Comet
Lets an AI assistant in your web browser research, compare pages and handle web tasks
by Perplexity
Agent TARS
by ByteDanceAgentGet organised
Lets you hand web and research tasks to ByteDance's agent, which works in a browser
About Agent TARS
Agent TARS is an open-source AI agent you start from a terminal and then use in a web page. You describe a task in plain words and it works through it by controlling a Chrome browser, using your own AI model key.
For Technical makersPeople automating browser tasks
Made by ByteDance
ByteDance, the technology company, publishes this open-source project on GitHub as part of its UI-TARS-desktop repository.
- Stars
- ★ 39k
- License
- Apache-2.0
- Cost
- Usage-based
- Platforms
- Mac · Win · Linux
- Level
- Intermediate
Screenshots
From the project's own repository · 2 images, swipe or use the arrows.
Overview
Agent TARS is ByteDance's open-source AI agent that you start from the command line and then use in a web page on your computer. You describe a task in plain words; it controls a local Chrome browser by reading the page or by looking at screenshots, and it can call extra MCP tools, for example to draw a chart. You bring an API key for a vision-capable model. It suits technical makers who are comfortable running one command in a terminal.
It can work a browser by looking at the screen, by reading the page structure, or by mixing the two, and shows each step in its web interface.
Try these first
Start with the first one: it is a small task you can check easily.
- Open a public weather page for [city] and list the forecast for the next three days.First try · read-only
- Turn these numbers into a bar chart: [paste numbers].uses an MCP tool
- Find the three most recent posts on [public blog] and summarise each in two sentences.
- Search [site] for [item] and compare the top three results by price and rating.
Example prompts written by Agent Flows Hub.
Features
From the creator's documentation.
| Feature | What it does | Runs |
|---|---|---|
| Web interface | Chat with the agent in a local web page that shows each step and tool result | Your computer |
| Browser control | Works a local Chrome by looking at screenshots, by reading the page, or both | Your computer |
| MCP tools | Add MCP servers so it can use more tools, such as chart makers | Other services |
| Model choice | Use Volcano Engine, Anthropic or OpenAI with your own key | Provider API |
| Replays | Save a replay of a session on your computer, or share it to a server you set up | Your computer |
Good to know
You want an agent to carry out a multi-step web task and show you each step. Agent TARS drives a local Chrome browser with the model you choose and can call MCP tools, for example to turn data into a chart.
Limits
- You start it from a terminal with your model provider, model name and API key. The official quick start examples name older models, so check which model your provider offers today.
- It controls your local Chrome by default, so a task can click and type on real sites. The setup steps are in the GitHub README.
Notes by Agent Flows Hub (original project name: Agent TARS (in the UI-TARS-desktop repository)), written from the creator's documentation, not from running it ourselves. Agent TARS: overview and quick start · Agent TARS v0.3.1 release notes
Stuck setting it up?
Tell us what happened. We take a look first; if it needs paid work, you get an estimate before anything starts.
More to get organised
Other agents
- AionUi Puts an AI agent on your desktop that works on your files and makes Office documents
- AIPOCH Open-Science Hands research tasks to AI agents that read papers, run analyses and record each step
- Base44 Builds a working app or website with sign-in and a database from a chat description
- Bolt Builds websites, web apps and mobile apps from a chat prompt, ready to publish
- BrowserOS Gives you a Chromium-based browser with an AI agent in every new tab
- Cindy Hands work on your computer to an AI agent you can reach from your desktop or phone