Controlling Context in Client for Sampling Requests

[yarnabrina]

Question

I am exploring the sampling components in MCP. Based on my understanding of the protocol, this should allow the server to utilise the LLM(s) available on the client side. Unfortunately, none of the examples provide a working demonstration of this.

I have managed to get something working here, but I had to invoke the LLM directly using only part of the information from the CreateMessageRequestParams object. I have not been able to determine the purpose of RequestContext or how it can be leveraged.

In summary, these are my questions:

1. How can I access the LLM(s) available to the client directly, rather than making an explicit call?
2. How can I use modelPreferences to suggest that the client changes the choice of LLM based on priorities or name hints?
3. How can I use the RequestContext to pass available context to the LLM for improved generation?
4. How can I use the includeContext attribute of CreateMessageRequestParams to filter what information is included in the context?

Any guidance or examples would be greatly appreciated.

Additional Context

• MCP Python SDK version: 1.12.2
• Python version: 3.13.5
• Operating System: Ubuntu 22.04.5 LTS on Windows 11 Home via WSL 2

[yarnabrina]

@dsp-ant @ihrpr @Kludex may I kindly request you to have a look at this question when you have a moment? I'd appreciate your inputs.

[Kludex]

I am exploring the sampling components in MCP. Based on my understanding of the protocol, this should allow the server to utilise the LLM(s) available on the client side. Unfortunately, none of the examples provide a working demonstration of this.

Check https://ai.pydantic.dev/mcp/server/#mcp-sampling.

[yarnabrina]

Thanks for the response. Sorry if I was unclear, what I meant was none of the examples actually show how does the MCP client callback calls already configured LLM (or use hints to switch the model). I am not using pydantic-ai yet, but in that example also assumes a LLM response as a TODO comment only.

As linked, I do have a sample implementation, but I am not sure if this is the intended way to use as I failed to use some properties, some guidance will be much appreciated.

https://github.com/yarnabrina/learn-model-context-protocol/blob/770728148a8aa423a1467785bdb695927cf3428d/src/mcp_learning/mcp_client/orchestrator.py#L253-L334

[felixweinberger]

Should add a sampling example to our mkdocs

[yarnabrina]

@felixweinberger thanks for looking into it. Can you please have a look at above link and see if this is the ideal way to use sampling? If so, I'm happy to send a PR if you can guide me about the docs/examples contribution process here.

[maxisbey]

You could start with a draft PR by adding an example server for sampling, and an example client for sampling in the examples/ folder, each with a README.md stating how to run the tests. We're in the process of adding a proper documentation structure so you can also try to add some docs in the docs/ folder as well if you'd like.

Give it your best shot then it makes it easy for someone to give feedback afterwards if anything needs to be changed :)