Ship a ChatGPT App in 2 Commands (April 2026)

With sunpeak, you can go from zero to a working ChatGPT App in two commands. This guide walks through the full process: scaffolding your project, testing locally without a ChatGPT account, connecting to the real ChatGPT, and deploying to production.

TL;DR: Run npx sunpeak new to scaffold a complete ChatGPT App project, then pnpm dev to start building. The sunpeak Inspector replicates the ChatGPT runtime at localhost:3000. When you’re ready to test in real ChatGPT, expose your server with ngrok and add the URL in ChatGPT’s developer mode settings.

Command 1: Create your project

Scaffold a new ChatGPT App:

npx sunpeak new

This creates a project with everything you need:

• src/resources/ for your UI components (React views that render inside ChatGPT)
• src/tools/ for tool handlers (the actions ChatGPT’s model can invoke)
• src/server.ts for server configuration and optional auth
• tests/ for Playwright test files and simulation fixtures
• A pre-configured MCP server, TypeScript, Vite, and hot module replacement

Command 2: Start developing

cd sunpeak-app && pnpm dev

This starts two things:

1. Your MCP server on port 8000, which handles tool calls and serves your App resources
2. The sunpeak Inspector on port 3000, a local replica of the ChatGPT runtime

Open localhost:3000 and you’ll see your App running in a simulated ChatGPT environment. Every code change hot-reloads instantly, no manual refresh needed.

The Inspector also supports Claude, so you can switch hosts from the sidebar dropdown and see how your App renders in both environments from one codebase. This is one of the bigger time-savers because your MCP App works across hosts without code changes.

Testing your App locally

Before connecting to the real ChatGPT, you can thoroughly test your App right from the Inspector.

sunpeak’s testing framework runs against the local Inspector, which means your tests replicate real ChatGPT behavior without a paid account or API credits. Simulation files in tests/simulations/ provide deterministic mock data so your tests produce the same result every time.

pnpm test           # run unit + e2e tests
pnpm test:visual    # visual regression testing
pnpm test:eval      # multi-model LLM evals

You can also add these tests to your CI/CD pipeline so every PR gets validated automatically.

Connecting to the real ChatGPT

When you’re ready to see your App in ChatGPT, you need to expose your local MCP server with a public URL. ngrok works well for this (a free account is enough):

ngrok http 8000

Copy the Forwarding URL from ngrok’s output, then add it in ChatGPT:

1. Go to Settings > Apps > Advanced settings
2. Enable Developer mode (requires a Plus, Team, or Enterprise account)
3. Click Create app and enter your ngrok URL with the /mcp path, for example: https://abc123.ngrok-free.dev/mcp

Start a new conversation and invoke one of your tools. ChatGPT will call your MCP server, and your App UI will render right in the chat.

If you don’t have a paid ChatGPT account, the Inspector fully covers local development and testing. You can build and ship your App without connecting to ChatGPT until you’re ready to deploy.

Deploying to production

For production, you don’t need ngrok. Build and start your server directly:

pnpm build && pnpm start

pnpm build compiles your resources into optimized bundles and your tools into production-ready modules. pnpm start launches the MCP server on port 8000 (configurable via the PORT environment variable).

Deploy to any Node.js host, or use createMcpHandler for custom Express setups and createHandler for serverless platforms. The deployment guide covers production configuration, scaling, health checks, and publishing to OpenAI’s platform.

What you get out of the box

The scaffolded project includes a sample carousel App that demonstrates tools, resources, and display modes. From there you can build any UI that ChatGPT Apps support: dashboards, forms, data visualizations, multi-step workflows, and more.

Because sunpeak Apps are MCP Apps, they work in any MCP Apps-compatible host. The same App you build for ChatGPT also runs in Claude as a connector, with no code changes needed. Write once, test across hosts with the Inspector, and ship to production with pnpm build && pnpm start.

Get started with the quickstart guide, explore the ChatGPT App framework, or check out the full tutorial for a deeper walkthrough.