App Template

CI

The AI-agent-first MCP App template. Built so coding agents (Claude Code, Codex, Cursor, ...) can modify, test, and go as far as possible autonomously - without human in the loop.

Works with any MCP Apps host: Claude, ChatGPT, VS Code, Goose, and many more to come.

What are MCP Apps?

credit: image generated using Nano Banana through Ormind

MCP Apps are an extension to the Model Context Protocol that let MCP servers return interactive UIs -- React widgets rendered directly inside AI hosts like ChatGPT, Claude, VS Code, and so on... Instead of tools returning plain text, they return rich, interactive experiences displayed in sandboxed iframes within the conversation.

This is a platform shift. AI chatbots are becoming app platforms -- ChatGPT alone has 800M+ weekly users, and apps now run inside that distribution channel. One MCP server works across every host.

For more context, see the MCP Apps blog post and the protocol specification (SEP-1865).

Why This Template?

Most templates assume a human developer. This one is designed for AI agents to work as much as possible autonomously:

1. Orthogonal Test Suite

~450 tests verify infrastructure (MCP Apps compliance, protocol format, accessibility) plus 12 browser tests that render widgets in real Chromium. Tests are auto-discovered per widget (input validation, build output, browser rendering) — you never write them, they appear automatically when you add a widget. Modify widgets, change data, add features — tests still pass. Automated grading generates reports with actionable FIX: hints to steer coding agents.

2. Hierarchical Documentation

AGENTS.md for quick onboarding → docs/README.md for a step-by-step building guide → deep docs covering MCP best practices, widget patterns, and complete SDK reference (llms-full.txt).

3. Automated Visual Testing

AI agents can test widgets and capture screenshots - no API key required:

pnpm run ui-test --tool show_carousel  # Renders tool, saves screenshot

Agents read /tmp/ui-test/screenshot.png to verify their changes work.

4. Zero-Config App Tester

Local app tester for manual testing - works instantly without API key via Puter.js. Add an OpenAI key for AI-in-the-loop testing.

5. Working Examples

12 production-ready widgets demonstrating state management, theming, 3D visualization, drag-and-drop, real-time monitoring, and more.

Demos

Claude Code autonomously building an app end-to-end from the template

https://github.com/user-attachments/assets/407393de-c9d8-4da3-b541-54b8b9a2d7dc

Showcase of widgets within ChatGPT

https://github.com/user-attachments/assets/f5877544-0dce-4c31-979e-50b5533f9a16

Quick Start

# Clone and setup
git clone https://github.com/sebderhy/mcp-app-template.git my-app
cd my-app
./setup.sh        # Installs deps, builds widgets, runs tests

# Start the server
pnpm run server   # Starts on http://localhost:8000

pnpm install
cd server && python3 -m venv .venv && .venv/bin/pip install -e ".[dev]" && cd ..
pnpm run setup:test   # Install Playwright
pnpm run build
pnpm run test

Open the app tester: http://localhost:8000/assets/apptester.html

That's it! The app tester uses Puter.js for free AI - no API key required for manual testing.

Creating Your Own App

The template includes 12 example widgets. To start fresh with only the widgets you need:

# Remove all example widgets (start from scratch)
./create_new_app.sh

# Keep specific widgets as starting points
./create_new_app.sh --keep carousel
./create_new_app.sh --keep carousel,todo

# Set your app name
./create_new_app.sh --keep boilerplate --name my-app

The script:

• Deletes widget directories from src/ and server/widgets/
• Removes unused dependencies (three.js, framer-motion, chart.js, etc.)
• Updates the lockfile and runs build + tests to verify

After running, add your own widgets following the Adding Your Own Widget section.

Local App Tester

The app tester lets you test widgets without deploying to a real MCP host. It works in two modes:

Zero-Config Mode (No API Key)

Just start the server and open the app tester - it works immediately:

pnpm run build
pnpm run server
# Open http://localhost:8000/assets/apptester.html

The app tester automatically detects when no API key is configured and uses Puter.js - a free, browser-based AI service. This is perfect for:

• Quick prototyping sessions
• Demos and presentations
• Manual testing without credentials

Note: Puter.js is only used in the local app tester. In production, your app connects to real MCP hosts (Claude, ChatGPT, etc.) which provide the AI.

Full Mode (With OpenAI API Key)

For production-quality testing with your preferred model:

1. Add your OpenAI API key to .env:

   OPENAI_API_KEY=sk-...
   

2. Optionally configure the model in server/simulator_config.json:

   {
     "model": "gpt-4o-mini",
     "mcp_server_url": "http://localhost:8000/mcp"
   }
   

3. Restart the server and open the app tester

The app tester shows which mode is active in the header.

App Tester Features

Try prompts like "Show me a carousel of restaurants" or "Display a dashboard"

Test with Real MCP Hosts

Claude (Desktop or Web)

1. Expose your local server via a tunnel:

   npx cloudflared tunnel --url http://localhost:8000  # or: ngrok http 8000 --host-header=rewrite
   

2. Add the tunnel URL as a custom connector in Claude settings
3. Ask: "Show me the boilerplate widget"

ChatGPT

1. Enable Developer Mode
2. Expose your local server via a tunnel:

   npx cloudflared tunnel --url http://localhost:8000  # or: ngrok http 8000 --host-header=rewrite
   

3. Add connector in ChatGPT Settings → Connectors (tunnel URL + /mcp)
4. Ask: "Show me the boilerplate widget"

Note: The server disables DNS rebinding protection by default to support tunneling with random hostnames. For production deployments with a fixed domain, re-enable it in server/main.py (TransportSecuritySettings).

MCP Apps Basic Host

For testing with the reference implementation:

git clone https://github.com/modelcontextprotocol/ext-apps
cd ext-apps/examples/basic-host
npm install && SERVERS='["http://localhost:8000"]' npm start

How It Works

User Prompt → MCP Host → MCP Tool Call → Python Server → Widget renders in host

• Widgets (src/) - React/TypeScript UIs that render inside MCP hosts
• Server (server/main.py) - Python MCP server that handles tool calls
• App Tester (internal/apptester/) - Local development UI with Puter.js fallback
• Assets (assets/) - Built widget bundles (generated by pnpm run build)

Included Examples

These widgets are adapted from the OpenAI Apps SDK Examples and MCP Apps Examples.

Try them: "Show me the carousel", "Show me the dashboard", etc.

Adding Your Own Widget

1. Create src/my-widget/index.tsx and src/my-widget/App.tsx (entry point must target my-widget-root)
2. Create server/widgets/my_widget.py with WIDGET, INPUT_MODEL, and handle() exports (auto-discovered)
3. Run pnpm run build && pnpm run test
4. Test in the app tester: http://localhost:8000/assets/apptester.html

When customizing this template for your own app, follow the guidelines in docs/mcp-development-guidelines.md for tool naming conventions, descriptions, and error handling best practices.

Testing

pnpm run test          # Server + UI tests (fast, run after every change)
pnpm run test:all      # All tests including browser (requires Playwright)
pnpm run test:server   # Server tests only
pnpm run test:ui       # UI unit tests only
pnpm run test:browser  # Browser compliance tests only (requires Playwright)

Tests are orthogonal to your app. They verify:

• MCP Apps protocol compliance
• SDK format requirements
• Build output structure
• React hooks work correctly
• Widgets render without errors in real browsers

They don't verify your specific widgets, data, or business logic. Modify anything - tests still pass.

Browser Compliance Tests

Browser tests run each widget in a real Chromium browser to verify:

• No JavaScript errors when rendering
• Widget renders visible content
• Works in both light and dark themes
• No unhandled promise rejections
• Images have alt text (accessibility)
• No duplicate HTML IDs
• All callTool invocations from widgets are callable on the server
• Text contrast meets WCAG guidelines (warnings only)
• Interactive elements are keyboard accessible (warnings only)

Setup (one-time):

pnpm run setup:test           # Install Playwright browsers
npx playwright install-deps   # Install system dependencies (may need sudo)

Run:

pnpm run test:browser

Tests skip gracefully if browser dependencies aren't installed, so they won't break CI pipelines that lack browser support.

Automated Grading

The test suite includes automated grading against best practices. After running tests, check these reports:

MCP Best Practices (server/tests/mcp_best_practices_report.txt): Grades against MCP server guidelines - tool naming, descriptions, error handling.

MCP App Guidelines (server/tests/mcp_app_guidelines_report.txt): Grades against app design guidance - Know/Do/Show value, model-friendly outputs, ecosystem fit.

Output Quality (server/tests/output_quality_report.txt): Grades tool output quality - response size limits, schema stability, null handling, ID consistency, boundary value handling.

Example report:

============================================================
MCP APP GUIDELINES GRADE REPORT
============================================================

1. Value Proposition: 100.0%
  ✓ Clear Know/Do/Show value: 100%

2. Model-Friendly Outputs: 94.3%
  ✓ List items have IDs: 100%
  ✗ Complex outputs have summary: 88%
      FIX: Add 'summary' field: {"summary": "3 items", "items": [...]}

OVERALL SCORE: 95.5% (Grade: A)
============================================================

When a check fails, the report includes a FIX: hint explaining exactly how to resolve it - useful for AI agents improving the server.

Current limitation: The grading tests use heuristic checks (keyword matching, regex patterns, length thresholds) rather than LLM-based evaluation. This means some checks can produce false positives/negatives for edge cases. Upgrading to LLM-based test evaluation is on the roadmap — contributions welcome!

Automated UI Testing (for AI Agents)

AI coding agents can visually test widgets using the built-in UI test tool. This enables AI agents to verify their changes work correctly by examining screenshots.

Setup (One-time)

pnpm run setup:test   # Install Playwright browsers (~150MB)

Two Testing Modes

1. Direct Mode (no API key) - Test specific tools directly:

pnpm run ui-test --tool show_carousel           # Test a tool
pnpm run ui-test --tool show_carousel --theme dark  # Test in dark mode

This is the recommended mode for AI coding agents.

2. AI-in-the-loop Mode (requires OpenAI API key) - Full app tester with AI:

pnpm run ui-test "Show me the carousel widget"

The AI receives your prompt, decides which widget to show, and the tool captures the result.

Output

Both modes save artifacts to /tmp/ui-test/:

• screenshot.png - Visual capture of the rendered widget
• dom.json - Structured data about what rendered
• console.log - Browser console output

AI agents can read the screenshot to verify widgets rendered correctly:

Read tool → /tmp/ui-test/screenshot.png

Example Workflow (Coding Agents)

User: "Add a new stats card to the dashboard"

Coding Agent (Claude Code, Codex, Cursor, etc.):
1. Modifies src/dashboard/index.tsx
2. Runs pnpm run build
3. Runs pnpm run ui-test --tool show_dashboard
4. Reads /tmp/ui-test/screenshot.png
5. Verifies the new card appears correctly

Key APIs

// Read data from server
const { title, items } = useWidgetProps({ title: "", items: [] });

// Persist state across tool calls
const [state, setState] = useWidgetState({ count: 0 });

// Respond to theme/display changes
const theme = useTheme();           // "light" | "dark"
const mode = useDisplayMode();      // "inline" | "fullscreen" | "pip"

// Call tools from widget
await window.openai.callTool("my_tool", { arg: "value" });

Deployment

Local Development

No configuration needed - just run:

pnpm run build
pnpm run server

VPS / Remote Server

When running on a VPS or remote server, set the BASE_URL environment variable so widgets load correctly:

# Build once (no BASE_URL needed - uses placeholder)
pnpm run build

# Set BASE_URL when starting the server
BASE_URL=http://YOUR_SERVER_IP:8000/assets pnpm run server

# Or use a .env file at the repo root
cp .env.example .env
# Edit .env and set BASE_URL=http://YOUR_SERVER_IP:8000/assets

The server replaces the __BASE_URL__ placeholder at runtime, so you can rebuild once and deploy anywhere by just changing the environment variable.

Production

For production with a domain and HTTPS:

BASE_URL=https://your-domain.com/assets pnpm run server

Deploy the Python server to any platform (Fly.io, Render, Railway, Cloud Run, etc.). Requirements: HTTPS, /mcp endpoint, SSE streaming support.

Roadmap

Planned features and improvements:

• Authentication support - Apps that require user login and secure sessions
• Agentic commerce - Enable purchases and transactions within AI conversations
• LLM-based test evaluation - Use LLMs to determine test pass/fail for complex verifications
• Interactive browser tests - Browser tests that simulate user interactions with widgets
• Multi-step agentic workflows - Support agent loops chaining multiple tool calls

Have ideas? Open an issue or PR.

Resources

• MCP Apps Protocol - Official MCP Apps specification (see also local copies: docs/mcp-apps-docs.md, docs/mcp-apps-specs.mdx)
• MCP Protocol - Model Context Protocol
• OpenAI Apps SDK - OpenAI's implementation
• Apps SDK UI Components
• Puter.js - Powers the app tester's zero-config mode (free for dev/testing)

Acknowledgments

Based on OpenAI's Apps SDK Examples and the MCP Apps Examples.

The qr, system-monitor, scenario-modeler, and map widgets are ported from modelcontextprotocol/ext-apps (MIT License).

License

MIT