I am impressed with the support for the new MCP Servers—great work! I’d like to suggest a couple of improvements that would personally make my experience even better:
Implement support for JSON configuration, similar to Claude’s format.
Introduce a ‘YOLO’ mode for tools, allowing users to whitelist tools that can be run without confirmation. For instance, I use a Pinecone database as shared memory across my projects and would love for the cursor agent to automatically check it as it does with the project codebase.
Add an ‘@’ mention feature for tools to easily instruct the agent to use a specific tool.
10000% needed
Configuration shouldn’t be this tedious
Should be configurable the same for all MCP clients to streamline it allowing plug-n-play use
Whitelist is mandatory too
MCPs are essential for real autonomous agents, can’t afford staying around the composer to avoid a potential timeout because you didn’t accept the tool use request. Would also need an increase in tools usage per prompt since it reaches 25 use quite fast (or make a separate count)
Really appreciate all the feedback on MCPs! We wanted to get this out early so people could start using it, and your suggestions are super helpful for our next iterations. The YOLO mode for tools is definitely something we’re looking at - totally get how manual confirmations can slow down autonomous agents. JSON config support and @ mentions for tools are solid ideas too
For now you can work around the 25 tool limit by just asking the agent to continue, but we’re keeping an eye on how this affects different workflows
Been working with the Cursor MCP support a bit and it’s great! Something I would really appreciate is better control over MCPs for extensions.
We’re building an extension that automatically spins up an SSE MCP server and the only tool we have to communicate with Cursor right now is mcp.json. If there was a way to register/deregister/refresh/list servers from an extension, that would make the experience so much smoother!