Updated October 2026
AI agents that control your computer
A computer-use agent operates software the way you do: it looks at the screen, moves the mouse, types and clicks. That lets it work with apps that have no API, but it also makes it the riskiest kind of agent, because a mistake or a malicious instruction can reach anything you can. Two questions matter most: where does it run (your own desktop, a cloud machine, or only a browser), and what stops it doing the wrong thing.
Every entry below was checked against the developer's own documentation or repository in October 2026.
At a glance
| Agent | Open source | Where it runs | What it controls | Platforms |
|---|---|---|---|---|
| CoWork OS | Yes (MIT) | Your own computer | Native desktop apps, plus browser and files | macOS, Windows |
| Claude computer use (API) | Reference code only | An environment you provide, such as a VM or container | A full desktop | Your choice; examples use Linux |
| Manus | No | A cloud sandbox, or your computer with Manus Studio | Cloud computer, local files and apps | Web; macOS (Apple Silicon) and Windows app |
| UI-TARS Desktop | Yes (Apache 2.0) | Your own computer | Desktop and browser | Windows, macOS, browser |
| Agent S3 | Yes (Apache 2.0) | Your own computer | Full desktop | Linux, macOS, Windows |
| Microsoft UFO | Yes (MIT) | Your own computer | Windows apps through UI and APIs | Windows (UFO²); multi-device (UFO³) |
| Browser Use | Yes (MIT) | Locally or in Browser Use Cloud | Browser only | macOS, Linux (Windows via WSL) |
| Perplexity Comet | No | Inside the Comet browser | Browser only | Mac, Windows, iOS, Android |
The agents
1. CoWork OS
CoWork OS includes computer use as one tool among many, not the default. It drives native apps on macOS through a bundled helper (enabled in Settings → Tools → Computer use with Accessibility and Screen Recording permissions) and on Windows one visible window at a time. The planner prefers APIs, MCP connectors, the in-app browser and scripts, and only reaches for computer use when a task clearly needs a desktop UI.
Safety controls: computer use must be allowed by the task's access profile and stays subject to approvals and admin policy; only one session runs at a time; Esc aborts immediately; risky key combinations are blocked; and it will not operate over private browser prompts such as passwords or checkout. Computer use docs
Best for: occasional desktop automation inside a wider agent that mostly uses safer tools.
2. Claude computer use (API)
Anthropic's computer-use tool lets Claude operate a full desktop through screenshots, mouse and keyboard. It is a developer API: "your application runs every call in an environment you control", and the reference implementation is a Docker container with a Linux desktop. The current toolset is generally available and billed at normal token rates. Anthropic recommends a dedicated VM or container with minimal privileges, a domain allowlist and human confirmation for important decisions, and scans screenshots for prompt injection. Anthropic docs
Best for: developers building their own computer-use product.
3. Manus
Manus "operates in a complete sandbox environment—a virtual computer with internet access" in the cloud. Its desktop app, Manus Studio for Apple Silicon Macs and Windows, adds computer use that "works in its own workspace on your computer while you stay in yours". manus.im
Best for: handing off long tasks to a hosted agent.
4. UI-TARS Desktop
ByteDance's Apache-licensed "native GUI agent for your local computer", with mouse and keyboard control and a browser operator. The model can be self-hosted. On macOS it needs Accessibility and Screen Recording; it supports one monitor. Its free Remote Operator cloud service was discontinued in August 2025. GitHub
Best for: a free, local GUI agent you can self-host end to end.
5. Agent S3
Simular's open framework for agents that operate "a real GUI — mouse, keyboard, and screen" on Linux, macOS and Windows. The repository warns that the agent runs Python code to control your computer; its arbitrary-code option is off by default. Simular's hosted product, Sai, runs on a cloud computer. GitHub
Best for: researchers and developers experimenting with GUI agents.
6. Microsoft UFO
A Microsoft research project. UFO² is a "Desktop AgentOS" for Windows that mixes GUI actions with Windows APIs and is in long-term support; UFO³ Galaxy extends orchestration to Linux and Android. GitHub
Best for: Windows automation research.
7. Browser Use
An MIT-licensed Python library for AI agents that work in a browser, run locally with any model or on Browser Use Cloud (pay as you go). It controls the browser only. GitHub
Best for: web automation where you don't need the rest of the desktop.
8. Perplexity Comet
A browser with a built-in assistant that acts on web pages for you, available on Mac, Windows, iOS and Android, with a control to pause the assistant. Browser only. perplexity.ai/comet
Best for: everyday web tasks without setting anything up.
Browser-only or full desktop?
Most tasks people want automated happen on websites, and browser agents are easier to contain. Full desktop control is worth it for native apps with no web or API alternative. A good rule: use an API or connector first, a browser agent second, and full computer use only when nothing else works. CoWork OS applies that order automatically.
Staying safe with any computer-use agent
- Run it in a separate account, VM or container when you can.
- Require approval before purchases, sign-ins, sending messages and deleting anything.
- Keep a way to stop it instantly.
- Be careful with tasks that read untrusted pages or emails, which can carry hidden instructions.
We did not include OpenAI's ChatGPT agent because we could not verify its current details from OpenAI's site at the time of review.
Frequently asked questions
What is a computer-use agent?
An AI agent that operates software through the screen, mouse and keyboard, like a person would, so it can use apps that have no API.
Which AI agents can control my Mac?
CoWork OS, UI-TARS Desktop, Agent S3 and Manus Studio can operate a Mac. On macOS they need Accessibility and Screen Recording permissions.
Are computer-use agents safe?
They carry more risk than other agents because they can reach anything you can. Use one with approvals and a quick stop, run it in a separate account or VM where possible, and avoid giving it untrusted content.
Is there an open-source AI agent that controls the computer?
Yes. CoWork OS (MIT), UI-TARS Desktop (Apache 2.0), Agent S3 (Apache 2.0) and Microsoft UFO (MIT) are open source, and Browser Use (MIT) covers the browser.
CoWork OS itself is free and open source. Model providers, gateways, connected services, and account plans may charge separately and apply their own eligibility and usage rules.
Last reviewed 2026-10-20 against official documentation. Features, plans and prices change; check the linked sources before deciding. CoWork OS is not affiliated with any other product listed here.