OpenAI Apps Template (Todo App)

Build, preview, and ship a ChatGPT App that renders a responsive React widget and talks to a Model Context Protocol (MCP) server—all from one repository.

• Single-file widget delivery – every widget is bundled into dist/<name>.html with inline JS + CSS so ChatGPT can embed it via ui://widget/....
• Batteries-included MCP server – exposes SSE + POST endpoints, advertises tools/resources, and returns structured todo data plus widget metadata.
• Local-first DX – Vite dev server with HMR, Tailwind styles, React 19, TypeScript, and helper hooks that mirror the ChatGPT window.openai runtime.

Use this template as a starting point for any small app that needs both an interactive UI and tools the ChatGPT agent can call.

---

At a Glance

Piece	Tech	Purpose
src/widgets/todo.tsx	React + hooks	Interactive todo widget that can optimistically update local state while syncing with the agent
server.ts	Node + @modelcontextprotocol/sdk	Minimal MCP server exposing todo list, add_todo, toggle_todo, delete_todo, and the widget resource
build.ts	Vite + esbuild	Discovers widget entries, bundles them, and emits inline HTML files ready for ui://widget/...
dev.ts	Vite loader	Lets you hot-reload any widget via http://localhost:5173?entry=<name>

---

Prerequisites

• Node.js 18+
• npm (or pnpm/yarn if you adapt the scripts)
• (Optional) ngrok or similar tunnel to share the MCP server with ChatGPT

Install dependencies:

npm install

---

Quick Start

1. Build the widget bundle

   npm run build
   # → dist/todo.html (inline CSS + JS)

2. Run the MCP server

   npm start
   # SSE  GET  http://localhost:8000/mcp
   # POST POST http://localhost:8000/mcp/messages?sessionId=...

3. Preview the widget locally (optional)

   npm run dev
   # open http://localhost:5173?entry=todo

4. Expose to ChatGPT (optional)

   ngrok http 8000
   # use https://<subdomain>.ngrok-free.app/mcp as the connector URL

---

Development Workflow

Widget iteration

• Widgets live under src/widgets/<name>.tsx. Each file should call createRoot and mount itself into __WIDGET_ROOT_ID__.
• npm run dev spins up Vite with HMR. Navigate to http://localhost:5173?entry=<name> to load that widget.
• Hooks under src/hooks provide typed access to window.openai globals:
  • useWidgetProps pulls structured tool output (toolOutput.todoList in this case).
  • useWidgetState keeps local widget state in sync with the ChatGPT host via window.openai.setWidgetState.
  • useCallTool, useRequestDisplayMode, etc. wrap MCP APIs.
• The todo widget reads theme, safeArea.insets, displayMode, and maxHeight so it can match ChatGPT’s light/dark palette, respect safe areas, and resize itself for PiP, inline, or fullscreen layouts without extra work on your end.

Building for ChatGPT

• npm run build auto-discovers every file in src/widgets (or nested index.tsx files inside directories) and emits dist/<name>.html.
• Each HTML file inlines:
  1. Shared Tailwind styles from src/styles/main.css
  2. A fallback window.todoData block for standalone preview
  3. The minified widget bundle
• The build script keeps the output self-contained so ChatGPT can fetch ui://widget/<name>.html with no extra assets.

MCP server

• server.ts wires up @modelcontextprotocol/sdk with an SSEServerTransport:
  • GET /mcp starts the SSE session.
  • POST /mcp/messages?sessionId=<id> streams tool messages/responses.
• Tools shipped by default:

  Name	Description	Arguments
  list_todos	Return the current todo list	none
  add_todo_item	Insert a new todo	{ title }
  toggle_todo_item	Flip completion state	{ todoId }
  delete_todo_item	Remove a todo	{ todoId }

• All tool responses include:
  • structuredContent.todoList (mirrors the widget data contract)
  • _meta.openai/* descriptors so ChatGPT knows to render the widget
• The server keeps a simple in-memory globalTodoState. Swap this out for a real database or API when you graduate from the demo.

---

Connecting to ChatGPT Apps

1. Enable Developer Mode in ChatGPT → Settings → Connectors.

2. Create a new “Model Context Protocol” connector.

3. Point the MCP URL at your server (local tunnel or deployed host):

   https://<your-ngrok-subdomain>.ngrok-free.app/mcp
   

4. In a ChatGPT conversation, ask something like “Show my todo list” or “Add ‘Review PR’ to my todos.” ChatGPT will call the appropriate tool(s), and the widget will render using the ui:// resource the server exposes.

---

Project Structure

openai-apps-template/
├─ src/
│  ├─ widgets/
│  │  └─ todo/
│  │     ├─ index.tsx              # Widget entry point (self-mounting)
│  │     └─ components/
│  │        ├─ NewTodo.tsx         # Input row for creating todos
│  │        └─ TodoItem.tsx        # Presentational list item
│  ├─ hooks/              # window.openai + MCP helpers
│  ├─ styles/main.css     # Tailwind layer shared by widgets
│  └─ utils/              # UI utilities (media queries, etc.)
├─ build.ts               # Inline widget bundler
├─ dev.ts                 # Vite preview loader (?entry=<name>)
├─ server.ts              # MCP SSE server + todo tools
├─ dist/                  # Generated ui:// HTML widgets
├─ package.json
└─ README.md

---

Scripts & Configuration

Command	Description
npm run build	Bundle every widget into dist/*.html
npm start	Launch MCP server (defaults to port 8000)
npm run start:all	Build widgets, then start the server
npm run dev	Vite dev server for widget HMR preview

Environment variables:

• PORT – overrides the MCP server port (default 8000).

---

Extending the Template

• Add another widget: drop src/widgets/notes.tsx, ensure it mounts to __WIDGET_ROOT_ID__, run npm run build, then expose it as a new resource/tool in server.ts.
• Persist todo data: replace globalTodoState with a database call or REST client. The only contract the widget needs is { todoList: TodoList }, where TodoList is just { title: string; todos: TodoItem[] }.
• Enhance UI state: use useWidgetState for optimistic updates; it already mirrors state back to ChatGPT so the host stays in sync between agent turns.
• Add more tools: register descriptors in server.ts (ListToolsRequestSchema handler) and implement the behavior inside the CallToolRequest handler.

---

Troubleshooting

• Widget shows stale data – ensure the MCP response includes fresh structuredContent.todoList and _meta.openai/outputTemplate. The widget only resyncs when the host sends new structured data.
• “Build not found” error – run npm run build before starting npm start so dist/todo.html exists.
• No widget in ChatGPT – verify the connector is allowed to render widgets (openai/widgetAccessible: true) and that the ui://widget/<name>.html resource is listed via ListResources.
• CORS or tunnel issues – the server enables Access-Control-Allow-Origin: *, but your tunnel must forward both GET /mcp and POST /mcp/messages.

---

Happy shipping! Open issues or PRs if you add new widgets, storage adapters, or deployment recipes.