English |
TARS* is a Multimodal AI Agent stack, currently shipping two projects: Agent TARS and UI-TARS-desktop:
MCP.Pizza Chef: bytedance
Lets you connect an AI assistant to a virtual computer or browser so it can look at what's on the screen, click buttons, type text, and carry out multi-step tasks on its own, the same way a person would. It works as its own desktop app and can also connect to apps like VS Code, and it supports controlling either your own computer or a remote computer and browser over the internet. You will need an API key for the underlying AI model to power it.
Coding
Other
Web/Research
Use This MCP server To
Let my assistant click through a website for me Automate repetitive computer tasks like filling out forms Test a website by having my assistant click through it Control a remote computer or browser without local setup Complete multi-step tasks that span several apps automatically Have my assistant look at my screen and describe what it sees
README
UI-TARS-desktop FAQ
Can I use this to automate clicking through a website?
Yes — it can see a browser window and click, type, and scroll just like a person would, so you can hand off repetitive browsing tasks.
Do I need to write any code to use this?
No coding is required, but setup does involve pasting a small configuration snippet into your AI app's settings.
Which apps can I use this with?
It works with VS Code and other apps that support the Model Context Protocol, and it also ships as its own standalone desktop application.
Do I need an API key?
Yes, you need an API key for the underlying AI model, such as one from OpenAI, Claude, or Gemini, to power the agent's decisions.
Is this hard to set up?
Setup is moderate — you paste a short configuration snippet into your app rather than writing code, though more advanced remote setups may need a developer's help.
Can this control a computer I don't own or one running somewhere else?
Yes — it includes a remote computer and browser mode that lets the assistant control a separate machine over the internet without installing anything locally.
Can I use this to test how a website behaves?
Yes — since it can navigate and interact with a page the way a visitor would, you can use it to click through a site and confirm things work as expected.
Is it safe to let this run on my own computer?
It's meant to be given specific tasks and typically lets you review what it's about to do, but as with any tool that can click and type on your behalf, only run it with a model and setup you trust.