MCP is plumbing

It’s been a year this week since Google handed A2A to the Linux Foundation, and six months since Anthropic did the same with MCP. Both protocols now live under a neutral foundation, AWS, Google, Microsoft, Anthropic and OpenAI have all signed up, and the specs get versioned by working groups instead of by whoever shipped last.

I think this is great. I also think it’s the least exciting thing to happen in AI this year, and those are the same thought.

MCP is plumbing. It’s a standard way for a model to find out what tools exist, what arguments they take, and how to call them. A2A is plumbing one floor up: how an agent finds another agent and hands it a job. Both are the kind of thing that should be boring, governed by a foundation, and forgotten about by everyone except the people maintaining the spec. That’s what happened to HTTP. Nobody puts “built on HTTP” on a landing page.

And yet. I’ve sat through more vendor pitches this year than I’d like where “MCP support” was a headline feature. One had it on a slide by itself. Fine, your product has a plug on it. That tells me nothing about whether the thing on the other end of the plug is any good.

The protocol doesn’t make your tools well described. It doesn’t make your agent pick the right one. It doesn’t make the data behind the tool current, or the permissions sensible, or the failure modes survivable. Every hard problem I’ve hit building agent systems has been on one side of the pipe or the other. Never in the pipe.

So yes, put the pipes in a foundation. Standardize them, version them, argue about the auth spec in a working group. That’s exactly where pipes belong. Just don’t invoice me for them.