The Birth of a New Platform
It’s not often that we see the birth of a new platform. We’re seeing one right now, with an app store for the AI age being built from scratch right before our eyes. OpenAI recently announced that developers can now submit apps to ChatGPT.
When that happens, it’s worth understanding the mechanics of it. With that in mind, I went through building a very simple app and getting it to run and render inside ChatGPT:
Some observations:
These are very early days, in every sense: specs, tooling, patterns, and monetization. It really does feel similar to the early days of the iOS App Store. It isn’t clear yet what kind of apps will feel at home on this new platform.
Specs. OpenAI has simply made up its own conventions on top of the Model Context Protocol (MCP). Meanwhile, Anthropic has a proposed extension to MCP that is currently in the early stages of community feedback.
Tooling. There isn’t any. Other than whatever exists to write a Python or TypeScript server, or more specifically, an MCP server. Developers need to ensure they’re conforming to the OpenAI-specific conventions that the MCP server must obey (e.g., serving ui://... resources) that will make it a valid app in ChatGPT.
There isn’t even good (or any) tooling to test and iterate on your app server. Anthropic’s MCP Inspector will let you introspect and execute a generic MCP server. MCPJam claims to have an MCP inspector with support for ChatGPT apps, but I couldn’t get it working. I had to vibe-code a simple “emulator” to render ChatGPT apps locally. OpenAI’s recommended process is to turn on “developer mode” on your real ChatGPT account (which is possible only on paid plans!) and tunnel your local MCP server into it. I’ll be polite and say that is… not ideal.
Monetization. No official word yet from OpenAI, other than indirect monetization via distribution or transactions due to traffic coming from ChatGPT.
But in spite of all these shortcomings, I’m betting developers will flood the gates. One cannot just pass up an opportunity to reach an audience of 800 million weekly active users. This is right out of the aggregation theory playbook.
From the point of view of apps that have established brands, there is a deep strategic tension here. LLMs are driven by intent expressed in natural language. This leads to decoupling intent from brand name recognition. For example, on my phone, I translate the intent “I want to check out houses for sale” into finding and tapping the icon for the Redfin app. In ChatGPT, that intent will lead to the LLM selecting an appropriate app which can satisfy that intent. It could be Redfin or Zillow or any number of other real estate apps. I as a user don’t need to know or care. I’m trusting ChatGPT to simply pick the right one and the best one. I riffed on this a bit more here:
There are many unknowns. As Ilya Sutskever said: the answer will reveal itself.