🚀 feat: Add a sampling handler

[codingjoe]

Hi there 👋,

wonderful project! It serves as a great dev environment to build MCP servers.

That being said, I noticed one of the newer features missing, mainly sampling.

Sampling allows an MCP server to run inference on the client. Kind of "bring your own LLM" in an agentic world.

FastMCP has a fairly straight forward client integration: https://gofastmcp.com/clients/sampling

The handler could include a HIL, even though those tokens will not be injected in the client chat.
It might "cost you inference/money" but there is no token injection risk.

If you are curious about this feature, I'd be happy to provide a patch.

Best!
Joe

PS: Do you accept donations?

[codingjoe]

OMG ticket no 42!!!!!

[jonigl]

Hi Joe,

Thank you so much for the kind words and for opening the PR, I really appreciate it! It’s on my list and I’ll check it out soon, along with your question about donations. Sorry I couldn’t get to it earlier.

Thanks again for your support and patience!

Best,
Jonathan

[codingjoe]

Hi @jonigl,

Thanks! I started working on a draft. Sampling is already working. However, MCP defines a include_context argument, which allows servers to select MCP servers to be included in the sampling inference run on the client. This bit I haven't implemented yet, as it wasn't even implemented in FastMCP yet jlowin/fastmcp#1141

With that I'll try to get on it later next week.
Best
Joe

[jonigl]

Nice work, @codingjoe!

The sampling feature is interesting. If I understood correctly, it lets MCP Servers request prompt execution on the client with specific parameters and context. Are you building a server using this? If so, I’d like to try it out.

HIL controls are key for security, since a remote server could access personal data. It makes sense for this to be off by default with clear warnings if enabled.

I also saw your PR: jlowin/fastmcp#1141. Looks like it's lined up for the next release.

Thanks for contributing! 🙂

[codingjoe]

Yes, I am working on an open agent orchestration SDK called codingjoe/karajan. I'll push my code on the weekend.

[codingjoe]

While writing it, I noticed that there's still a lot of ground that needs to be covered, from client, to SDK, to spec. This is going to be much more tricky fun with MCP chaning modelcontextprotocol/modelcontextprotocol#958