UI-TARS-desktop

This ByteDance repository hosts TARS, an open-source multimodal AI agent stack with two projects. Agent TARS is a general multimodal agent with a CLI and web UI that combines GUI-agent vision with DOM-based browser control and MCP tools, aiming to complete tasks the way a person would. UI-TARS Desktop is a native desktop application that lets a vision-language model, driven by the UI-TARS and Seed-VL series models, operate your local computer, a remote computer or a browser with mouse and keyboard.

It suits developers and researchers exploring computer-use agents: automating GUI tasks in desktop apps that have no API, multi-step browser tasks such as comparing and booking travel, or prototyping agents that see the screen. UI-TARS Desktop works on Windows and macOS and in the browser, with screenshot-based recognition and real-time status feedback; Agent TARS runs headful in a web UI or headless as a server and can mount additional MCP servers.

Getting started: run Agent TARS with `npx @agent-tars/cli@latest` (Node.js 22+ for a global install) and pass your model provider and API key, for example Volcengine or Anthropic. UI-TARS Desktop is a separate desktop app; follow its quick start to download it and configure a vision-language model endpoint.