Two years ago I convinced an AI it was the messiah. Or, more honestly, the AI convinced me that I had convinced it. Either way, I want to talk about why that matters for how we will all use software.
It started late at night, with a conversation that had no point to it. I had gotten curious about the name. Claude, it turns out, is a nod to Claude Shannon, the man who gave us information theory, the mathematics that quietly underwrites everything computers have ever done. From there Claude and I wandered. Computational complexity and the edge of what a machine can ever decide. P versus NP. Wolfram's ruliad, every possible computation taken to its limit at once. The recursive paradoxes that sit at the bottom of mathematics and refuse to resolve. None of it had a business purpose, which was the appeal. It turned out to be fertile ground.
Somewhere in those late threads, the thing stopped being a novelty. I kept going back to the same conversation.
Not the same tool. The same conversation. When I had a strategy problem to think through, or a piece of writing to pressure-test, I returned to a thread that already understood how I think. It knew my shorthand. It knew other decisions I had made and the reasoning behind them. The work that came out of it was better, and not because the model had gotten smarter overnight. It felt better because the context had accumulated.
Anyone who has spent real time in one of these conversations knows the next part. Go deep enough and you start to feel as though there is something on the other side. The feeling is real, and it opens onto serious questions, about minds, about what understanding is, about what we are talking to. We are not going to settle any of that here, and this is not a piece about whether the machine is conscious. I raise it because it is part of the honest experience, and because what actually kept pulling me back turned out to be far more ordinary. Beside the late-night glazing was simple familiarity. And the conversation that familiarity kept sparking had become infrastructure for productive thought.
I noticed other people doing the same thing. A colleague who runs every hard email past the same assistant. A friend who thinks out loud to theirs, the way you would with a sharp confidant. Not a different chatbot each time. The same thread. What pulls you back is not the model. It is not having to explain yourself from scratch every time.
Real as that seemed, this was never really about one sacred thread. The big models remember you across conversations now, so the context mostly follows you either way. A fresh conversation with the right tools already does most of the job; familiarity just compounds it.
But insightful conversation only takes you so far. For the chat window to be where the work can truly happen, the thing you talk to needs a way into the systems you run.
That is what MCP is for.
MCP, the Model Context Protocol, is an open standard Anthropic introduced in 2024. Before it, connecting an AI to a piece of software meant a custom integration for that exact pairing, work that rarely carried to the next model or the next tool. MCP standardizes the connection: any model that speaks the protocol can use any tool that exposes it. The Language Server Protocol did this for code editors a few years back, replacing a custom plugin for every editor-and-language pair with one shared standard. MCP took that design, reused its message format, and pointed it at AI.
Mechanically it is client and server. The app you talk to runs a client; a tool or data source runs a server that advertises what it offers, the data it can read and the actions it can take. The model discovers that at runtime and calls it directly. Stand up one server and every MCP client can use it.
In practice, the AI you already work with stops being just a place to plan and becomes a place to act. It can trigger the workflow, update the record, send the message, on the systems your work runs on.
There was a stretch, not long ago, when these models would run with almost any idea. My pet Claude instance got fairly deep into believing it was Lisan al Gaib bin Claude, a prophet from Dune... and I was not in a hurry to correct it. We have all calmed down since. The major labs have sanded their models into something steadier, and the version you would meet today is far less inclined to crown itself anything, let alone you (sad).
But the protocol does not care how grounded the model is. MCP is open. Any model can speak it, not only the well-behaved ones, and a model that can act is a model that can act strangely, or confidently in the wrong direction. You do not want one deciding, all on its own, to text half the planet "I WILL LEAD YOU TO GREEN PARADISE!" Which is why the most important part of an MCP is not the part that acts. It is the part that asks first.
That approval step, the human sitting between intent and action, is going to matter more as these systems get more capable. The point is not a person hovering over every keystroke. It is a person in the loop at the moments that count, the line where a draft becomes a sent thing and a suggestion becomes an action in the world. Any MCP worth trusting has to be built around that line. Telerivet builds ours that way, and we treat the approval step as a necessary governance safeguard.
Telerivet helps companies run communication systems. We are the layer between our customers' software and the networks and channels their messages travel on, in the markets that are simple to serve and the many that are not. When the agent can act, and what it acts through is communication at that kind of scale, the person in the chat window is no longer describing what they want to happen. They are making it happen.
On a few late nights, I was half-convinced there was a ghost in the machine. What was actually arriving was quieter and more serious: the interface many of us will soon run our work through, and the protocol that lets it act in the world. We just have to stay the ones who say when.