OpenAI plugin extensions: your app is now a panel

For two years the integration question was "can ChatGPT call our API?" As of DevDay 2026 on September 29, the question is "does our product have a home in the sidebar?" OpenAI's plugin extensions turn integrations from a tool call into a surface — and the plumbing underneath is MCP.

What actually happened

Plugin extensions give an app its own spot in the ChatGPT sidebar and let developers ship interactive panels that sit alongside the conversation, according to TechCrunch's writeup. You can also register viewers for the file types your product handles, so a user opening one of your files inside ChatGPT sees your renderer instead of a text dump. On stage the demos were Figma comments, Photoshop editing, and OpenAI's own Meetings app running inside the chat window. Altman's framing: an editor, a dashboard, or a whole workspace built directly into ChatGPT and Codex.

The developer-side pieces matter as much as the surface. OpenAI shipped a Plugin Creator tool, a redesigned submission flow with clearer review feedback, better ranking and recommendations in the directory, and per-permission approval so users grant access piece by piece rather than all at once. Sites can now host plugins themselves.

The one to actually read the spec on: OpenAI is adding support for the proposed MCP Events specification, which lets a plugin kick off automations in response to events in a connected app. That flips the direction of the integration. Today your MCP server waits to be called. With events, your system emits — a paid invoice, a failed sync, a new order — and the assistant reacts.

Why plugin extensions matter for your business

Distribution moved, and it is not your website. If your customers live in ChatGPT for four hours a day, a sidebar panel is closer to them than your dashboard is. That is worth building for. It is also a channel you rent, on terms set by a company that has already retired a browser, an API, and a handful of models this year.

MCP Events is the piece worth building on. Events are a protocol proposal, not an OpenAI product. Build your event emitters against the spec, keep them in your own service, and point them at whichever assistant your customers use. Do the work once and it survives the vendor. Wire it only into ChatGPT's flavor and you will do it again for the next surface.

Per-permission approval is a scoping deadline. Granular consent means users will finally see exactly what your plugin asks for, one item at a time. Audit your scopes before that screen is what sells your integration. "Read and write everything" reads very differently on its own line.

Sources: TechCrunch, OpenAI.