[Feature]: Add support for sampling capability

[0xKoller]

Problem statement

MCP provides a standardized way for servers to request LLM sampling ("completions" or "generations") from language models via clients. This flow allows clients to maintain control over model access, selection, and permissions while enabling servers to leverage AI capabilities—with no server API keys necessary.

Currently, xmcp tool handlers have access to extra.sendRequest() which could technically be used to send a sampling/createMessage request, but there's no ergonomic helper for it. Developers would need to manually construct the full JSON-RPC payload, know the method name, and handle the response parsing themselves.

The xmcp client already declares the sampling capability during initialization (packages/xmcp/src/client/index.ts:28), but the server-side framework provides no abstraction for tool handlers to leverage it.

Proposed solution

The framework should:

• Declare sampling server capability (with optional tools sub-capability) during initialization
• Support all message content types: text, image, audio
• Support modelPreferences (hints, cost/speed/intelligence priorities)
• Support tools and toolChoice in sampling requests for agentic workflows
• Handle the multi-turn tool loop pattern (tool_use → tool_result → continue)
• Export the SampleResult type from xmcp

Package

xmcp (core framework)