We need some more client and sampling examples and use-cases.

[z29]

The client examples provided are vague and the sampling examples, non-existant. The best I could find was:

  async def handle_sampling_message(
      message: types.CreateMessageRequestParams,
  ) -> types.CreateMessageResult:
      return types.CreateMessageResult(
          role="assistant",
          content=types.TextContent(
              type="text",
              text="Hello, world! from model",
          ),
          model="gpt-3.5-turbo",
          stopReason="endTurn",
      )

this barely explains how it works, and worse, doesn't show how to invoke sampling on the server-side. The documentation on the mcp website is even more vague.

Then there seem to be multiple ways to create a client:

• creating an MCPClient class like in https://modelcontextprotocol.io/quickstart/client
• aysnc with ClientSession like in your readme https://github.com/modelcontextprotocol/python-sdk/blob/main/README.md#writing-mcp-clients

I'm just trying to learn, but its confusing

[dacox]

I agree. Reading https://modelcontextprotocol.io/specification/2025-03-26/client/sampling, I was hoping the python sdk supported this. However, it seems like it doesn't? It's quite unclear!

Is the sampling callback supposed to explicitly call the LLM?

[Finndersen]

Also that example sampling callback function doesn't even match the actual required protocol:

class SamplingFnT(Protocol):
    async def __call__(
        self,
        context: RequestContext["ClientSession", Any],
        params: types.CreateMessageRequestParams,
    ) -> types.CreateMessageResult | types.ErrorData: ...

[felixweinberger]

How about https://github.com/modelcontextprotocol/python-sdk/blob/main/examples/snippets/servers/sampling.py?

[Gerr1ck]

@felixweinberger thanks for sharing this. But this is just the server implementation. What is the client doing? What do we have to implement there? Who is responsible to call "generate_poem" function at the server? I have also a lot of questions in terms of MCP sampling...

[felixweinberger]

@felixweinberger thanks for sharing this. But this is just the server implementation. What is the client doing? What do we have to implement there? Who is responsible to call "generate_poem" function at the server? I have also a lot of questions in terms of MCP sampling...

Ah sorry, missed the point about client here!

Looks like #1436 is attempting to create an example for just that - is that going in the right direction in answering your questions about how to set this up?