A team is dividing work on a new integration. Someone proposes that their server should decide which of its own capabilities the assistant ought to use for a given request, and return only that one. What is wrong with putting that responsibility on the server?

The principle — Every participant in MCP has one job, and most design mistakes are a job crossing a boundary.

Why the key is correct — MCP servers are programs that expose specific capabilities to AI applications through standardized protocol interfaces. Exposing is the verb. The primitives on either side define what clients and servers can offer each other, and offering is not the same as choosing: the server publishes what it can do and the host, with its model, decides what to invoke. A server that pre-selects has quietly taken a decision the client was going to make with far more context about the user's actual request.

Why the others are wrong — Selection is not part of the server's defined role, however well it knows its own domain. Whether it could run a model of its own is beside the point. And the interface belongs to the host, but that is a different boundary from this one.

Remember this — Servers offer; hosts and models choose.

Sources — MCP first-party documentation.