UI-TARS Desktop

What is UI-TARS Desktop?

UI-TARS Desktop is an open-source multimodal agent released by ByteDance in 2025. It is built around the UI-TARS vision-language model family, which was trained to understand screenshots of interfaces the way a person sees them, and it packages that model into a desktop application that can actually operate your computer.

The agent works by taking a screenshot of your screen, deciding what to do next, and then performing the action with real mouse and keyboard input. That means it does not need special APIs or hooks inside applications; if a human can see it and click it, the agent can work with it, which is why it handles both desktop apps and browsers equally well.

This style of tool is often called computer use or GUI agent technology, and UI-TARS Desktop is notable for being open source and for supporting both local models and cloud model APIs, so you can keep your automation private or get stronger performance from hosted models.

Key Features

• Full desktop control: clicking, typing, dragging and keyboard shortcuts
• Browser automation through the same visual interface
• Task planning from natural-language prompts
• Screen understanding via the UI-TARS vision-language model
• Agent memory so it can remember context from earlier tasks
• Runs with local models (Ollama and others) or cloud model APIs
• Cross-platform desktop app for Windows, macOS and Linux
• Open source, so you can audit, extend and self-host it

How to get started

Download the installer for your operating system from the GitHub releases page and install it like any desktop app. On first launch you choose how the agent gets its intelligence: point it at a local model, or configure a cloud model API key.

Once it is running, you simply describe a task in natural language, such as 'open the mail app and compose a message to Sam'. The agent takes over, shows you its plan and reasoning, and carries out the steps while you watch.

• Download the installer from GitHub releases
• Install and launch the desktop app
• Choose a local model or configure a cloud model API
• Type a task description in plain language
• Watch the agent execute and review the results

Use cases

• Automating repetitive desktop chores like file organization and app setup
• Browser research and data collection across many pages
• Testing applications the way a real human user would
• Filling long forms that span multiple apps
• Accessibility support for people who cannot use a mouse easily
• Prototyping personal automations without writing scripts

Pricing and licensing

UI-TARS Desktop is free and open source, so the software itself costs nothing and you can run it on your own hardware. If you use a local model, inference is free too, though you will want a reasonably powerful GPU for smooth performance.

If you prefer stronger results from larger hosted models, you connect a cloud model provider and pay for the tokens you use at that provider's normal rates. See the official GitHub repository for the current release details and model recommendations.

Pros and cons

The standout strengths are its multimodal nature and its openness. Because it works from screenshots, it handles anything visible on screen, and because it is open source, you can run it privately and extend it however you like. For many repetitive tasks it is genuinely faster than writing a bespoke script.

The main limitations are reliability and speed. GUI agents occasionally misclick or misunderstand unusual interfaces, they are slower than scripted automation, and running capable models locally needs decent hardware. For critical production workflows you may still prefer deterministic tools.

Alternatives

• OpenAI Operator: hosted agent that browses the web for you
• Anthropic Computer Use: API-based agent that controls a computer via screenshots
• Stagehand: natural-language browser automation built on Playwright
• SikuliX: scripted visual automation that matches on-screen images
• Apple Shortcuts: built-in automation for macOS and iOS