MCP guide
An agent can be given a model context protocol server. The tools it exposes are treated as untrusted input.
What is stored
A list of server entries on the agent row. Each entry is a name and a URL; there is no separate credential store for a server, so a server that needs a token is out of scope today.
Guards
| Control | Behaviour |
|---|---|
| Scheme | Only https. Anything else is refused before a socket is opened. |
| Destination | Private, loopback, link-local and metadata ranges are refused, and the check runs again after every redirect. |
| Reply size | Capped, so a hostile server cannot exhaust memory. |
| Spend | A paid tool call is bounded by the agent's per-call cap and its allowlist. An unlisted host cannot be paid. |
| Steps | The run has a step cap and a budget cap. A loop stops rather than running until the credit is gone. |
The same destination policy applies to the generic fetch tool, so a URL that appears in a prompt cannot reach an internal service.