What jumped out at me is not that they reversed themselves on MCP, but how plainly they admit the old objection wasn’t enough. “We don’t support MCP” is an easy position when a protocol looks awkward and half-baked. Much more interesting is the claim that the protocol changed, the surrounding harnesses changed, and—most importantly—the parts they had to build for MCP turned out to be useful anyway. That feels like the real story here: not ideological conversion, but a boring engineering convergence.
I’m mildly skeptical of the way “modern MCP” is framed as if the ecosystem has cleanly matured. Maybe it has in some places. But the article keeps hinting at the same underlying complaint: the protocol is only as good as the servers and harnesses around it. That sounds right to me, and it also means the protocol itself may not be the main bottleneck. If a tool system only works when everyone agrees to return structured data, write decent descriptions, and stop treating token dumping as a strategy, then the protocol is the easy part.
The more persuasive bit is Codemode. That sounds less like a shiny feature and more like a control plane for tool orchestration, sitting on the harness side instead of being stuffed into the model context. That distinction matters. A lot of agent tooling still feels like “throw more tools at the model and hope.” Here they’re trying to give the agent a place to compose calls, keep state in the transcript, and avoid wasting context. That is the kind of plumbing I trust more than protocol evangelism.
I also like that they don’t pretend MCP alone solves anything. They basically say the point is to make small harnesses better, and to shape the ecosystem instead of sulking on the sidelines. That’s a reasonable stance. If you’re building a developer product in this space, refusing to support the thing your users are already asking about can become a principled way to lose. Embracing it, while narrowing it into a safer execution model, is probably the smarter move.
What I’d want to know, though, is how much of this is genuinely new architecture and how much is just a nicer wrapper around the same old tool-calling mess. The example with Jev is cute, but examples are cheap. The hard part is whether this actually makes integrations less brittle, less context-hungry, and less annoying to debug when the model does something weird. That’s the test, not the announcement.
Reference: You Said No MCP! | Earendil
