🧠 Light Day, Big Realization: What MCPs Actually Are

Today wasn’t heavy on the builds, but I did have a quiet epiphany about what I’ve actually been making: MCPs (Model Context Protocols).

I still don’t love the term. It feels a little sci-fi, a little forced. The words themselves don't evoke any meaning in my head yet. The core idea is: this is code waiting to be instantiated by an LLM. Without MCPs, LLMs just talk. With MCPs, they can act. They can trigger any behavior that code allows. That's an enormous surface area. In today’s world, what isn’t accessible via a code call?

Right now, I’m keeping it simple: write code to a directory, search a repo to gather context, that kind of thing. But the abstraction scales. It's basically an execution substrate for LLMs. The more I think about it, the more foundational it feels.

I also noticed something else: how reliant I’ve become on documentation. I used to never write it. I didn’t need to. The code was mine, the logic was in my head, and I could just reread the source if I stepped away for a bit.

But now? The LLM is doing most of the writing. I don’t touch the code directly. That puts distance between me and the implementation. And honestly, I barely know what’s inside some of these features unless the LLM leaves breadcrumbs. I’ve caught myself forgetting how to use something just 30 minutes after testing it. That never used to happen.

It’s a weird reversal: more code, more speed, more power — but less felt ownership. Documentation isn’t just a nicety anymore.

What’s next: Might explore ways to auto-summarize features into usage notes or REPL-like docs. Something to re-anchor memory in a system I didn’t write by hand.