Client Sampling

When the server requests an LLM completion, the client should show the request to the user (human-in-the-loop) and then call the LLM.

async fn handle_create_message_request(
    &self, params: CreateMessageRequestParams, _runtime: &dyn McpClient,
) -> Result<CreateMessageResult, RpcError> {
    // params contains the conversation so far, system prompt, etc.
    // The client selects a model, calls the LLM, and returns the result

    Ok(CreateMessageResult {
        role: Role::Assistant,
        content: CreateMessageContent::TextContent(TextContent::new(
            "Here is the completion from the LLM...".into(),
            None,
            None,
        )),
        model: "gpt-4".into(),
        stop_reason: Some("endTurn".into()),
        meta: None,
    })
}

The client must declare sampling in its capabilities for this to work:

let client_details = InitializeRequestParams {
    capabilities: ClientCapabilities {
        sampling: Some(ClientSampling::default()),
        ..Default::default()
    },
    // client_info: Implementation { name: "my-client".into(), version: "0.1.0".into(), .. },
    // protocol_version: "2025-11-25".into(),
    ..Default::default()
};