I keep seeing Italian AI influencers saying “MCP is dead, now there are CLIs,” “MCP is dead, now there are skills.” So let me start with some encouragement. Good job, folks - spread the word about artificial intelligence, you have my full support. And you also have two smacks on the back of the neck, the educational kind that hurt but hurt for your own good, to remind you of something.
The secretary doesn’t install a CLI
If you want a secretary to use artificial intelligence, you don’t have her install the CLI, you don’t have her install skills. You give the secretary a web interface. And if you give her a web interface, that web interface in turn has to deal with a little server that you or someone else controls, and those in turn have to deal with other network services - that is, the MCP servers of Google, of Stripe, the one you built at home, the one your grandma made.
So you have this idea of AI as something that sits on your little computer. You do your experiments, even creative ones, you do your things, you share local runners - a Llama, all the fun little bits, the frameworks, the Cheshire Cat. People don’t give a damn about this stuff.
What MCP is actually for
MCP is made for composable AI agents that are available on the network and whose components are available on the network. So the next person who comes and tells me “ah, but why skills, why the CLI” - folks, artificial intelligence isn’t just that thing sitting on your computer. When you move to protocol-level stuff like MCP, we’re talking about artificial intelligence, about agents used by groups of people. It’s not your damn laptop. I hope that’s clear.