Skip to main content

MCP Integration

If your organization has a Model Context Protocol (MCP) server configured, chat completions automatically gain tool use against it — no extra request parameters needed.

How it works​

  • Tools always come from your org's configured MCP server. You cannot supply your own tool definitions in a request; there's no tools/tool_choice field on the completions API.
  • When the model decides to call a tool, the call is executed server-side against your MCP server, the result is fed back into the conversation, and the loop continues automatically (up to a small fixed number of round-trips per request).
  • You only ever see a final answer. The response never contains a raw, unexecuted tool_calls array — by the time you get a response, any tool use has already happened and been resolved into the assistant's reply.
  • If the MCP server is unreachable mid-request, the API returns 503 — see Errors.

Enabling it​

MCP integration is not self-serve. It's provisioned per organization by r-mad.ai, since it involves registering your MCP server's URL, credentials, and an explicit allow-list of which tools the model may call.

To set one up, contact us at support@r-mad.ai with:

  • Your MCP server's URL (must speak Streamable HTTP)
  • The tool names you want available to the model
  • Any authentication your server requires

We'll configure it on our side; once enabled, every chat completion for your org automatically has access — no client-side changes required.

Orgs without MCP configured​

Requests are answered as a plain single-shot completion with no tools offered at all — behavior is otherwise identical to an org with MCP enabled.