# MCP is a power strip (ep. 8)

Posted 2026-10-27 · Video: https://www.youtube.com/watch?v=oNo0hmZ5adA

> Everyone's talking about MCP right now. Basically, it's a standard way to plug tools into your AI.

**The prior it challenges:** More tools make an AI agent more capable

## Transcript

Rethink your priors.

Everyone's talking about MCP right now. Basically, it's a standard way to plug tools into your AI.

I think of it kind of like a power strip. Your calendar and your database plug in the same way, so nobody has to build a custom cord for each one.

I've built a bunch of these, and getting it connected is the easy part. What takes me way longer is deciding what it's allowed to touch. Reading my calendar, fine. Deleting something in my database, I want it to ask me first.

And like a real power strip, you can plug too much into it. Every tool adds a description the model has to read before it gets to your question.

Cloudflare ran into this. Their API has over 2,500 endpoints, and an MCP server for all of that, done the usual way, would eat about 1.17 million tokens. That's more than the whole context window of the most advanced models.

So they built Code Mode, and I think that's the architecture to follow once you have a lot of tools. The model writes a little code against a typed API made from the tools, the code runs in a sandbox, and only the result comes back.

Cloudflare's own docs say Code Mode doesn't handle permissions for you. You still build those into the tools.

So the question isn't, should we use MCP. It's, what should the robot be allowed to plug into.

Anyway. Start from the outcome. Or don't. I'm not your manager. Follow for more.

---
Rethink your priors by Eduardo Martínez. AI implementation. Outcomes over tools. And everything else.
YouTube: https://www.youtube.com/@rethinkyourpriors · Instagram: https://www.instagram.com/rethinkyourpriors/ · Work with Eduardo: https://edumartinez.ai/