Project Structure

app/                        Next.js pages
components/
├── examples/               Example widgets (POI Map)
├── workbench/              Workbench UI components
└── ui/                     Shared UI components
lib/
├── sdk/                    SDK exports for production
├── workbench/              React hooks + dev environment
└── export/                 Production bundler
server/                     MCP server (if included)

Building Your Widget

1. Create a component

// components/my-widget/index.tsx
import {
  useToolInput,
  useCallTool,
  useTheme,
  useCapabilities,
  useUpdateModelContext,
  useWidgetState,
} from "@/lib/sdk";

export function MyWidget() {
  const input = useToolInput<{ query: string }>();
  const callTool = useCallTool();
  const theme = useTheme();
  const capabilities = useCapabilities();
  const updateModelContext = useUpdateModelContext();
  const [widgetState, setWidgetState] = useWidgetState();

  const handleSearch = async () => {
    const result = await callTool("search", { query: input.query });
    console.log(result.structuredContent);
  };

  return (
    <div
      data-theme={theme}
      className={theme === "dark" ? "dark bg-background text-foreground" : "bg-background text-foreground"}
      style={{ colorScheme: theme }}
    >
      <p>Query: {input.query}</p>
      <button onClick={handleSearch}>Search</button>

      {/* Platform-specific features */}
      {capabilities.modelContext && (
        <button
          onClick={() =>
            updateModelContext({ structuredContent: { query: input.query } })
          }
        >
          Update model context (host-dependent)
        </button>
      )}
      {capabilities.widgetState && (
        <button
          onClick={() =>
            setWidgetState({
              ...(widgetState ?? {}),
              savedAt: Date.now(),
            })
          }
        >
          Save widget state (ChatGPT extensions)
        </button>
      )}
    </div>
  );
}

2. Register in the workbench

Add your component to lib/workbench/component-registry.tsx.

3. Add mock data

Configure mock tool responses in lib/workbench/mock-config/.

React Hooks Reference

Full documentation: lib/workbench/README.md

Universal Hooks (recommended)

These hooks work identically across MCP hosts (including ChatGPT):

Platform Detection (when needed)

Host-Dependent / Extensions (advanced)

These hooks only work on specific platforms. Check availability first:

useWidgetState() is not a standard MCP Apps persistence primitive. For portable MCP Apps, use app-managed persistence such as localStorage or server-backed tools.

Tool result metadata (_meta) is available via readToolResponseMetadata() when the host exposes window.openai.toolResponseMetadata.

Platform-Specific Features

MCP App Studio is MCP-first: prefer the MCP Apps bridge (ui/*) and feature-detect optional ChatGPT extensions (window.openai) when needed.

Use useCapabilities() or useFeature() to conditionally enable features.

Modal Guidance

• Prefer local/in-widget modals for cross-host compatibility.
• Use window.openai.requestModal() only when you specifically need a ChatGPT-hosted modal template.
• Always feature-detect and provide a fallback:

if (typeof window !== "undefined" && window.openai?.requestModal) {
  await window.openai.requestModal({ title: "Details", params: { id } });
} else {
  // Fallback: local modal state or route navigation
}

Checkout Guidance (ChatGPT beta extension)

• window.openai.requestCheckout(...) is currently a ChatGPT private beta extension.
• Treat checkout as optional: feature-detect and provide a fallback (for example, external checkout).
• If you expose a companion deep link in ChatGPT menus, use window.openai.setOpenInAppUrl({ href }).

import { requestCheckout, setOpenInAppUrl } from "@/lib/sdk";

setOpenInAppUrl("https://your-app.com/orders/123");

const outcome = await requestCheckout(
  { id: "checkout_123", payment_mode: "test" },
  () => window.open("https://your-app.com/checkout/123", "_blank"),
);

if (outcome.mode === "fallback") {
  // Non-ChatGPT host or checkout beta unavailable.
}

Exporting for Production

npm run export

Defaults for --entry and --export-name are read from mcp-app-studio.config.json (written by the CLI when you scaffold a project). You can override them via flags.

Generates:

export/
├── widget/
│   └── index.html      Self-contained widget
├── manifest.json       App manifest
└── README.md           Deployment instructions

The exported widget uses the mcp-app-studio SDK which automatically detects the host platform and uses the appropriate bridge.

MCP Metadata Defaults (Export)

• ui.* metadata is canonical in exported server code.
• Legacy metadata keys from older ChatGPT Apps integrations are not supported in exported MCP metadata.
• Visibility defaults to host default (["model","app"]) when no visibility keys are set.
• Widget resources are emitted with MCP Apps MIME type text/html;profile=mcp-app.

Widget resource CSP must use MCP-standard keys:

ui: {
  csp: {
    connectDomains: ["https://api.example.com"],
    resourceDomains: ["https://cdn.example.com"],
    frameDomains: ["https://www.youtube.com"],
    baseUriDomains: ["https://cdn.example.com"],
  },
}

Deploying

Widget

Deploy export/widget/ to any static host:

# Vercel
cd export/widget && vercel deploy

# Netlify
netlify deploy --dir=export/widget

# Or any static host (S3, Cloudflare Pages, etc.)

MCP Server

If you have a server/ directory:

cd server
npm run build
# Deploy to Vercel, Railway, Fly.io, etc.

Register with Platform

For ChatGPT:

1. Update manifest.json with your deployed widget URL
2. Go to ChatGPT Apps dashboard
3. Create a new app and connect your MCP server
4. Test in a new ChatGPT conversation

For Claude Desktop:

1. Configure your MCP server in Claude Desktop settings
2. The widget will render when tools with UI are invoked

Configuration

SDK Guide (Optional)

The workbench includes an AI-powered SDK guide. To enable:

cp .env.example .env.local
# then set:
# OPENAI_API_KEY="your-key"

MCP Server CORS

For production, restrict CORS to your widget domain:

cp server/.env.example server/.env
# then set:
# CORS_ORIGIN=https://your-widget-domain.com

Dark Mode

Exported widgets inherit host theme and token variables. Follow the framework-agnostic contract in lib/workbench/THEMING_CONTRACT.md.

At minimum, support data-theme / .dark and semantic tokens:

.my-element {
  background: var(--background);
  color: var(--foreground);
  border-color: var(--border);
}

Learn More

• MCP App Studio — CLI and SDK documentation
• MCP Specification — Model Context Protocol
• ChatGPT MCP Apps — ChatGPT as an MCP host