I'm trying to wrap my head around the idea of "MCP Apps", but I'm having difficulty knowing if any of the metaphors track. It's an app interface in an iFrame in your AI chat. So if it's a coding environment app, it is sort of like an inversion of "an agent in your IDE" to "an IDE in your agent". Does that matter? It feels like it might just because conversation is the dominant interface, and that is going to be more approachable for more people. But I don't know what it means about the ephemerality of the IDE inside the chat. Like, at what point is it no longer expected to maintain state with the repository? When it scrolls out of view? When there's a newer one? Does anyone have an intuition for it they can share? Interactive read-only views seem easy enough to integrate into that workflow, but having stateful apps feels really weird. Conversations are append only.

had to look that up https://modelcontextprotocol.io/extensions/apps/overview apps that render inside other harnesses like Claud Desktop, and provide an actual UI within a dialogue UI.

In a coding context it seems like an interesting way to give source code some UX. What if the AI implements the stripe API and it uses an MCP App to show me exactly what the end user UX looks like and how the UX triggers the API and then the state changes the API goes through as the payment is handles (so some sort of visual representation that is not the end user UX). One would be able to click around to learn how the API works, I would be able to see if the user can get into a bad state. Beyond MCP Apps, we need a new material for source code, that is 100x easier to comprehend.

So I have been experimenting with putting IDEs in apps, and I was considering a json-rpc interface too, so the MCP apps thing makes sense to me. With MCP you can dynamically add to the MCP tool list, so the inner app can offer services as well. So you could code an integration inside the app, and then that integration can be offered a service to the dialog bot, or the outer-bot can code an integration inside the app... to extend its reach. The value is the MCP app is a sandbox, so its safer, maybe it holds credentials? It becomes a substrate for the bot to use, that is somehow a safer boundary, but it can offer the extensibility of the python code environment but without being tied to python. I can imagine putting an embedded linux as an MCP app. The statefulness is however much you want, I have an MCP that offers an export service to a bot, so then it chooses whether to save state or not, and can save that file or not. (this is not with MCP Apps though, this is with channel server but I think it ends up similar)