Illustration of what UI-TARS-desktop does: Open-source agent stack that drives browsers and desktops with vision models and MCP

bytedance /

UI-TARS-desktop

Open-source agent stack that drives browsers and desktops with vision models and MCP

Open on GitHub (opens in a new tab) Website (opens in a new tab)

TARS is a multimodal agent stack. It lets an AI model look at a screen, decide what to click or type, and carry out a task such as booking a hotel or drawing a chart. The repo ships two projects: Agent TARS and UI-TARS Desktop.

Agent TARS runs from a command line or a Web UI. It can control a browser by reading the page structure (the DOM), by looking at screenshots (a GUI agent), or by mixing both. Its core is built on MCP, so you can attach MCP servers to give it extra tools. It works with model providers such as Volcengine and Anthropic.

UI-TARS Desktop is a native app that runs a GUI agent on your own computer, powered by the UI-TARS model and Seed-1.5-VL/1.6 models. It has local and remote operators for computers and browsers. The README points to Midscene for browser-only use.

What it does

  • CLI with Web UI or headless server

    Agent TARS runs with a visual Web UI or as a headless server, started from one CLI command.

  • Hybrid browser control

    The agent can operate a browser by visual grounding, by DOM, or by a mix of the two.

  • MCP at the core

    The kernel is built on MCP, and you can mount extra MCP servers to connect real-world tools.

  • Event stream

    A protocol-driven event stream feeds context engineering and the agent UI. A viewer helps trace data flow when debugging.

  • Native desktop operators

    UI-TARS Desktop offers local and remote computer and browser operators built on the UI-TARS model.

What you could build

  • Ask an agent to search travel sites and help book flights or hotels from a plain-language request.
  • Generate charts or reports by attaching extra MCP servers to Agent TARS.
  • Automate clicks and typing in desktop apps with the UI-TARS Desktop local operator.
  • Run Agent TARS as a headless server to embed a browser agent in your own product.
  • Compare DOM-based and screenshot-based browser control on the same task.

Get started

You need Node.js 22 or newer for global install. Run Agent TARS with npx or install it globally, then start it with your chosen model provider, model name and API key. The README links a fuller quick start guide.

npx @agent-tars/cli@latest
npm install @agent-tars/cli@latest -g
agent-tars --provider anthropic --model claude-3-7-sonnet-latest --apiKey your-api-key

Commands are copied from the project's README. Read it on GitHub for the full setup.

More in Browser & computer use

See all
browser-use Python library and CLI that lets an LLM drive a web browser to finish tasks 118.9k1.9k Agent-Reach Setup layer that lets AI agents read the web, YouTube, GitHub, Reddit and more 101.1k13.5k playwright One API to test and automate Chromium, Firefox and WebKit, for scripts and agents 98.3k1.3k Scrapling Python web scraping framework, from one request to large crawls 89.2k4.2k crawl4ai Python crawler that turns web pages into clean Markdown or JSON for LLMs 86.1k1.4k agent-browser A command-line tool that lets AI agents drive a Chrome browser 44.4k994 playwright-mcp Browser automation server that lets AI assistants drive web pages through Playwright 38.5k766 CloakBrowser Patched Chromium that works as a drop-in Playwright replacement for bot-protected sites 32.4k566 page-agent JavaScript library that lets people control a web page with plain-language commands 29.6k242 ego-lite A macOS browser where AI agents work in their own Spaces while you keep browsing 18.8k2.1k

    ↑↓ to move, Enter to open, Esc to close