@mcp -> @github, @tavily ?

[ravitemer]

Hi guys, I need your opinion with a design decision with regard to MCP Hub and CodeCompanion

Currently we add MCP abilities to CodeCompanion using @mcp tool group which has use_mcp_tool and access_mcp_resource tools. We then add the running MCP Servers with their tools, resources to the system prompt.

With the current approach, there are situations where some models get confused between the use_mcp_tool and actual tool of an MCP server and responds with tool call where the name of the tool called is not use_mcp_tool but e.g get_issues a tool from github MCP server. Although this is model specific and not that common, I am wondering how many of you are encountering this issue. So I am thinking of providing each MCP server as well as each tool in a MCP Server as an individual tools just like how we currently have the resources as variables and prompts as slash_commands.

Are there any one who wants each server or each tool of a MCP Server available as individual tool like for example if we have github, tavily MCP servers running in the hub

[ljwall]

With the second option "Provide each server and each tool in the server as tools like @github, @github:get_issues, @tavily etc " how would you see it working to start / stop local MCP servers automatically?

[ravitemer]

We can't. This is for users who wants more control or save some tokens with better reliability.

If we want to let the LLM handle toggling servers automatically or just call correct tools first option is ideal.

However, we can have both @mcp tool for these scenarios along with individual servers or tools

[ljwall]

thanks, that makes sense. I think I'd still lean towards individual tools if possible

[Keyruu]

Would it be possible to have both?

[ravitemer]

Yes.

[Keyruu]

Then I would prefer that option!