Apps are moving inside the chat, and it's worth paying attention

There was a moment last year when the question closed for me for good: the models already build interface better than I do. I asked one of them to rewrite my blog in a single sentence, and what came back was a clean search box with a blur on the backdrop and accessibility baked in, the kind of work I would have billed a full day for not long ago. Since then one fairly simple thought has refused to leave me alone, which is that if the model can assemble the interface on its own, I genuinely don’t understand why I am still sitting here staring at a wall of chat text.

The answer started taking shape right now, and it has a name, MCP apps. I think this is the next big turn, the kind that resets how we use these tools, and what follows is me walking through it calmly and in order, explaining why I am watching this more closely than almost anything else in AI.

What it actually is, in plain terms

Plain MCP is the back end of an AI agent: you ask for the weather, the agent calls the right tool, and you get a line of text back, something like “eighteen and clear,” which is useful and completely flat. MCP apps are the part of that picture that grew a face, because the same tool call can now return a live interface right inside the chat, a weather card with an hourly graph and buttons that actually do something, and it is a working widget that the agent can talk to and that you can reach out and touch.

If you have ever used Telegram Mini Apps you already hold the right mental model, where a small application lives inside a host you already have open and stays wired to its own back end, and MCP apps are that same idea with the host quietly swapped from a messenger to an AI assistant.

It already works today

The thing worth sitting with is that none of this is a roadmap promise. Open Claude or ChatGPT today and the surface is already there, Anthropic’s own visualizer is a first-party MCP app, and the same standard draws widgets inside ChatGPT, VS Code, Cursor and Goose, so while everyone argues about what the interface for this new class of assistant will eventually look like, pieces of that interface are already shipping in production and spinning on your screen.

Why this is a genuine turn

Andrej Karpathy framed it well when he said we have a new kind of computer and that talking to it today feels like talking to a terminal, where you have direct access to an enormous intelligence and yet nobody has invented the mature graphical language for it, as if we were back in the seventies when everything was text. By the same logic we are living through the radio era of this thing, the moment when the first television shows were simply radio broadcasts with a camera pointed at them because no one had yet imagined what the new medium could do on its own terms. MCP apps are the first serious attempt to find an interface language for this new computer, and the most honest signal that the attempt is real looks like this: Anthropic and OpenAI, who compete on quite literally everything, sat down and co-authored a single standard together, and companies at that level only agree to share a layer once the layer has become ground that all of them have to stand on.

Google is walking in from the other side

This is where it gets genuinely interesting, because Google arrived at the same problem from its own angle and brought A2UI, which stands for agent-to-user-interface. Its approach is native-first: instead of handing over a finished web page, the agent describes the interface declaratively and the client draws it with its own native components, whether that is React, Flutter or SwiftUI. Put crudely, MCP apps hand the host a finished web page inside a sandbox while A2UI hands over a blueprint and lets the host build the picture from its own parts, and the two are not necessarily rivals, since there is already a guide for running A2UI on top of MCP. The signal that matters most lives in the roster of who showed up: when Anthropic, OpenAI and Google all pour into the same layer at the same time, that is an industry agreeing the interface is now generated on the fly, and three of the largest labs very rarely land on the same bet at the same moment without a serious reason.

Where we actually are right now

Here comes the honest part, without which the whole picture would read like an ad. All of the strength today sits on the supply side, where the standard exists, the hosts exist, the public servers already number close to ten thousand, partners on the scale of Figma and Canva have shown up, and a payment rail through Stripe and OpenAI arrived back in February. The demand and distribution side, by contrast, is standing almost empty, because the open spec still has no built-in checkout and no shared place to browse all these apps, finding the right server is still its own small quest, and ordinary users have not arrived in anything like the numbers the developers already have. I read that as the clearest sign that we are very early, because the window is open precisely because half the market has not been built yet, and it will close the moment somebody finishes building it.

And why I won’t soften this one

I hold most of my AI takes carefully, but on this one I have no caution left at all. Almost no one is writing about MCP apps yet, and I want to put it on your radar early, because it is rare for me to feel this sure about something. A standard that Anthropic, OpenAI and Google all agreed on within half a year, hosts with hundreds of millions of people inside them, a payment rail that already works, and at the same time a whole half of the market that handles demand, distribution and trust still standing empty and waiting for someone to build it. The window here is wide open, and it is genuinely big.

It is quiet around this right now, the topic has not surfaced yet, and that makes it a calm moment to actually dig in without the rush. If you are building something, I would look straight into that empty half, into discovery, into trust, into the one job where a live interface genuinely beats a wall of text. Anyone can ship server number nine thousand today, and the people who walk into the unbuilt half and plant themselves there will be the ones who got in early.

I am watching this more closely than anything else in AI right now, and I wanted to share it with you early. Take a look :)

— Ilya