What is MCP sampling?

Sampling reverses the usual direction of an MCP interaction. Normally a client asks a server for something. With sampling, a server asks the client to get a completion from its language model, which lets the server run nested LLM calls inside its own features without holding any model API key. The server returns an InputRequiredResult containing a sampling/createMessage request with messages and optional settings. The client gets the completion from its model and retries the original call with the assistant message, model name and stop reason attached. Servers can also pass a tools array so the model can use specially defined tools during that single flow, but only if the client declared the sampling.tools capability.

Because servers cannot pick a model by name for certain, they describe preferences: priorities for cost, speed and intelligence, plus optional model hints. The client makes the final choice. The client may also change or ignore the system prompt, and it must respect the maximum token limit.

The specification says there should always be a human in the loop who can deny sampling requests, review prompts before they are sent, and check responses before delivery.

Status note: as of the 2026-07-28 revision (checked 2026-10-26), the page marks sampling as deprecated. It stays in the specification for at least twelve months, and new implementations should not adopt it. Existing ones should move to calling LLM provider APIs directly.