Build Hour: Apps in ChatGPT

ChatGPT apps are moving from “chat-only” to full, interactive experiences by combining three pieces: an apps SDK that injects UI into ChatGPT, an MCP server layer that supplies tools/actions and contextual data, and (optionally) a web component that can react in real time. The practical payoff is clear in the demos: users can ask for hiking recommendations that pull from AllTrails data, generate a styled flyer through Adobe Express templates, and take dynamically rendered quizzes—without the model needing direct access to those external systems on its own.

Since DevDay, OpenAI has shipped the building blocks for this ecosystem. The apps SDK defines how ChatGPT hosts UI widgets, while MCP servers let apps fetch context and perform actions on the user’s behalf, including read/write capabilities. Developers also get an apps docs site on developers.openai.com to learn the components needed to build experiences, plus a public app submission flow and an app marketplace for discovery and installation. A UI component library and sample repositories aim to reduce the time spent on styling and scaffolding, and a new “Docs MCP server” is positioned as a faster on-ramp for development.

The demos make the architecture tangible. In the AllTrails example, ChatGPT interprets intent (“find me dog friendly trails in Marin”), routes it to an AllTrails MCP app, and returns structured results that the UI presents as a refined, engaging experience. The Adobe Express demo shows how an MCP server can trigger template selection and then fill event details into a generated flyer. The quizzes app demonstrates a different pattern: UI content is dynamically rendered from structured tool output, so generation happens on demand and doesn’t require storing quiz content in a backend.

A key conceptual section explains what MCP contributes. MCP provides tools, actions, and resources so an AI client can call external logic—potentially returning a URI for a web component that ChatGPT renders. The most powerful setup pairs MCP-driven data and logic with an interactive web component that can use the window.openai capability API to update state, react to context, and call tools from inside the widget. Display modes like inline, picture-in-picture, and full-screen are highlighted as ways to keep users engaged.

The build walkthrough then shifts from concepts to workflow. Developers are encouraged to use the traditional path—read docs, grab samples, code, deploy—but Codex plus the Docs MCP server is pitched as an AI-first alternative that can handle scaffolding and documentation ingestion. The live process installs Codex via npm, adds the OpenAI developer docs MCP server into Codex, and then prompts Codex to generate a “ping pong” ChatGPT app with a simple UI and MCP server. Codex produces a minimal app repo, which is run locally and then tunneled so ChatGPT can load it in developer mode.

The culminating demo upgrades the game into a real-time multiplayer experience with scores, a lobby, and postgame analysis powered by additional tool calls. After the match, the model provides stats like win rate, rally length, and targeted improvement tips, turning gameplay data into actionable coaching. The session closes with best-practice guidance—extract value instead of porting entire apps, treat ChatGPT as the “home” experience, and design for multi-turn refinement—along with Q&A on MCP compatibility, local development tooling (tunneling with services like Enro or Cloudflare), design freedom, authentication requirements, and monetization approaches (auth flows, in-app checkout links, and emerging agent commerce protocol/instant checkout).