can add your docs to their own AI tools. Every question they ask now runs through content you control rather than a stale training snapshot.
- 3
Read what the dashboard tells you
Every request records the question, the passage that answered it, a retrieval confidence score, and whether it was answerable at all. Low-confidence clusters are your content gaps, ranked by how often they cost someone an answer.
What to index
- Your published documentation site
- The docs repository, for content that never made it to the site
- Your API reference, kept separate so it can be measured on its own
- Release notes and changelog
- Support macros and the knowledge base your team already maintains
- PDFs — spec sheets, integration guides, compliance documents
Tools worth enabling
- search_docs
- The workhorse. Its query log is the raw material for every content decision that follows.
- ask_question
- Captures questions in the words your users actually use, which is rarely the wording of your headings.
- find_api_reference
- Separates reference lookups from conceptual questions, so you can tell a navigation problem from a coverage problem.
- get_changelog
- Shows how often people are asking what changed, which tells you whether your release notes are doing their job.
Questions
- How do I turn my documentation site into an MCP server?
- Paste the documentation URL into the MCP Studio wizard as a source, choose which tools to expose, and deploy. The site is crawled and indexed in the background, up to 5,000 pages per source, and you get an MCP endpoint you can share with your developers.
- What can I actually measure once the docs are an MCP server?
- Every request records the question asked, which page and section answered it, an excerpt of the passage used, a retrieval confidence score, and whether the question was answerable. Aggregated, that produces ranked content gaps, per-page confidence, and a view of which sources carry the most load.
- Do I need engineering help to set this up?
- No. Creating and deploying the server is a no-code wizard. The only step that involves a developer is optionally connecting a private docs repository, which needs a GitHub sign-in.
- Does upgrading the analytics tier lose my earlier history?
- No. Per-request telemetry is written on every request regardless of which tier you are on. The tier controls what the dashboard displays, so upgrading backfills the new views with the history you already accumulated.
- How is this different from search analytics on my docs site?
- Site search only sees people who visit your site and use the search box. An MCP server sees the questions asked through AI assistants, which increasingly never load the page at all, and it records which passage was actually used to answer rather than just which result was clicked.