When to use llms.txt, ARD, and MCP together
Use llms.txt to guide agents to your docs. Add Agentic Resource Discovery (ARD) when agents or directories that support it need to find your integration. Use Model Context Protocol (MCP) to connect a compatible AI app to your tools or data.
Use all three when agents need to read your docs, find your integration, and access it through the app. Publishing these formats does not establish agent readiness.
What each one is for
Reading. Publish llms.txt at your site root, or /docs/llms.txt for pages under /docs/. Link to Markdown versions of those pages when available. It sits beside robots.txt and sitemaps. It does not replace them. It does not declare a tool or integration by itself.
Discovery. Publish ARD to help agents or searchable directories find a tool or integration you run. The agent or directory must support ARD. Advertise with /.well-known/ard.json and <link rel="ard">.
Connection. Implement MCP when the AI app you want to support uses it. Run an MCP server that gives the app access to your tools or data. Current protocol revision: 2026-07-28. MCP can also list tools on a connected server. That in-protocol listing is not the same as site-wide ARD discovery.
Worked example (fictional, not tested)
Northwind Docs Cloud wants agents to read the docs, find a "create project" integration, and call it from an AI app.
- Publish
/docs/llms.txtwith links to markdown API and auth pages. Open the file. Follow the links. Check that the file and linked markdown pages load. - Publish
/.well-known/ard.jsonwith anentriesarray. Add<link rel="ard" href="/.well-known/ard.json">to the page's<head>.
Each entry needs identifier, displayName, type, and exactly one of url or data. Include representativeQueries with examples of user requests the integration can handle.
For Northwind's MCP entry, use a server card: a JSON document describing the MCP server. Link to that document with url, or include it directly in data.
Validate the manifest and check that referenced documents load. See Agentic Resource Discovery Specification, v0.91 Proposal, August 26, 2026.
- Run an MCP server with a create-project tool. Test that tool with your own client in staging.
HonestClaw measures observational agent readiness from public evidence. A free scan publishes a public record. For Use, HonestClaw checks whether interfaces are published and whether software can read their structure. The score excludes live API and MCP call success, completed jobs, and whether your site appears or is cited in AI answers. We do not sell scores, reviews, or placement.
When one layer is enough
Docs only, no tool or integration to advertise: start with llms.txt.
For a tool already connected to one AI app, MCP or that app's connector may be enough without public ARD discovery.
Sources
- The /llms.txt file, v2, Jeremy Howard, 2024; updated 2026.
- Agentic Resource Discovery Specification, v0.91 Proposal, August 26, 2026.
- Model Context Protocol Specification, revision 2026-07-28.
Next action
Pick one domain you own. Add only what the agents you want to support need. Run a free scan to record the public evidence.