The ChatGPT apps framework. Define widgets and schemas - the framework handles MCP resources, tools, and handlers automatically.

Note: This project is a fork of OpenAI's Apps SDK Examples, enhanced with an automated framework to reduce boilerplate and simplify widget creation.

For AI Coding Tools: If you're using an AI coding tool like Cursor, Claude Code, or Codex, try prompting:

"Read the /docs/AGENTS.md and help me use this chatjs framework."

npx create-chatgpt-app my-app
cd my-app

Or clone this repository to use as a starting point.

pnpm install
pnpm run build

Terminal 1 - Frontend assets server:

pnpm run serve

Terminal 2 - MCP server:

cd server
pnpm start

Terminal 3 - Expose to internet:

ngrok http 8000

From the ngrok command, You will get a public URL.

For example: https://<custom_endpoint>.ngrok-free.app/mcp

💡 Tip: If you used npx create-chatgpt-app, your project already includes 6 example widgets (pizzaz map, albums, carousel, list, solar-system, and todo) ready to try!

To add these apps to ChatGPT, enable developer mode, and add your apps in Settings > Connectors.

Add your local public URL from ngrok(like 'https://<custom_endpoint>.ngrok-free.app/mcp') to ChatGPT in Settings > Connectors.

Create src/components/my-widget/index.jsx:

import React from 'react';
import { createRoot } from 'react-dom/client';

function MyWidget() {
  return <div>My Widget Content</div>;
}

const rootElement = document.getElementById('my-widget-root');
if (rootElement) {
  const root = createRoot(rootElement);
  root.render(<MyWidget />);
}

Edit server/src/server.ts:

const widgets = [
  // ... existing widgets
  {
    component: "my-widget",  // matches folder name
    title: "My Widget",
    schema: z.object({
      param: z.string().describe("Parameter description")
    }),
    handler: async (args) => ({
      text: "Widget rendered!",
      data: args
    })
  }
];

# Terminal 1: Rebuild
pnpm run build

# Terminal 2: Restart server
cd server
pnpm start

Done.

• component: "my-widget" maps to src/components/my-widget/
• Build script auto-discovers all components
• Server auto-syncs version from package.json for asset hashing
• Framework auto-generates MCP tools, resources, and handlers

404 on widget assets:

pnpm run build      # Rebuild
cd server && pnpm start  # Restart server

Component not found:

• Check src/components/your-widget/index.jsx exists
• Check widget definition in server/src/server.ts

chatjs/
├── src/components/          # React components (add here)
│   └── my-widget/
│       └── index.jsx
├── assets/                  # Built assets (generated)
├── server/src/
│   ├── framework/           # MCP framework
│   └── server.ts           # Widget definitions (edit here)
└── package.json            # Version (auto-synced)

{
  component: string;              // Component folder name (required)
  title: string;                  // Display name (required)
  description?: string;
  schema: ZodType;                // Input schema (required)
  handler: (args) => Promise<{   // Handler (required)
    text: string;
    data?: Record<string, any>;
  }>;
  meta?: {
    invoking?: string;            // "Loading..."
    invoked?: string;             // "Loaded!"
    widgetDescription?: string;   // For AI model understanding
  };
}

This project is built upon OpenAI's Apps SDK Examples, which provides the foundation for creating ChatGPT apps using the Model Context Protocol (MCP).

Key enhancements in Chat.js:

• Automated MCP framework that eliminates 300+ lines of boilerplate
• Convention-based widget registration
• Automatic asset URL generation with version syncing
• Simplified API for creating widgets
• npm initializer (create-chatgpt-app) for instant project setup

Credits to OpenAI for the original Apps SDK implementation and example widgets.

MIT