I don’t really want to use global MCP servers. They’re a waste of tokens. But with FastMCP, I can made a simple MCP server with just the tools I need, one per every project.
I want to be able to specify a config in the .cursor folder of my project which defines that project-specific MCP server (or multiple, but even just one would be fine) and then the Cursor LLMs would pick it up and use.
And I’d like to be able to decide which MCP server is executed YOLO-style, without confirmation.
My view is that an MCP server should expose extremely simple commands, similar to the commands that a tool like make or hatch does: build, test, fmt, clean etc. No more than maybe 6 commands. The LLM should leverage its trained knowledge of working with build tools. The LLM already has prior knowledge of using things like “make clean” or “hatch build” or watever. So it doesn’t need much explantion if the tool matches that. I’ll just call the stuff.
This means that with MCP, I can instruct the LLM in agentic mode to perform some higher-level operations, and I myself would provide the implementation. The implementation could really by “anything” of course.
Thanks for the detailed feature request! Currently project-specific MCP configs in the .cursor folder aren’t supported, but this is a really interesting idea for making MCP more efficient and project-focused
You’re spot on about keeping MCP commands simple and build-tool-like - that’s a great approach. The current docs (Cursor – Model Context Protocol) only cover global MCP setup, but project-level config would be super useful for cases like yours
I’ll pass this feedback along to the team - having project-specific servers is definitely on our roadmap for MCP in the near future!
I made an account here specifically to second this request. I’m currently fighting an uphill battle trying to get memory to be project specific. PLEASE make this feature. Big thanks, and love the great work you guys are doing.
Yes for certain MCP tools this will be a must (project based) as some projects are not to be allowed to access that specific tool, and others will need it.
It’d be nice to be able to load env variables somehow in that file to be able to commit the file without sensitive credentials, probably needs refactoring to change the type of file where mcp servers project configurations are stored tho.