Manual · Chapter 14 · Part C Channels

Agent-ready website

Purpose

More and more of the questions put to your website come not from people but from AI assistants: ChatGPT, Claude, Perplexity, the assistants in browsers and operating systems. They read pages, guess and summarise. With the answer endpoint, you give them your own assistant instead: the agent asks, dAI Pro answers from your content, with sources. What your website says about you then reaches AI assistants the way you meant it.

Where

The switch Answer endpoint for AI agents (MCP) is in the collection's settings under "Access and visitor functions". It is off by default.

The switch "Answer endpoint for AI agents" in the settings

What applies

  • Questions run through the same answer pipeline as the chat: knowledge rule, sources, answer cache. A question the chat has already answered does not cost a second time.
  • The collection's allowed websites apply here too.
  • Each sender is allowed 60 questions per hour. Anything more is rejected.
  • Every question is recorded in the log with the marker "mcp", and can be analysed separately from the chat in the quality cockpit (Chapter 16).
  • Human handover, actions and tools are not available to the endpoint. It only answers.

Making it known

For agents to find the endpoint, it must be named on the website. Exactly how depends on the service and is currently changing fast; the administration will find the current recommendations in the customer area (Chapter 29). Two routes are common today: an entry in the file in which the website lists its services for agents, and a link with the appropriate relation in the head of every page.

Frequently asked questions

Does it cost anything? Every new question is a call to the language model as in the chat. Repeated questions come from the cache. The per-sender throttle limits the effort.

Can I see which agents are asking? The log records question, answer and marker, not the service behind it. Agents rarely identify themselves.

Do the answers reveal more than the website? No. The endpoint answers from the same content as the chat and only from active documents.

See also