MCP vs. API: The Rest of the Homework

I've gotten the full writeup online.

You've already seen the scoreboard from August (you do read these, right?). I'm tying up the work. The overall MCP edge is ten-ish percent, not worth the maintenance. That part has not changed. My initial guess was mostly correct.

What I did not elaborate on earlier: the extra MCP file is not the cost to base this decision around. It's whether you should invest into MCP or a better web API. If you can get both, by all means. Keeping two contracts honest with each other is overhead, and that overhead grows with the polish you'd expect - and in many cases does not end. I've kept the mistakes, the ugly 66/70 is still in the record, the research wasn't one-shot. Both fixes moved toward the tie I'd already published - the fixture tighten helped the OpenAPI tests, and two of the three first-pass misses were MCP. I've also added a short list of when I think the second interface is actually worth paying for. Hint: "Agents need to use us" is not on it.

Homework, receipts, the boring methodology: https://tinthe.dev/case-study-openapi-vs-mcp

In other news: I've been wiring a small software factory for cloud agents. A parent that is not allowed to actively work (the loop), a builder that gets wiped per-ticket, and a reviewer on a different model family on purpose. The mission is the builder's whole memory. I'll run it overnight on a product I already wanted to build and I'll tell you what it invents.

Expect that post soon.