Computer Use Agents

Every agent that can use a computer

A directory of AI agents that operate a computer directly — reading the screen, moving the pointer and typing, rather than calling an API. Compared on what they control, how they are licensed, and how you actually get hold of them.

23Agents listed
10Open source
15Full desktop
8 OctLast reviewed
Agent S Open agentic framework for desktop control, with an emphasis on experience-based memory between runs. Simular Desktop Apache-2.0 Self-host GA
Browser Use Python library that hands a browser to any LLM. One of the most widely adopted open-source options. Browser Use Browser MIT Self-host GA
ChatGPT Agent Retired. Agent mode inside ChatGPT, withdrawn in August 2026; OpenAI now points users at ChatGPT Work for multi-step tasks and at the cloud browser for browser workflows. OpenAI Browser Proprietary Product Superseded
Claude Browser Use Anthropic's browser-only toolset. Reads the page's accessibility tree and acts on named elements rather than pixel coordinates, in a browser your own application hosts. Anthropic Browser Proprietary API GA
Claude Computer Use Screenshot-and-control toolset exposed through the Anthropic API. Moves the mouse, types, and reads the screen of a full desktop environment you host yourself. Anthropic Desktop Proprietary API GA
Comet An agentic browser rather than a model: the assistant is built into the browsing surface itself. Perplexity Browser Proprietary Product GA
Computer use tool OpenAI's computer use tool on the Responses API, for developers building their own control loop. The earlier computer-use-preview path is deprecated. OpenAI Both Proprietary API GA
Copilot Studio computer use Computer use tool inside Copilot Studio. Automates Windows desktop applications and websites for agents, billed per step. Microsoft Both Proprietary Product GA
Cua Containerised desktop sandboxes for agents, with a focus on running macOS and Linux environments safely. trycua Desktop MIT Self-host GA
Dots Persistent agents inside ChatGPT, each with its own cloud computer and browser. They keep working after you close the tab, and you can watch the screen while they do. OpenAI Both Proprietary Product GA
Gemini Computer Use Computer use model on the Gemini API and Vertex AI, covering browser, mobile and desktop interfaces. Google Both Proprietary API Preview
Meta Muse Meta's personal agent. The Mac app, shipped in September 2026, acts inside native macOS applications — files, Messages, Calendar, Notes and mail — with permission. Meta Desktop Proprietary Product GA
Open Interpreter Runs model-authored code locally to operate the machine — a code-first rather than pixel-first approach. The original repository has been repurposed; the Python project continues as a community-maintained fork. Open Interpreter Desktop AGPL-3.0 Self-host GA
OpenAdapt Learns processes from recorded human demonstrations rather than from written instructions. OpenAdapt.AI Desktop MIT Self-host GA
OpenClaw Open-source personal assistant that runs on your own devices, reaches you through the messaging channels you already use, and can take desktop control. OpenClaw Foundation Desktop MIT Self-host GA
Operator OpenAI's first consumer browser agent, built on the Computer-Using Agent model. Folded into ChatGPT Agent, which has since been retired too. OpenAI Browser Proprietary Product Supersededunconfirmed
Perplexity Computer Cloud-hosted general-purpose agent that orchestrates work across many models in parallel, spawning subagents per task. Perplexity Desktop Proprietary Product GA
Personal Computer Perplexity Computer moved onto your own machine — reaching local files and native applications rather than a cloud sandbox. Perplexity Desktop Proprietary Product GA
Project Mariner Google's browser-agent research prototype. Reporting puts its shutdown at 4 May 2026 and DeepMind's model index no longer lists it. Google DeepMind Browser Proprietary Product Supersededunconfirmed
Self-Operating Computer Minimal framework for pointing a multimodal model at your own screen. Useful as a reference implementation. OthersideAI Desktop MIT Self-host GA
Skyvern Vision-driven browser automation aimed at workflows that break conventional selector-based scripts. Skyvern Browser AGPL-3.0 Self-host GA
Stagehand TypeScript framework layering natural-language actions over Playwright, so deterministic code and model calls can be mixed. Browserbase Browser MIT Self-host GA
UI-TARS ByteDance's open-weights vision-language agent for GUI control. Prompt templates cover desktop and mobile; browser automation is a separate sibling project. ByteDance Desktop Apache-2.0 Self-host GA

23 agents

What changed

All news ›
8 October 2026 · Correction We were listing an agent that had been dead for five months Project Mariner appears to have shut down on 4 May 2026. It sat in this directory as a live research preview until today. 8 October 2026 · Correction Open Interpreter's repository is no longer Open Interpreter The canonical repo is now a Rust harness based on Codex. The Python project continues as a community fork — under a different licence. 8 October 2026 · Method Every entry re-checked: two products moved, one licence was wrong, one surface was wrong The first pass to check all 23 entries at once, in both directions. Here is everything it changed, including the things that make us look bad.

What people use them for

All use cases ›
Personal · 15–40 steps Ordering the weekly groceries The demo everyone shows. Fifteen to forty actions, of which two actually require judgement. Personal · 20–60 steps Booking a flight or a hotel Comparison rather than a fixed sequence, which is what these agents are genuinely good at. Personal · 20–50 steps Filling in a long application form Pure transcription across several pages. Unglamorous, low-risk, and one of the highest-value things here.

Start here

All writing ›
Analysis The agents got their own computers In one quarter, OpenAI, Meta and Anthropic each answered the same question differently: whose machine does the agent run on? That answer predicts your risk far better than desktop-versus-browser does. Explainer What is a computer use agent? An agent that operates software the way a person does — reading the screen, moving the pointer, typing. Why that matters, what the loop actually looks like, and where it breaks. Comparison Claude vs OpenAI vs Gemini: three approaches to computer use The three big labs ship computer use in genuinely different shapes, and at genuinely different stages of maturity — one beta, one GA, one preview. Comparing them on score misses the point.

Notes on method

  1. Entries are listed alphabetically by default. Every column heading sorts.
  2. Controls distinguishes agents that drive a full desktop from those confined to a browser tab. Both means the vendor documents each.
  3. Access records how you actually obtain the agent: as a developer API, as a finished Product, or by running it yourself.
  4. Benchmark scores and pricing are absent by policy. Published figures vary by harness, step budget and date, and secondary sources mostly cite each other rather than a leaderboard.
  5. Every entry links to its primary source — vendor documentation or the project repository. Stage, control surface and licence are checked against that source, and each profile page carries the date it was last verified. Anything we could not confirm is marked unconfirmed rather than quietly asserted.