How to make your developer docs agent-ready
Developer docs are agent-ready when a coding agent like Claude Code, Cursor or Codex can find the right page, load it in one request as clean markdown, and get from it to a working API call. In practice: an llms.txt index, every page available as .md or through Accept: text/markdown, pages short enough to fetch whole, a stable OpenAPI spec, API keys a developer can create without a sales call, and ideally a docs MCP server.
From our data: As of 3 Nov 2026, our scan found developer docs for 178 of the 214 software vendors in our rankings. 101 served docs as markdown, 174 published llms.txt (28 of those on the docs host), 60 had llms-full.txt and 35 published an OpenAPI spec at a URL we could find.
Steps
Publish a lean llms.txt
An H1, a one-paragraph summary, and H2 sections linking your key pages with a short note each. Link the .md version of each page. Vercel's guide recommends keeping it a described index, not a dump of every URL.
Serve every page as markdown
Either page.md next to every page, or the same URL answering Accept: text/markdown with Content-Type: text/markdown and Vary: Accept. Advertise it with <link rel="alternate" type="text/markdown">.
Add llms-full.txt
Your whole docs in one markdown file, for agents that want everything in one load. Keep it current from the same build.
Keep pages fetchable in one go
The Agent-Friendly Documentation Spec (AFDocs) checks page size, client-side rendering and tabbed content that bloats a page. Split giant reference pages and server-render the rest.
Return real status codes
404 for missing pages (not a 200 "not found" page), redirects for moved ones, and 429 with Retry-After when you rate-limit. Agents treat a 200 as content.
Publish an OpenAPI spec
At a stable URL like /openapi.json, linked from llms.txt. Agents generate correct calls from it.
Put key creation on the first page
The quickstart should say where keys come from, with no sales step. Offer test keys or free credits so an agent can verify an integration.
Run a docs MCP server
A search-and-fetch MCP server over your docs lets agents query them directly. GitBook and Mintlify host one for every published site.
Don't gate public docs
Login walls and bot challenges on docs pages are AFDocs' authentication failures. Keep reference docs public.
Platforms that do most of it for you
GitBook serves .md for every page, /llms.txt, /llms-full.txt and a read-only MCP server at /~gitbook/mcp on published sites. Mintlify hosts llms.txt and a docs MCP server, and generates a skill.md that describes what agents can do with your product, also listed under /.well-known/agent-skills/. If you build docs yourself, Vercel's guide and AFDocs are the two best checklists.
Vendors whose docs serve markdown to agents
101 of the vendors in our rankings, from our scan on 3 Nov 2026.
Check it yourself
You want markdown from the first, content-type: text/markdown from the second and a 404 from the third. The last runs the open AFDocs checks against your docs. Then run the free check with the Get an API key job.
curl -s https://docs.example.com/llms.txt | head -15
curl -s -D - -o /dev/null -H "Accept: text/markdown" https://docs.example.com/quickstart | grep -i "^content-type"
curl -s -o /dev/null -w "%{http_code}\n" https://docs.example.com/this-page-does-not-exist
npx afdocs check https://docs.example.comQuestions
llms.txt or llms-full.txt for docs?
Both. llms.txt is the index an agent reads first; llms-full.txt is the whole thing for agents with room to load it. Generate both from the same build.
Do I need a docs MCP server if I have llms.txt?
No, but it helps agents that search rather than browse. A docs MCP server answers "how do I paginate?" without the agent loading pages it doesn't need. It's separate from a product MCP server, which performs actions in your product.
What is the Agent-Friendly Documentation Spec?
An open spec created by Dachary Carey that defines testable checks for how well docs work for agents, across discoverability, markdown, page size, structure, URL stability, observability and authentication. Its CLI is npx afdocs check, and Fern's Agent Score is built on it.
Does serving markdown hurt SEO?
Not when done properly: same URL with Vary: Accept, or a separate .md URL that isn't in your sitemap. Google says Search doesn't use markdown copies, so they neither help nor hurt rankings.