UI-TARS-desktop and Agent TARS: A Multimodal Agent Stack That Puts GUI Control in the Terminal and Browser
GitHub describes it as The Open-Source Multimodal AI Agent Stack: Connecting Cutting-Edge AI Models and Agent Infra. The repository metadata lists TypeScript as its primary language. The metadata lists the Apache-2.0 license. This article stays within the project description and details documented in the GitHub repository README.
At a glance
- What is it?
- Bytedance's UI-TARS-desktop repository now hosts two projects: a desktop GUI agent and a general-purpose Agent TARS stack with CLI and Web UI. This review covers what each does, how they are built, and where the rough edges are.
- Who is it for?
- Adopt this stack if you need a GUI agent that can operate a real computer or browser, or if you want a CLI-based multimodal agent that can call MCP tools and run in an isolated sandbox. Do not adopt it if you need a production-hardened agent with a stable API and mature documentation, because the repository is moving fast, with beta releases and v0.3.0 (November 2025), which introduced significant changes like streaming and a sandbox, followed by v0.3.1 in October 2026.
- Can I use it commercially?
- Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- No. The owners have archived the repository on GitHub, so it is read-only and no longer receives changes.
- What is it written in?
- Mainly TypeScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- bytedance/UI-TARS-desktop39,599 ★ · +1,020
| Repository | 2026-08-13 | 2026-11-15 |
|---|---|---|
| bytedance/UI-TARS-desktop | 38.6k | 39.6k |
1,020 Stars in the last 94 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
What This Repository Actually Ships
The repository bytedance/UI-TARS-desktop is not a single application. It is the home for two related projects under the TARS name. UI-TARS Desktop is a desktop application that runs a GUI agent based on the UI-TARS model, and it provides local and remote operators for both computer and browser. Agent TARS is a broader stack that brings GUI agent and vision capabilities into the terminal, computer, browser, and your own product. It ships primarily as a CLI and a Web UI. The README describes Agent TARS as aiming for a workflow closer to human-like task completion, which is a strong claim. The repository is written in TypeScript, licensed under Apache-2.0, and the default branch is main. The most recent release is v0.3.1 from October 2026, preceded by v0.3.0 from November 2025 and several beta releases. This is an active project with a clear split: one tool for end users who want a desktop agent, another for developers who want to embed agent capabilities or run them from a terminal.
The Two Operators: Local and Remote, Computer and Browser
UI-TARS Desktop's core feature is the ability to operate a computer or browser. The README distinguishes between local and remote operators. The local operator runs on your machine, presumably controlling the same desktop session. The remote operator, introduced in v0.2.0, lets you control a computer or browser over the network. The release notes for v0.2.0 say both remote operators are free and require no configuration. The distinction matters for use cases: a local operator is for automating your own machine, while a remote operator could handle a cloud VM or a shared testing environment.
Agent TARS: CLI, Web UI, and MCP Integration
Agent TARS is the more developer-facing half. It provides a CLI and a Web UI, and it integrates with MCP (Model Context Protocol) tools. The README shows a showcase example where the agent books a flight on Priceline, and another where it generates a chart with extra MCP servers. This indicates that Agent TARS is not just a GUI agent; it can call external tools that are not part of the core model. The CLI is published as an npm package, @agent-tars/cli, which means installation is likely via npm. The README references a quick-start guide and a documentation site at agent-tars.com. The v0.3.0 release notes mention streaming support for multiple tools, including shell commands and multi-file structured display, plus runtime settings with timing statistics for tool calls and deep thinking. There is also an Event Stream Viewer for tracking data flow and debugging. This suggests that Agent TARS is designed for observability, which is a practical concern for any agent that makes multiple tool calls.
Getting It Running: Commands and Configuration
The README links to a quick-start guide for UI-TARS Desktop and to the Agent TARS CLI documentation. For UI-TARS Desktop, the quick-start guide is at docs/quick-start.md, and it covers both local and remote operators. The local operator requires getting a model and running it locally, while the remote operator is described as requiring no configuration. For Agent TARS, the npm package @agent-tars/cli is the entry point. The documentation site, agent-tars.com, has pages for CLI, Web UI, and MCP integration.
What the v0.3.0 Release Adds and What It Breaks
The v0.3.0 release, dated November 2025, was followed by v0.3.1 in October 2026. The release notes highlight streaming support for multiple tools, meaning the agent can output shell command results or file contents as they arrive, rather than waiting for completion. It also adds runtime settings with timing statistics for tool calls and deep thinking, which is useful for profiling where the agent spends its time. The Event Stream Viewer is a debugging feature that shows the data flow between the model and tools. More significantly, v0.3.0 introduces exclusive support for the AIO agent Sandbox, an isolated environment for tool execution. That is a major architectural change: instead of running shell commands directly on your machine, the agent can run them inside a sandbox. This is a good security move, but it also means that existing workflows that depend on direct filesystem access may need adjustment. The beta releases before v0.3.0, such as v0.3.0-beta.12, suggest that the API may have changed during development. Anyone upgrading from an earlier version should check the release notes for breaking changes.
Licence and Maintenance Cost
The repository is licensed under Apache-2.0, which is permissive and allows commercial use, modification, and redistribution, provided you preserve the license notice. That is a positive for adoption, as it avoids the restrictions of some other agent frameworks. However, the project is under active development, with a last push on 2026-10-09 and a release cadence that includes multiple betas. That means you should expect frequent updates and potential API instability. The README mentions a contribution guide and a citation section, indicating an open development model. The maintenance cost is not trivial: you will need to track releases, update your installation, and adapt to changes in configuration or CLI flags. The documentation is split between the GitHub repo and agent-tars.com, so you may need to check both places for current information.
Limitations and When This Is the Wrong Tool
The most obvious limitation is that UI-TARS Desktop is built around the UI-TARS model. The README mentions the UI-TARS model and the advanced UI-TARS-1.5 model, and .env.example configures a VLM provider, base URL, API key and model name. Another limitation is that the remote operator is described as free and configuration-free, so check how it handles latency and security before using it for sensitive tasks. Also, the v0.3.0 sandbox is exclusive to AIO, which means if you want isolation, you are locked into that specific sandbox implementation. For users who need a simple desktop automation tool without a heavy agent stack, the repository might be overkill. The CLI and Web UI add complexity that a basic screen-scraping script would not need.
A Real Alternative: UI-TARS SDK and Other GUI Agents
Within the same repository, there is the UI TARS SDK, introduced in February 2025. The README describes it as a cross-platform toolkit for building GUI automation agents. This is a different approach: instead of using the ready-made desktop app or CLI, you build your own agent using the SDK. That gives you more control but requires more development effort. Outside this repository, the obvious alternative is a traditional RPA tool like Selenium or Playwright for browser automation, but those do not use a multimodal model; they rely on selectors and scripts. The difference is that UI-TARS Desktop can understand a screen image and decide what to click, whereas Playwright requires you to specify selectors. Another alternative is a general-purpose agent framework like LangChain, which can call tools but does not have built-in GUI understanding. The choice depends on whether you need visual reasoning or just structured automation.
Editorial conclusion
Adopt this stack if you need a GUI agent that can operate a real computer or browser, or if you want a CLI-based multimodal agent that can call MCP tools and run in an isolated sandbox. Do not adopt it if you need a production-hardened agent with a stable API and mature documentation, because the repository is moving fast, with beta releases and v0.3.0 (November 2025), which introduced significant changes like streaming and a sandbox, followed by v0.3.1 in October 2026. Before committing, verify that your target model is supported by the UI-TARS model family, check the latest release notes for breaking changes, and test the Event Stream Viewer to understand how the agent makes decisions.
Frequently asked questions
What is UI-Tars desktop?
It is the desktop application in this repository, a native GUI agent built on the UI-TARS model that ships local and remote computer operators plus browser operators. The same repository also ships a second project, Agent TARS, delivered as a CLI and a Web UI, so the repository name covers both.
How to use UI-tars?
UI-TARS Desktop drives the computer through local, remote computer and browser operators, and the model it drives with is configured through four variables: VLM_PROVIDER, VLM_BASE_URL, VLM_API_KEY and VLM_MODEL_NAME, shown with placeholder values in .env.example. The root dev:ui-tars script runs a monorepo command, turbo run ui-tars-desktop#dev.
how to install ui tars desktop
The source tree is a pnpm workspace: packageManager is [email protected], pnpm-workspace.yaml and pnpm-lock.yaml define it, and .node-version pins the Node release. Running pnpm i at the root (the bootstrap script) gives you a private monorepo package versioned 0.0.1 that publishes nothing, and model access is configured separately through the four VLM variables.
What are the key differences between UI-TARS desktop and Agent TARS desktop?
There is no project named Agent TARS desktop in the repository, so the two things it ships are Agent TARS, a multimodal agent stack delivered as a CLI and a Web UI, and UI-TARS Desktop, the desktop application with local and remote computer and browser operators.
Is there an AI that can control your desktop?
Yes, that is the stated purpose of UI-TARS Desktop, which provides a native GUI agent based on the UI-TARS model, and since v0.2.0 on 2025-06-12 it includes a Remote Computer Operator and a Remote Browser Operator described as free and needing no configuration.
ui tars desktop alternative
The README points at the separate UI-TARS model repository, the Agent TARS project that ships alongside the desktop app, and the agent-infra/sandbox AIO agent Sandbox that v0.3.0 supports as an isolated tool execution environment.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/bytedance-ui-tars-desktop)Star history chart for your README
[](https://hysenlabs.com/projects/bytedance-ui-tars-desktop)Was this useful?