bytedance/UI-TARS-desktop

An open-source multimodal AI agent stack that drives computers and browsers through natural language, shipping both a desktop GUI agent and a CLI-based agent.

What it is

UI-TARS-desktop is a repository that ships two related projects under one multimodal AI agent stack: Agent TARS and UI-TARS Desktop. Agent TARS is a general multimodal agent that runs from a terminal CLI or Web UI and integrates with real-world tools over MCP, combining GUI-agent and vision capabilities. UI-TARS Desktop is a native desktop application that operates the local computer and browser as a GUI agent, driven by the UI-TARS and Seed-1.5-VL/1.6 models. It is aimed at developers who want agents that complete tasks by seeing and controlling a real screen rather than only calling APIs.

Key features

Natural-language control of a computer powered by a Vision-Language Model, with screenshot capture and visual recognition, and precise mouse and keyboard control.
Local and remote operators for both computer and browser use; the remote computer and browser operators are described as free and require no configuration.
Agent TARS ships a one-click CLI supporting both a headful Web UI and headless server execution.
Hybrid browser agent that can control browsers via GUI Agent (visual grounding), DOM, or a hybrid strategy.
MCP-based kernel that also supports mounting external MCP servers to connect to real-world tools.
Cross-platform desktop support (Windows, macOS, Browser) with real-time feedback and status display, running fully local processing.

Tech stack

TypeScript, organized as a pnpm monorepo (pnpm@9.10.0) using Turbo (turbo ^2.4.4).
Electron desktop application (Electron Forge, electron-vite, electron-builder configs present).
Node.js >=20.x at the monorepo level; the Agent TARS CLI requires Node.js >=22.
Testing with Vitest and Playwright; linting/formatting via ESLint and Prettier; Changesets for release management.
Distributed on npm as @agent-tars/cli; the UI-TARS SDK is provided as a cross-platform toolkit for building GUI automation agents.

When to reach for it

You need an agent that operates real GUI applications and browsers via vision rather than through structured APIs.
You want a desktop app that can control your local machine with natural-language instructions.
You want to run browser or computer automation tasks against a remote operator without local setup.
You are building GUI automation on top of the UI-TARS model family and want a CLI, Web UI, or SDK entry point.

When not to reach for it

Your automation targets have stable APIs or DOM selectors; a vision-driven GUI agent adds cost and latency you may not need.
You need a small dependency-light library; this is a large Electron monorepo plus model integration.
You cannot supply a capable multimodal model endpoint (for example Volcengine, Anthropic, or a self-hosted UI-TARS deployment), which the agent depends on.
You require a mature, low-churn tool; the project is young and moves quickly.

Maturity signal

The repository has over 38,000 stars accumulated since its creation in January 2025, and its most recent push was July 2026, making it a fast-moving, actively maintained project backed by ByteDance and licensed under Apache-2.0. The large open-issue count (over 400) reflects heavy real-world use and rapid iteration rather than abandonment, and the frequent versioned releases noted in the README (UI-TARS Desktop v0.1.0 through v0.2.0, Agent TARS CLI v0.3.0) reinforce an early-but-active trajectory. Treat APIs and features as still evolving.

Alternatives

browser-use — use it instead when you only need browser automation driven by an LLM and do not need native desktop control.
Midscene — use it instead when you want in-browser GUI automation; the README itself links it as the browser-focused counterpart.
Self-Operating Computer — use it instead when you want a lightweight, model-agnostic script for computer control rather than a packaged desktop app.
Anthropic Claude computer use — use it instead when you are already in the Claude ecosystem and want a hosted computer-use capability.

Notes

The repository is a stack rather than a single app: Agent TARS (CLI/Web UI, MCP-centric) and UI-TARS Desktop (Electron desktop GUI agent) are shipped together but serve different usage modes. Agent TARS's kernel is built directly on MCP, and it uses a protocol-driven Event Stream to power both its context engineering and its Agent UI. The v0.3.0 release adds an AIO Agent Sandbox as an isolated all-in-one tool-execution environment.

Wiki: bytedance/UI-TARS-desktop