.Net: Python: MCP Client support for Agents

[rohit-ganguly]

---

Feature Request: MCP Client support with Python Agents

Hi all,

I've been going deep into building custom agents in Python for a while, and noticed a huge feature gap that I think SK could benefit from filling. Anthrophic's modelcontextprotocol/python-sdk doesn't currently provide a good client implementation that acts as a 'translation layer' from MCP tools to the way model providers expect tools to be passed in. It actually just forces the docstrings into system prompt which is not great.

This causes a huge headache making boilerplate to make my MCP tools understandable by Claude, OpenAI, etc.

The way this is done by other community-authored MCP clients/agents like Pydantic AI is by smartly reading docstrings in Google, Sphinx, or NumPy format. I think it would be super valuable to tackle this - with the rise of MCP, people know how to build servers - the 'agent' (client) is the new area that people need guidance on, which is where SK can help a lot here.

C# doesn't have this problem as there's interfaces via Microsoft.Extensions.AI, but Python really feels this problem. I think there's a lot of value in solving this!

[nmoeller]

I am currently working on something like this in Python #10778. I would be very happy to hear your feedback about this, the idea is pretty simple converting tools into KernelFunctions, internally SemanticKernel will send the Functions than as tools parameter to the corresponding API's.

[eavanvalkenburg]

This is done and released in 1.28.1