Coding agents in Claude, Cursor and Codex are good at writing code. They're weaker at everything around it: a machine to run the app on, a live preview, a database, file storage, secrets and a public URL at the end. Usually you wire that up yourself.
This post shows how to hand all of that to the agent through one remote MCP server. I build Mythex, an AI app builder from Dubai, and this is how its MCP server works. The pattern applies to any MCP server that exposes a sandbox.
Once connected, the client can:
bash
set_run_config, start_preview, get_preview)
Every tool result includes a workspace_url, so you can watch the live preview in the browser while the agent works in your editor.
https://mcp.mythex.ai/mcp
Add the server with only the URL (.cursor/mcp.json):
{
"mcpServers": {
"mythex": {
"url": "https://mcp.mythex.ai/mcp"
}
}
}
Toggle it on, click Connect when Cursor prompts, and approve in the browser. There are no API keys to paste: it's OAuth with refresh tokens.
Create a new Mythex project called "booking-demo": a React + TypeScript app where clients book 30-minute slots. Store bookings in Postgres, start the preview, check it in the browser and send me a screenshot. When it works, publish it.
The agent creates the project, writes the code in the sandbox, adds the database, starts the preview, checks it with a browser run, and publishes. You get back a preview URL, screenshots and a public URL.
upgrade_url instead of running.
If you'd rather build in a chat UI, the same projects open in the Mythex web app. It also works in Arabic and 68 other languages.