Sampling

What it does

Sampling allows an MCP server to request LLM completions through the client, enabling agentic behaviors where the server can leverage AI reasoning. The server sends sampling/createMessage to the client with a prompt, and the client returns the LLM's response. This creates a human-in-the-loop pattern where the client controls which model is used and can approve requests before sending them.

How it works

Example implementation

const result = await server.requestSampling({
  messages: [{ role: "user", content: { type: "text", text: "Summarize this data" } }],
  modelPreferences: { hints: [{ name: "claude-3-5-sonnet" }] },
  maxTokens: 500
});