Reference apps for embedding live widgets in ChatGPT

What it does The repository is a gallery of example widgets—pizza menus, 3D solar systems, shopping carts—plus the Model Context Protocol (MCP) servers that power them. Each server advertises tools via JSON Schema, handles call_tool requests, and returns structured responses with _meta.ui.resourceUri metadata so the Apps SDK can render HTML, CSS, and JavaScript inline next to ChatGPT’s text replies. It is explicitly positioned as copy-paste inspiration rather than a production framework.

The interesting bit The clever part is the two-way binding between the LLM and the UI: widgets can read host state like toolInput and theme, mutate shared state with setWidgetState, and even trigger follow-up tool calls via callTool, so the conversation and the interface stay in sync without the developer hand-rolling a custom bridge. The build pipeline hashes and bundles each widget with Vite so the resulting MCP servers serve them as self-contained static assets.

Key highlights

• Ships paired Node and Python MCP servers for most examples, letting you pick your stack.
• Demonstrates stateful widgets (shopping cart) that persist widgetState across turns using widgetSessionId.
• Includes an authenticated server example showing OAuth-gated tool calls.
• Bundles every widget entrypoint through a single build-all.mts Vite orchestrator into versioned assets/.
• Covers the full window.openai surface: requestDisplayMode, openExternal, sendFollowUpMessage, and theme awareness.

Caveats

• The Python Pizzaz server caches widget HTML with functools.lru_cache, so you must restart it after rebuilding assets.
• Chrome 142 users need to disable the #local-network-access-check flag to see local widget UI during development.
• The shopping-cart demo keeps state client-side via metadata; the README notes you should persist server-side in production.

Verdict Worth a look if you are building ChatGPT connectors that need more than text—buttons, maps, carousels, and checkout flows. Skip it if you are not targeting the OpenAI Apps SDK ecosystem; the examples are tightly coupled to ChatGPT’s host APIs and MCP wire format.