Connecting an AI client over MCP

Noverva exposes an MCP endpoint at /api/mcp, so an AI client can work with your sites directly rather than you copying figures into a chat window.

It inherits your authorisation exactly

There is no service account and no ambient authority. An MCP session acts as you, with your permissions, in your workspace. Anything you could not do in the interface fails over MCP with the same error.

If you are restricted to certain sites, so is the assistant. If your plan does not include API access, MCP is refused — it is the same entitlement, and treating it as a separate door would make it a way around the first one.

Consequential actions ask first

Reading is free. Anything that spends money or changes state — requesting a crawl, for instance — requires explicit confirmation rather than happening because a model decided it was a good idea.

The prompts carry the rules

The server publishes prompts alongside its tools, and those prompts contain the reporting rules from this documentation. An assistant reading your opportunity scores is told, in the same breath, that they contain no demand data and must not be presented as search volume.

This matters more over MCP than anywhere else. A model summarising your data will smooth over caveats unless the caveats travel with the data, and "your top opportunity is X" is exactly the kind of confident sentence that loses the qualification that made it honest.

Rate limits

MCP is limited per user and at a lower ceiling than the REST API, because an agent in a retry loop is the expected failure mode here rather than a busy integration. Limits scale with your plan.

Current limitation

MCP authenticates by browser session, which means it works with clients that can carry your login. Unattended agents authenticating by API key are not supported yet.