UI-TARS Desktop

Two complementary projects share the UI-TARS-desktop repository: UI-TARS Desktop, a native application that turns natural language into GUI actions on your local computer and browser, and Agent TARS, a general multimodal agent framework with CLI and web interfaces plus MCP tool integration. Both are powered by vision-language models that ground their actions in what is on screen, either the open UI-TARS 1.5 and 1.6 models ByteDance publishes on Hugging Face for local deployment or hosted options such as the Doubao vision series on VolcEngine and Claude. Browser control works through DOM inspection or pure visual grounding, and the stack supports local and remote computer operators, streaming tool output, and runtime statistics. Getting started takes Node.js 22 and npx @agent-tars/cli@latest, or a packaged desktop build. The code is Apache-2.0 licensed with about 38,000 GitHub stars, and the v0.3.0 release landed in November 2025, keeping this among the leading open computer-use agent stacks alongside the models that share its name.