Connecting to MCP

[ymuichiro]

OpenAI’s Function Calling is a highly flexible and excellent mechanism. When Anthropic announced MCP, I initially considered it similar to Function Calling and didn’t engage with it much. But how about now? The ecosystem has grown, and many MCP servers have been published.

https://glama.ai/mcp/servers

While it’s possible to develop the same functionality as an MCP using Function Calling, such a rich ecosystem is attractive. You can extend various services like Google Drive, Box, GitHub, Local File System, etc., in a loosely coupled manner. Does Semantic Kernel plan to add this MCP as a connection destination data source? While it may not be technically necessary, I feel it would be highly convenient given the expansion of the ecosystem.

[sophialagerkranspandey]

Hi @ymuichiro thanks for the feedback. MCP is on our radar for sure and is in our backlog of items to work on.

[zetomatoz]

That sounds fantastic! Do you have a rough timeline for the MCP support implementation?

[sophialagerkranspandey]

At this time, we don't have an estimated timeline.

[erichexter]

cant this be implemented as a plugin?

[balaji830]

@erichexter I was thinking the same. How is different from the plugin? Anyone can please provide some insight?

[erichexter]

so mcp is different from a plugin, it allows local and remote tools/functions, that work with the standard, so you get tools that work universally. I was suggesting that mcp client could be implemented in the sem kern plugin.. rather then having to be fully integrated.

[stephentoub]

There are MCP client libraries on nuget, including support for importing as Microsoft.Extensions.AI AIFunctions, e.g.
https://www.nuget.org/packages/McpDotNet.Extensions.AI
That should make it easy to use with SK, as well.

[nmoeller]

There was one PR finished that demonstrated how to use MCP Server Tools with Semantic kernel here f8ee3ac

Currently I have a PR open that demonstrates this capability in Python.