Skip to content

Test Your App

frac apps expose a normal HTTP MCP endpoint, so you can test locally with MCP tooling and then connect the same endpoint to compatible clients.

Start your development server:

Terminal window
npm run dev

Your MCP server will be available at http://localhost:3000/mcp.

When you’re ready to test your AI App with an LLM, deploy the server or expose it through your own tunnel so the client can reach your /mcp endpoint.

ChatGPT uses the Apps SDK runtime and requires a public URL to connect to your app.

ChatGPT needs a public URL to access your AI App. Deploy your server or use your preferred tunnel provider, then keep the /mcp path on the URL.

  1. In ChatGPT, go to Profile → Apps → Create app
  2. Enter your public /mcp URL:
    https://your-domain.example/mcp
  3. Click Create
  1. Start a new conversation
  2. Select your app using the + button
  3. Prompt the model to trigger your tools

View changes in src/views/ appear instantly without reconnecting.

Server changes in src/server.ts require starting a new conversation to take effect.

Claude uses the MCP Apps runtime and requires a public URL to connect to your app.

Claude needs a public URL to access your AI App. Deploy your server or use your preferred tunnel provider, then keep the /mcp path on the URL.

  1. In Claude, go to Settings → Connectors → Add Custom Connector and configure your MCP server connection
  2. Enter your App name and public /mcp URL:
    https://your-domain.example/mcp
  3. Click on Add
  1. Start a new conversation in Claude
  2. Make sure your Connector is selected in the Connectors dropdown
  3. Test view interactions and tool responses

Desktop MCP clients like VSCode, Goose, and others can connect directly to your local server without needing a HTTP tunnel or a public URL.

Make sure your development server is running:

Terminal window
npm run dev

Your MCP server will be available at http://localhost:3000/mcp.

Connect your server to any desktop MCP client that supports the ext-apps specification, such as VSCode, Goose, Postman, or MCPJam.

  1. Configure your MCP client to connect to http://localhost:3000/mcp
  2. Start a conversation in your MCP client
  3. Test your app by invoking tools and interacting with the views