chatgpt-app-builder（chatgpt 应用程序构建器）

ChatGPT App Builder

Build production-ready ChatGPT apps with interactive widgets using the mcp-use framework and OpenAI Apps SDK. This skill provides zero-config widget development with automatic registration and built-in React hooks.

Quick Start

Always bootstrap with the MCP Apps template:

npx create-mcp-use-app my-chatgpt-app --template mcp-apps
cd my-chatgpt-app
yarn install
yarn dev

This creates a project structure:

my-chatgpt-app/
├── resources/              # React widgets (auto-registered!)
│   ├── display-weather.tsx # Example widget
│   └── product-card.tsx    # Another widget
├── public/                 # Static assets
│   └── images/
├── index.ts               # MCP server entry
├── package.json
├── tsconfig.json
└── README.md

Why mcp-use for ChatGPT Apps?

Traditional OpenAI Apps SDK requires significant manual setup:

• Separate project structure (server/ and web/ folders)
• Manual esbuild/webpack configuration
• Custom useWidgetState hook implementation
• Manual React mounting code
• Manual CSP configuration
• Manual widget registration

mcp-use simplifies everything:

• ✅ Single command setup
• ✅ Drop widgets in resources/ folder - auto-registered
• ✅ Built-in useWidget() hook with state, props, tool calls
• ✅ Automatic bundling with hot reload
• ✅ Automatic CSP configuration
• ✅ Built-in Inspector for testing
• ✅ Dual-protocol support (works with ChatGPT AND MCP Apps clients)

MCP Apps vs ChatGPT Apps SDK

mcp-use supports multiple widget protocols, giving you maximum compatibility:

Why MCP Apps?

MCP Apps is the official standard (SEP-1865) for interactive widgets in the Model Context Protocol:

• Universal: Works with ChatGPT, Claude Desktop, Goose, and all MCP Apps clients
• Future-proof: Based on open specification, ensuring long-term compatibility
• Secure: Double-iframe sandbox with granular CSP control
• Zero config: With type: "mcpApps", mcp-use automatically generates metadata for BOTH protocols

Key Point: When you use type: "mcpApps" in your server configuration, your widgets automatically work with both ChatGPT (Apps SDK protocol) and MCP Apps clients. You write the widget once, and mcp-use handles the protocol translation.

Creating Widgets

Simple Widget (Single File)

Create resources/weather-display.tsx:

import { McpUseProvider, useWidget, type WidgetMetadata } from "mcp-use/react";
import { z } from "zod";

// Define widget metadata
export const widgetMetadata: WidgetMetadata = {
  description: "Display current weather for a city",
  props: z.object({
    city: z.string().describe("City name"),
    temperature: z.number().describe("Temperature in Celsius"),
    conditions: z.string().describe("Weather conditions"),
    humidity: z.number().describe("Humidity percentage"),
  }),
};

const WeatherDisplay: React.FC = () => {
  const { props, isPending } = useWidget();

  // Always handle loading state first
  if (isPending) {
    return (
      <McpUseProvider autoSize>
        <div className="animate-pulse p-4">Loading weather...</div>
      </McpUseProvider>
    );
  }

  return (
    <McpUseProvider autoSize>
      <div className="weather-card p-4 rounded-lg shadow">
        <h2 className="text-2xl font-bold">{props.city}</h2>
        <div className="temp text-4xl">{props.temperature}°C</div>
        <p className="conditions">{props.conditions}</p>
        <p className="humidity">Humidity: {props.humidity}%</p>
      </div>
    </McpUseProvider>
  );
};

export default WeatherDisplay;

That's it! The widget is automatically:

• Registered as MCP tool weather-display
• Registered as MCP resource ui://widget/weather-display.html
• Bundled for Apps SDK compatibility
• Ready to use in ChatGPT

Complex Widget (Folder Structure)

For widgets with multiple components:

resources/
└── product-search/
    ├── widget.tsx          # Entry point (required name)
    ├── components/
    │   ├── ProductCard.tsx
    │   └── FilterBar.tsx
    ├── hooks/
    │   └── useFilter.ts
    ├── types.ts
    └── constants.ts

Entry point (widget.tsx):

import { McpUseProvider, useWidget, type WidgetMetadata } from "mcp-use/react";
import { z } from "zod";
import { ProductCard } from "./components/ProductCard";
import { FilterBar } from "./components/FilterBar";

export const widgetMetadata: WidgetMetadata = {
  description: "Display product search results with filtering",
  props: z.object({
    products: z.array(
      z.object({
        id: z.string(),
        name: z.string(),
        price: z.number(),
        image: z.string(),
      })
    ),
    query: z.string(),
  }),
};

const ProductSearch: React.FC = () => {
  const { props, isPending, state, setState } = useWidget();

  if (isPending) {
    return (
      <McpUseProvider autoSize>
        <div>Loading...</div>
      </McpUseProvider>
    );
  }

  return (
    <McpUseProvider autoSize>
      <div>
        <h1>Search: {props.query}</h1>
        <FilterBar onFilter={(filters) => setState({ filters })} />
        <div className="grid grid-cols-3 gap-4">
          {props.products.map((p) => (
            <ProductCard key={p.id} product={p} />
          ))}
        </div>
      </div>
    </McpUseProvider>
  );
};

export default ProductSearch;

Widget Metadata

Required metadata for automatic registration:

export const widgetMetadata: WidgetMetadata = {
  // Required: Human-readable description
  description: "Display weather information",

  // Required: Zod schema for widget props
  props: z.object({
    city: z.string().describe("City name"),
    temperature: z.number(),
  }),

  // Optional: Disable automatic tool registration
  exposeAsTool: true, // default

  // Optional: Unified metadata (works for BOTH ChatGPT and MCP Apps)
  metadata: {
    csp: {
      connectDomains: ["https://api.weather.com"],
      resourceDomains: ["https://cdn.weather.com"],
    },
    prefersBorder: true,
    autoResize: true,
    widgetDescription: "Interactive weather display",
  },
};

Important:

• description: Used for tool and resource descriptions
• props: Zod schema defines widget input parameters
• exposeAsTool: Set to false if only using widget via custom tools
• metadata: Unified configuration that works for both protocols (recommended)

Content Security Policy (CSP)

Control what external resources your widget can access using CSP configuration:

export const widgetMetadata: WidgetMetadata = {
  description: "Weather widget",
  props: z.object({ city: z.string() }),
  metadata: {
    csp: {
      // APIs your widget needs to call
      connectDomains: ["https://api.weather.com", "https://weather-backup.com"],

      // Static assets (images, fonts, stylesheets)
      resourceDomains: ["https://cdn.weather.com"],

      // External content to embed in iframes
      frameDomains: ["https://embed.weather.com"],

      // Script CSP directives (use carefully!)
      scriptDirectives: ["'unsafe-inline'"],
    },
  },
};

CSP Field Reference:

• connectDomains: APIs to call via fetch, WebSocket, XMLHttpRequest
• resourceDomains: Load images, fonts, stylesheets, videos
• frameDomains: Embed external content in iframes
• scriptDirectives: Script-src CSP directives (avoid 'unsafe-eval' in production)

Security Best Practices:

• ✅ Specify exact domains: https://api.weather.com
• ❌ Avoid wildcards: https://*.weather.com (less secure)
• ❌ Never use 'unsafe-eval' unless absolutely necessary
• ✅ Test CSP in development before deploying

Metadata Configuration Options

Modern Unified Approach (Recommended)

Use the metadata field for dual-protocol support:

export const widgetMetadata: WidgetMetadata = {
  description: "Weather widget",
  props: propSchema,
  metadata: {
    // Works for BOTH MCP Apps AND ChatGPT
    csp: {
      connectDomains: ["https://api.weather.com"],
      resourceDomains: ["https://cdn.weather.com"],
    },
    prefersBorder: true,
    autoResize: true,
    widgetDescription: "Displays current weather",
  },
};

Legacy Apps SDK Approach (Deprecated)

The old ChatGPT-only format (still supported but not recommended):

export const widgetMetadata: WidgetMetadata = {
  description: "Weather widget",
  props: propSchema,
  appsSdkMetadata: {
    // ChatGPT only - snake_case with openai/ prefix
    "openai/widgetCSP": {
      connect_domains: ["https://api.weather.com"],
      resource_domains: ["https://cdn.weather.com"],
    },
    "openai/widgetPrefersBorder": true,
    "openai/toolInvocation/invoking": "Loading...",
    "openai/toolInvocation/invoked": "Loaded",
  },
};

Migration Note: The old format uses appsSdkMetadata with openai/ prefixes and snake_case (e.g., connect_domains). The new format uses metadata with camelCase (e.g., connectDomains) and works for both protocols.

Using Both for Custom ChatGPT Features

You can combine both fields to use standard metadata plus ChatGPT-specific overrides:

export const widgetMetadata: WidgetMetadata = {
  description: "Weather widget",
  props: propSchema,
  // Unified metadata (dual-protocol)
  metadata: {
    csp: { connectDomains: ["https://api.weather.com"] },
    prefersBorder: true,
  },
  // ChatGPT-specific overrides/additions
  appsSdkMetadata: {
    "openai/widgetDescription": "ChatGPT-specific description",
    "openai/customFeature": "some-value", // Any custom OpenAI metadata
    "openai/locale": "en-US",
  },
};

Use Case: When you need to pass custom OpenAI-specific metadata that doesn't exist in the unified format, add it to appsSdkMetadata. The fields will be passed directly to ChatGPT with the openai/ prefix

useWidget Hook

The useWidget hook provides everything you need:

const {
  // Widget props from tool input
  props,

  // Loading state (true = tool still executing)
  isPending,

  // Persistent widget state
  state,
  setState,

  // Theme from host (light/dark)
  theme,

  // Call other MCP tools
  callTool,

  // Display mode control
  displayMode,
  requestDisplayMode,

  // Additional tool output
  output,
} = useWidget<MyPropsType, MyOutputType>();

Props and Loading States

Critical: Widgets render BEFORE tool execution completes. Always handle isPending:

const { props, isPending } = useWidget<WeatherProps>();

// Pattern 1: Early return
if (isPending) {
  return <div>Loading...</div>;
}
// Now props are safe to use

// Pattern 2: Conditional rendering
return <div>{isPending ? <LoadingSpinner /> : <div>{props.city}</div>}</div>;

// Pattern 3: Optional chaining (partial UI)
return (
  <div>
    <h1>{props.city ?? "Loading..."}</h1>
  </div>
);

Widget State

Persist data across widget interactions:

const { state, setState } = useWidget();

// Save state (persists in ChatGPT localStorage)
const addFavorite = async (city: string) => {
  await setState({
    favorites: [...(state?.favorites || []), city],
  });
};

// Update with function
await setState((prev) => ({
  ...prev,
  count: (prev?.count || 0) + 1,
}));

Calling MCP Tools

Widgets can call other tools:

const { callTool } = useWidget();

const refreshData = async () => {
  try {
    const result = await callTool("get-weather", {
      city: "Tokyo",
    });
    console.log("Result:", result.content);
  } catch (error) {
    console.error("Tool call failed:", error);
  }
};

Display Mode Control

Request different display modes:

const { displayMode, requestDisplayMode } = useWidget();

const goFullscreen = async () => {
  await requestDisplayMode("fullscreen");
};

// Current mode: 'inline' | 'pip' | 'fullscreen'
console.log(displayMode);

Custom Tools with Widgets

Create tools that return widgets with dual-protocol support:

import { MCPServer, widget, text } from "mcp-use/server";
import { z } from "zod";

const server = new MCPServer({
  name: "weather-app",
  version: "1.0.0",
});

server.tool(
  {
    name: "get-weather",
    description: "Get current weather for a city",
    schema: z.object({
      city: z.string().describe("City name"),
    }),
    // Widget config (registration-time metadata)
    widget: {
      name: "weather-display", // Must match widget in resources/
      invoking: "Fetching weather...",
      invoked: "Weather data loaded",
    },
  },
  async ({ city }) => {
    // Fetch data from API
    const data = await fetchWeatherAPI(city);

    // Return widget with runtime data
    return widget({
      props: {
        city,
        temperature: data.temp,
        conditions: data.conditions,
        humidity: data.humidity,
      },
      output: text(`Weather in ${city}: ${data.temp}°C`),
      message: `Current weather for ${city}`,
    });
  }
);

server.listen();

Key Points:

• baseUrl in server config enables proper asset loading
• Widget works with BOTH ChatGPT and MCP Apps clients automatically
• widget: {name, invoking, invoked} on tool definition
• widget({props, output}) helper returns runtime data
• props passed to widget, output shown to model
• Widget must exist in resources/ folder

Static Assets

Use the public/ folder for images, fonts, etc:

my-app/
├── resources/
├── public/              # Static assets
│   ├── images/
│   │   ├── logo.svg
│   │   └── banner.png
│   └── fonts/
└── index.ts

Using assets in widgets:

import { Image } from "mcp-use/react";

function MyWidget() {
  return (
    <div>
      {/* Paths relative to public/ folder */}
      <Image src="/images/logo.svg" alt="Logo" />
      <img src={window.__getFile?.("images/banner.png")} alt="Banner" />
    </div>
  );
}

Components

McpUseProvider

Unified provider combining all common setup:

import { McpUseProvider } from "mcp-use/react";

function MyWidget() {
  return (
    <McpUseProvider
      autoSize // Auto-resize widget
      viewControls // Add debug/fullscreen buttons
      debug // Show debug info
    >
      <div>Widget content</div>
    </McpUseProvider>
  );
}

Image Component

Handles both data URLs and public paths:

import { Image } from "mcp-use/react";

function MyWidget() {
  return (
    <div>
      <Image src="/images/photo.jpg" alt="Photo" />
      <Image src="data:image/png;base64,..." alt="Data URL" />
    </div>
  );
}

ErrorBoundary

Graceful error handling:

import { ErrorBoundary } from "mcp-use/react";

function MyWidget() {
  return (
    <ErrorBoundary
      fallback={<div>Something went wrong</div>}
      onError={(error) => console.error(error)}
    >
      <MyComponent />
    </ErrorBoundary>
  );
}

Testing

Using the Inspector

1. Start development server: yarn dev
2. Open Inspector:

- Navigate to http://localhost:3000/inspector

1. Test widgets:

- Click Tools tab - Find your widget tool - Enter test parameters - Execute to see widget render

1. Debug interactions:

- Use browser console - Check RPC logs - Test state persistence - Verify tool calls

Testing in ChatGPT

1. Enable Developer Mode:

- Settings → Connectors → Advanced → Developer mode

1. Add your server:

- Go to Connectors tab - Add remote MCP server URL

1. Test in conversation:

- Select Developer Mode from Plus menu - Choose your connector - Ask ChatGPT to use your tools

Prompting tips:

• Be explicit: "Use the weather-app connector's get-weather tool..."
• Disallow alternatives: "Do not use built-in tools, only use my connector"
• Specify input: "Call get-weather with {city: 'Tokyo'}"

Dual-Protocol Note: When using type: "mcpApps" in your server configuration, your widgets automatically work in both ChatGPT (via Apps SDK) and MCP Apps clients (like Claude Desktop, Goose). You can test the same widget in multiple clients without any code changes!

Best Practices

Schema Design

Use descriptive schemas:

// ✅ Good
const schema = z.object({
  city: z.string().describe("City name (e.g., Tokyo, Paris)"),
  temperature: z.number().min(-50).max(60).describe("Temp in Celsius"),
});

// ❌ Bad
const schema = z.object({
  city: z.string(),
  temp: z.number(),
});

Theme Support

Always support both themes:

const { theme } = useWidget();

const bgColor = theme === "dark" ? "bg-gray-900" : "bg-white";
const textColor = theme === "dark" ? "text-white" : "text-gray-900";

Loading States

Always check isPending first:

const { props, isPending } = useWidget<MyProps>();

if (isPending) {
  return <LoadingSpinner />;
}

// Now safe to access props.field
return <div>{props.field}</div>;

Widget Focus

Keep widgets focused:

// ✅ Good: Single purpose
export const widgetMetadata: WidgetMetadata = {
  description: "Display weather for a city",
  props: z.object({ city: z.string() }),
};

// ❌ Bad: Too many responsibilities
export const widgetMetadata: WidgetMetadata = {
  description: "Weather, forecast, map, news, and more",
  props: z.object({
    /* many fields */
  }),
};

Error Handling

Handle errors gracefully:

const { callTool } = useWidget();

const fetchData = async () => {
  try {
    const result = await callTool("fetch-data", { id: "123" });
    if (result.isError) {
      console.error("Tool returned error");
    }
  } catch (error) {
    console.error("Tool call failed:", error);
  }
};

Configuration

Production Setup

Set base URL for production:

const server = new MCPServer({
  name: "my-app",
  version: "1.0.0",
  baseUrl: process.env.MCP_URL || "https://myserver.com",
});

Environment Variables

# Server URL
MCP_URL=https://myserver.com

# For static deployments
MCP_SERVER_URL=https://myserver.com/api
CSP_URLS=https://cdn.example.com,https://api.example.com

Variable usage:

• MCP_URL: Base URL for widget assets and CSP
• MCP_SERVER_URL: MCP server URL for tool calls (static deployments)
• CSP_URLS: Additional domains for Content Security Policy

Deployment

Deploy to mcp-use Cloud

# Login
npx mcp-use login

# Deploy
yarn deploy

Build for Production

# Build
yarn build

# Start
yarn start

Build process:

• Compiles TypeScript
• Bundles React widgets
• Optimizes assets
• Generates production HTML

Common Patterns

Data Fetching Widget

const DataWidget: React.FC = () => {
  const { props, isPending, callTool } = useWidget();

  if (isPending) {
    return <div>Loading...</div>;
  }

  const refresh = async () => {
    await callTool("fetch-data", { id: props.id });
  };

  return (
    <div>
      <h1>{props.title}</h1>
      <button onClick={refresh}>Refresh</button>
    </div>
  );
};

Stateful Widget

const CounterWidget: React.FC = () => {
  const { state, setState } = useWidget();

  const increment = async () => {
    await setState({
      count: (state?.count || 0) + 1,
    });
  };

  return (
    <div>
      <p>Count: {state?.count || 0}</p>
      <button onClick={increment}>+1</button>
    </div>
  );
};

Themed Widget

const ThemedWidget: React.FC = () => {
  const { theme } = useWidget();

  return (
    <div className={theme === "dark" ? "dark-theme" : "light-theme"}>
      Content
    </div>
  );
};

Troubleshooting

Widget Not Appearing

Problem: Widget file exists but tool doesn't appear

Solutions:

• Ensure .tsx extension
• Export widgetMetadata object
• Export default React component
• Check server logs for errors
• Verify widget name matches file/folder name

Props Not Received

Problem: Component receives empty props

Solutions:

• Check isPending first (props empty while pending)
• Use useWidget() hook (not React props)
• Verify widgetMetadata.props is valid Zod schema
• Check tool parameters match schema

CSP Errors

Problem: Widget loads but assets fail

Solutions:

• Set baseUrl in server config
• Add domains to CSP via metadata.csp (modern) or appsSdkMetadata['openai/widgetCSP'] (legacy)
• Use HTTPS for all resources
• Check browser console for CSP violations

CSP Errors in Production

Problem: Resources blocked by Content Security Policy in production

Solutions:

1. Check browser console for CSP violation messages
2. Add missing domains to your CSP configuration: metadata: {csp: {connectDomains: ['https://api.example.com'], // Add missing API domain resourceDomains: ['https://cdn.example.com'], // Add missing CDN domain}}
3. Use exact domains - avoid wildcards in production
4. Test in Inspector before deploying to catch CSP issues early
5. Environment variable alternative: Set CSP_URLS environment variable with comma-separated domains

Protocol Compatibility Issues

Problem: Widget works in ChatGPT but not MCP Apps clients (or vice versa)

Solutions:

• Use type: "mcpApps" for dual-protocol support (recommended)
• Check baseUrl is set correctly in server config
• Verify metadata format: Use metadata (camelCase) not appsSdkMetadata (snake_case) for dual-protocol
• Test in Inspector which supports both protocols

When to use each type:

• type: "mcpApps" - Maximum compatibility (recommended)
• type: "appsSdk" - ChatGPT only (use if you need ChatGPT-specific features not in spec)

Learn More

• Documentation: https://docs.mcp-use.com
• MCP Apps Standard: https://docs.mcp-use.com/typescript/server/mcp-apps (dual-protocol guide)
• Widget Guide: https://docs.mcp-use.com/typescript/server/ui-widgets
• Apps SDK Tutorial: https://docs.mcp-use.com/typescript/server/creating-apps-sdk-server
• Templates: https://docs.mcp-use.com/typescript/server/templates
• ChatGPT Apps Flow: https://docs.mcp-use.com/guides/chatgpt-apps-flow
• Inspector Debugging: https://docs.mcp-use.com/inspector/debugging-chatgpt-apps
• GitHub: https://github.com/mcp-use/mcp-use

Quick Reference

Commands:

• npx create-mcp-use-app my-app --template mcp-apps - Bootstrap
• yarn dev - Development with hot reload
• yarn build - Build for production
• yarn start - Run production server
• yarn deploy - Deploy to mcp-use Cloud

Widget structure:

• resources/widget-name.tsx - Single file widget
• resources/widget-name/widget.tsx - Folder-based widget entry
• public/ - Static assets

Widget metadata:

• description - Widget description
• props - Zod schema for input
• exposeAsTool - Auto-register as tool (default: true)
• metadata - Unified config (dual-protocol, recommended)
• metadata.csp - Content Security Policy configuration
• appsSdkMetadata - ChatGPT-specific overrides (optional)

CSP fields:

• connectDomains - APIs to call
• resourceDomains - Static assets to load
• frameDomains - Iframes to embed
• scriptDirectives - Script policies

useWidget hook:

• props - Widget input parameters
• isPending - Loading state flag
• state, setState - Persistent state
• callTool - Call other tools
• theme - Current theme (light/dark)
• displayMode, requestDisplayMode - Display control