ChatGPT Apps SDK

The ChatGPT Apps SDK is the developer framework OpenAI announced at DevDay 2025 for building interactive applications that run inside ChatGPT. Unlike the earlier plugin system, retired in 2024, Apps can render rich UI (cards, carousels, forms, full-screen views) and keep context across a conversation.

The SDK is built on the Model Context Protocol (MCP), the open standard Anthropic originated and OpenAI adopted. This means an MCP server written for Claude or another MCP-compatible host can be adapted into a ChatGPT App with relatively modest additional work, and the same server can serve multiple surfaces.

Launch partners at DevDay 2025 included Canva, Expedia, Spotify, Figma, Booking.com, and Zillow. The stated design goal: users should not have to leave ChatGPT to book travel, create a design, browse listings, or buy something. The app renders inline inside the conversation, inherits the conversational context, and can take action on the user's behalf.

Apps complement the Agentic Commerce Protocol (ACP) rather than replace it. ACP handles the transaction layer, originally behind Instant Checkout, which OpenAI retired in March 2026 in favour of purchases inside retailer Apps. Apps handle the experience layer: how a catalog is browsed and, now, where the purchase happens.

Without an App, a retailer's presence in ChatGPT is whatever ChatGPT's retrieval chooses to show: text mentions, product cards from the Shopping feed, citations from the web. Retailers can improve the inputs (product information, feeds, schema) but do not own a surface inside the conversation.

A retailer with a ChatGPT App gets:

• An owned, branded surface. Brand name, visual identity and navigation live inside the conversation.
• Interactive discovery. Carousels, filters and comparison tools inside the chat flow.
• Conversational context. The App inherits what the shopper has already said, so recommendations start from the real conversation.
• Deeper catalog access. The App queries the full catalog through its own backend, not only what reached the Shopping feed.
• Checkout continuity. With ACP, the shopper can browse, select and pay inside the App without leaving ChatGPT.

An App is an investment most relevant to large brands with the engineering to build and maintain it. For most retailers the first and larger win is still appearing in ChatGPT's own answers, which depends on product information rather than code.

At the protocol level a ChatGPT App is an MCP server with an extension for rich UI. The developer defines:

Tools. The actions the App can take: search products, add to cart, check inventory, book an appointment. Each has a schema for inputs and outputs.

Resources. The data the App can provide: catalog, history, availability. Retrieved on demand.

UI components. The SDK extends MCP with a rendering layer. The App returns structured UI payloads that ChatGPT renders inline: cards, carousels, forms, confirmations, full-screen views.

During a conversation the model decides when a registered App is relevant, calls its tools, receives structured data and renders the UI inline. The shopper interacts with it and the App handles the response.

A retail App connects to the same catalog, inventory and checkout systems as the main storefront. It is a new surface for the same data, which means it inherits the same data problems. A product with a thin description or missing attributes is just as thin inside an App as it is in the Shopping feed. Getting the catalog complete comes first, and that is the work Paz handles.