Agents learning mxcp

[alexzerntev]

Hello,
I was thinking about how we could teach mxcp to Cursor (or any other agent). I created a branch: https://github.com/raw-labs/mxcp/tree/agent-help . It basically adds a discovery tool through a command mxcp agent-help which has its own internal navigation through better organized documentation. At some point when the models are trained on MXCP they will already know how it works, but until then...

The idea is to prompt Cursor like this(very short version): There is a cli tool called mxcp, you can read what it does through a command mxcp agent-help . Create an MCP server with with mxcp that does .

What are your thoughts?

[miguelbranco80]

Great idea, let's try this out!

My understanding is that Cursor has a biased towards executing tools. For instance, it will not, by default, lookup docs on the web on how a new tool works necessarily; it will prefer to run it and see the --help. So if the tool itself explains how it should be used - and it could do so, when running your mxcp agent-help - then this could work really well.

Give it a try and let's see!

[alexzerntev]

Ok, I tried it out, here is the report: https://medium.com/p/6e1a85d23f04