bytedance/UI-TARS-desktop ? reverse-engineered prompt

Build me an AI assistant app that can actually use a computer for me, not just chat. I want a desktop app where I can type a normal request like book a flight, compare hotels, open a website, fill forms, or make a chart, and the agent can look at the screen, click, type, use the browser, and explain what it is doing.

It should also have a simple web interface and a command line option for people who prefer that. Please support both local computer control and remote computer or browser control. Add a clear activity view so I can see each step, tool call, screenshot, result, and any errors. Let users connect external tools through MCP style integrations.

Keep setup friendly, with an env example, model configuration, and a quick start path. Make the UI feel polished and safe, with confirmations before risky actions. Look up current docs online if you need to.