How Do llms.txt and MCP Servers Make Technical Documentation Machine-Readable?

How do the options compare?

What does llms.txt do for documentation?

An llms.txt file is a Markdown index: a title, a one-paragraph summary, then lists of links with short notes, grouped by section. For documentation, it works like a table of contents for machines. Two companions make it more useful:

• Markdown versions of each page, published at the same address with .md added. Tools can read the page without navigation, cookie banners or scripts.
• llms-full.txt, which some platforms generate by combining the entire documentation into one file, so a tool can load everything at once.

Who actually reads these files? Ahrefs' May 2026 study of 137,210 domains found that most valid llms.txt files received no requests at all. The fetches that did happen came mainly from GPTBot, with Claude Code second. Google says it ignores llms.txt. Its Search Advocate John Mueller said it was “not done for search” and called it a “temporary crutch” for coding tools, “perhaps to save some tokens”. The evidence on visibility is covered in whether llms.txt improves AI search visibility, and writing one in how to write an llms.txt file.

What does an MCP server add?

The Model Context Protocol is an open standard for connecting AI applications to outside tools and data. A documentation MCP server turns your docs into tools the assistant can call while it answers. Instead of relying on training data, which may describe last year's API, it searches your current docs and reads the relevant page. Whether sites beyond documentation need one is covered in whether websites need MCP to be ready for AI agents.

Four examples show the range:

• Mintlify generates a search MCP server at the /mcp path of each docs site. It exposes three tools: search, a file-system style tool for reading pages and sections, and a feedback tool that lets an agent report a page it found wrong or outdated. That report appears in the docs team's analytics. Public content is open to anyone; for gated docs, users sign in and can only search what they are allowed to see.
• GitBook adds an MCP server to every published site, at the site URL plus /~gitbook/mcp, with a separate endpoint for partly gated sites. It stops working if page actions are turned off.
• Microsoft Learn runs a free MCP server, with no sign-in needed, that searches its documentation, fetches full articles and searches code samples.
• Google launched its Developer Knowledge API and MCP server in public preview in February 2026. It serves Firebase, Android and Google Cloud documentation as Markdown, re-indexed within 24 hours of an update during the preview.

For open-source libraries, third-party indexes such as Context7 collect documentation for coding assistants, and maintainers can submit their library.

How does Mintlify expose docs to in-browser agents?

Mintlify also supports WebMCP on its Pro and Enterprise plans. WebMCP registers tools directly on the page through a browser interface, so an agent working in the browser can use them while someone reads the docs, without connecting a server first. Mintlify's five in-page tools include one that returns the site's llms.txt, so the two approaches meet in one place. Browser support for WebMCP is still emerging, so treat it as an extra, not a replacement.

Mintlify also publishes discovery files under /.well-known/, describing the server and its tools. These help agents find the server automatically, although most people still add servers by hand.

What should a documentation team do first?

1. Make the HTML pages good first. Content rendered on the server, one topic per page, current code samples. Everything else builds on this.
2. Publish Markdown versions and llms.txt. Often a setting on your docs platform; otherwise a build step, as covered in how to add llms.txt on any CMS.
3. Turn on the MCP server if your platform offers one.
4. Tell developers how to connect it. Put a short “use these docs in your AI tool” section in your quickstart, with the server URL and the one-line command for common tools.
5. Submit open-source libraries to indexes coding assistants use.
6. Watch the feedback. Agent-reported errors and search queries show where docs are missing or wrong.

What do they not do?

• They do not make you rank in AI search. ChatGPT search, Google's AI features and Perplexity build answers from their own indexes. An MCP server only helps people who have connected it.
• They do not fix wrong documentation. They deliver outdated pages faster. How stale changelogs and scattered pages mislead assistants is covered in how fragmented documentation causes wrong AI answers.
• They can leak what you publish. A single combined file makes all public docs easy to copy. Keep internal material out, and check that gated docs stay behind sign-in.

How do you build your own if your platform has no built-in option?

For llms.txt, generate the index and Markdown copies from your navigation at build time; static site generators such as Docusaurus and VitePress have community plugins for this. Serving Markdown to tools that ask for it is covered in how to serve Markdown to AI user agents.

For an MCP server, keep the first version small and read-only:

• A search tool that calls your existing docs search and returns titles, snippets and URLs.
• A read-page tool that returns one page as Markdown.
• Rate limits. Mintlify, for example, allows 5,000 requests an hour per user and 10,000 searches an hour per site.
• No write access, and no access to anything that is not already public, unless users sign in.

We have not run a documentation MCP server ourselves. This page is based on Mintlify's, GitBook's, Microsoft's and Google's documentation, and Ahrefs' llms.txt study.

What might this look like in practice?

Consider an illustrative payments API company, Tessellate, whose support team kept seeing developers use an endpoint retired a year earlier. Coding assistants were suggesting it from old training data. Tessellate turned on llms.txt and the MCP server in its docs platform, and added a “connect your AI tool” box to its quickstart with the server URL. Within two months, the feedback tool had flagged eleven pages agents found confusing, including two that still mentioned the old endpoint. The team fixed them, and support tickets about the retired endpoint fell steadily. Its visibility in AI search answers did not change, as it had expected.

What do people ask about llms.txt and MCP for documentation?

Does llms.txt help my docs appear in ChatGPT answers?

There is no evidence that it does. It mainly helps coding tools and agents read your docs efficiently.

Do I need both llms.txt and an MCP server?

They complement each other. llms.txt is cheap and passive; an MCP server gives connected assistants live search. Start with llms.txt if you can only do one.

Is a documentation MCP server safe for private docs?

It can be, if it enforces sign-in and shows each user only what they may see. Hosted platforms handle this; check permissions carefully if you build your own.

How do developers connect to a docs MCP server?

They add the server URL in their tool's MCP settings, or with a one-line command in tools such as Claude Code.