Today I had a blast - I felt like a kid again. I’ve been playing with version two of the Cheshire Cat, building a few little things that will eventually turn into plugins, tutorials, and shared learning moments.
Skills inside the Cheshire Cat in 20 minutes
The idea for today was simple: can I use the new version to write a plugin that
gives the Cheshire Cat skills? The kind you drop into whatever folder you
like, loaded according to your own logic. And it worked. In about 20 minutes,
Cheshire Cat 2 had skills up and running - it read the skill.md, executed
the scripts, and did what it was told, on command.
Clearly this is a bit risky. Skills are wonderfully convenient, but - as you well know - they mean the agent has direct access to the terminal. That’s the trade-off you’re signing up for.
There’s also a different flavor of skill I came across, for cases where you’d rather not use MCP at all. You give the agent the skill’s markdown file, let it read it from the skill folder, and that’s it - but the moment you want it to execute something, those actions have to be exposed as proper tools. That’s an interesting approach too.
MCP: postponed on purpose
I’m holding off on the MCP (Model Context Protocol) integration for now. I want to wait for the explicit release of the protocol version due on July 28, which I’m looking forward to with real anticipation.
After that, we’ll figure out how MCP servers should actually be configured inside the Cheshire Cat, and that raises some genuinely tricky design questions:
- Who adds the servers - the administrator?
- Who filters them, and can the user filter them too?
- Do they belong to the agent, or to the user?
None of these are easy calls. Getting the ownership and permission model right is half the work.
That’s it for today - the power actually went out on me mid-session. Take care, everyone.