Making your website readable to large language models is the first half of the story. The second half is giving AI agents a way to call your capabilities directly. Three lightweight mechanisms cover most legacy setups.
llms.txt: a text map for models
An llms.txt file in the site root tells a model what the website is about, what its core sections are, and where each entry point lives. It is plain text, costs almost nothing to maintain, and is the cheapest way to make an existing site navigable for AI.
OpenAPI: describe what agents can call
If you already expose endpoints — search, pricing, booking, order status — publish an OpenAPI document that describes them. Agents can then discover the operations, their parameters and the expected responses without a human-written integration.
MCP: the next step
For deeper integrations, a Model Context Protocol server lets agents invoke authorized operations through a standard channel. It is more work than llms.txt or OpenAPI, so start with the first two and treat MCP as the progressive enhancement.
Sequence it
Add llms.txt today, publish OpenAPI when you have endpoints worth exposing, and evaluate MCP when agents become a measurable channel. Test with a simulated agent question: can it find your capability and call it correctly? The same discipline we apply in our [case studies](/sites/).