ChatGPT Plugin Extensions Are the New Mini Apps

On September 29 OpenAI used its DevDay keynote to open ChatGPT's own interface to outside developers. The feature is called plugin extensions. A plugin can now have a home in the ChatGPT sidebar, open as a panel next to a conversation, or become the viewer for a file type the plugin's product handles. ChatGPT has 1.2 billion weekly users. If you build software, that is a new place your product can live, and it comes with a landlord. We have been designing apps that live inside a host for a few years now, mostly inside Telegram. Most of what we learned there applies here.

What a plugin extension is

Until last week a ChatGPT plugin lived inside the thread. It could connect ChatGPT to Slack, SharePoint, Airtable or Google Drive, and whatever interface it had appeared inside the conversation itself. Plugin extensions add rooms of their own. OpenAI's developer docs list the surfaces: sidebar apps that open full screen, conversation panels that sit beside the chat, file viewers and editors, a settings page inside ChatGPT, deep links into a specific item, composer mentions on desktop, and rich forms that can show an image picker. There is also something called Model-App Context, which keeps the model and the app in sync in both directions. Under the hood it is an open specification on top of the Model Context Protocol, with TypeScript and Python SDKs and a public GitHub repo under an Apache 2.0 licence. Your logic runs on your server. The interface renders inside ChatGPT. OpenAI's Thibault Sottiaux said developers can now "build full native apps with plugin extensions," and Canva was the sidebar demo.

Three surfaces, three different products

The first design decision is not visual. It is which surface you ask for, because each one changes what the user is doing when your app shows up. A sidebar app opens full screen. The user picked you from a list, the chat is gone, and you are a normal app with a ChatGPT frame around it. Normal product design applies, with one twist: the model can still see your state through Model-App Context, so that state should be legible to a reader that is not a person. A conversation panel is the hard one. Your app sits next to a live chat, so it gets roughly half the width and a fraction of the attention. The user is mid-task with the model and glances over at you. A panel should do one job per view and show its current state without a tutorial. We learned this in Telegram, where a mini app shares the screen with the keyboard and a chat header, and every row of chrome you add is a row of content the user has to scroll past. A file viewer is a different contract again. The user opened a file and expected to see the file. Your interface should look like the document, not like your brand. OpenAI's docs say reading, live updates and saving are handled together, so the viewer needs honest save and sync states, because the person will not know whether ChatGPT or your server holds the latest version.

The model is your second user

In a normal app you design for one reader. Inside ChatGPT there are two. The person looks at your panel. The model reads your context, writes into it, and may act on it without a click. Composer mentions let a user pull a record from your plugin into a prompt the way they would mention a colleague. Rich forms let the model ask the user a structured question through your plugin and get a typed answer back. That changes copy and structure. Labels have to make sense to a language model as well as to a person skimming. An empty state has to say what is missing in words the model can relay. A status that is only a colour is invisible to the second user.

Automations mean your app can speak first

The part we think matters most is MCP Events, which OpenAI calls a proposed specification. An event in a connected app can trigger the plugin instead of a person typing a request. The examples Nerd's Chalk lists: a project board gets a new task, a feedback channel gets a bug report, a document gets a review comment, and ChatGPT reacts. Users decide what gets watched. For a product team this is a new kind of entry point. Most app design assumes the user opens the app. Here the app opens the conversation, and the first thing the person sees is the model's summary of whatever your plugin raised. If that summary is wrong or noisy, the user switches the automation off and you have lost the surface. Every event you emit is a push notification with a reasoning layer on top. Send fewer.

What Telegram taught us about living in someone else's house

In January 2025 Telegram told every mini app with crypto integrations to run on TON only, with a 30-day migration window and grants of up to $50,000 for teams that moved. Telegram had over 950 million users at the time, and plenty of teams had built real businesses on it. The rule changed anyway. That is the deal when your product lives inside a host, and ChatGPT is no different. The review terms are already specific. A public listing needs a verified developer or business identity, five positive and three negative test cases, domain verification for remote MCP servers, and reviewer access to reproduce the tests. Approval and publishing are separate steps. Plugins with strong utility and satisfaction may get stronger placement in the directory and in conversations, and OpenAI frames that as a possibility rather than a promise. Users approve each plugin's access individually, and permissions live in a Plugins tab shared across ChatGPT, Codex and the new Dots agents. So treat the plugin the way we treat a Telegram mini app: a thin, host-shaped layer over a product that also works without the host. Keep your logic on your own server, which the architecture already pushes you toward. Keep your design system in your own files and build the ChatGPT skin as one theme rather than the source of truth. That is how we approach mobile and mini app work at ClefDev, and it is the main reason a platform rule change is a sprint for our clients rather than a rebuild. Plugin extensions on the web are still "coming soon" for Free and Go users, and composer mentions only work in the desktop app, so the audience this autumn is paying users at a desk. A year ago an app inside ChatGPT was a widget inside the thread. Now it is a sidebar entry, a panel and a file viewer, with a model reading over the user's shoulder and the ability to start the conversation itself. The teams that do well here will design small and write for two readers. They will also keep the real product somewhere OpenAI cannot reorganise. We would start with the panel. It is the hardest surface and the one people will actually use.

Sources

1. OpenAI Developers, "Plugin Extensions", accessed October 7, 2026
2. TechCrunch, "OpenAI expands ChatGPT's plug-ins with app-like interfaces and automations", September 29, 2026
3. Engadget, "OpenAI Dev Day 2026: Live Updates On The Latest ChatGPT And Codex Announcements", September 29, 2026
4. Nerd's Chalk, "ChatGPT Opens Its Sidebar to Outside Apps With Plugin Extensions and Event-Triggered Automations", September 30, 2026
5. OpenTools, "OpenAI updates plugin creation and discovery, but a public listing still needs review", September 30, 2026
6. Decrypt, "Telegram Goes Exclusive With TON, Barring Other Networks From Apps and Games", January 21, 2025