trycua/cua

Analysis updated 2026-06-24

★ 16,526HTMLAudience · researcherComplexity · 4/5Setup · hard

Mindmap

mindmap
  root((cua))
    Inputs
      Agent code
      VM images
      Benchmark tasks
    Outputs
      Controlled desktop sessions
      Screenshots
      Benchmark results
    Use Cases
      Build computer-use agents
      Run mac and Linux sandboxes
      Benchmark agent models
    Tech Stack
      Python
      QEMU
      Docker
      Apple Silicon

mindmap root((cua)) Inputs Agent code VM images Benchmark tasks Outputs Controlled desktop sessions Screenshots Benchmark results Use Cases Build computer-use agents Run mac and Linux sandboxes Benchmark agent models Tech Stack Python QEMU Docker Apple Silicon

Click or tap to explore — scroll the page freely

What do people build with it?

USE CASE 1

Spin up a sandboxed Linux or macOS VM and drive it with a Python script that clicks, types, and screenshots.

USE CASE 2

Build an AI agent that automates desktop apps without taking over the user's foreground session via Cua Driver.

USE CASE 3

Give a coding assistant its own sandbox with browser and mobile simulation through CuaBot.

USE CASE 4

Benchmark a computer-use model on standardized tasks and export recordings to train it further.

What is it built with?

PythonQEMUDockerApple Silicon

How does it compare?

	trycua/cua	emilwallner/screenshot-to-code	windmill-labs/windmill
Stars	16,526	16,482	16,469
Language	HTML	HTML	HTML
Setup difficulty	hard	hard	hard
Complexity	4/5	4/5	4/5
Audience	researcher	researcher	ops devops

Figures from each repo's GitHub metadata at analysis time.

How do you get it running?

Difficulty · hard Time to first run · 1h+

Needs virtualization (QEMU or containers) plus Apple Silicon for the macOS VM features, expect non-trivial system setup.

In plain English

Cua (trycua/cua) is an open-source toolkit for building, testing, and running AI agents that can operate full desktop computers, clicking, typing, taking screenshots, and completing tasks autonomously across macOS, Linux, Windows, and Android. The project is structured as several interconnected tools. The core component is a sandbox SDK (a safe, isolated environment) that lets you spin up virtual machines or containers running any operating system, then programmatically control them. A code example shows creating a Linux sandbox and then running shell commands, taking screenshots, clicking at coordinates, and typing text, all with the same Python API regardless of which OS you target. Sandboxes can run locally via QEMU (a virtualization tool) or in the cloud. Cua Driver handles background computer-use on macOS, an AI agent can click and type in native apps without interrupting the user, even in apps where normal accessibility shortcuts don't work. CuaBot is a command-line tool that gives any AI coding agent a sandbox to work in, with support for browser control, mobile device simulation, and screenshot capture. Cua-Bench provides benchmarks and testing environments for measuring how well computer-use agents perform on standardized tasks, and can export recordings for training AI models. Lume handles macOS and Linux virtual machine management on Apple Silicon hardware. The project is written primarily in HTML with Python SDKs and suits AI researchers, developers building automation, and teams evaluating computer-use AI models.

Copy-paste prompts

Prompt 1

Walk me through installing trycua cua on a Mac with Apple Silicon and running my first Linux sandbox.

Prompt 2

Write a Python script using the cua SDK that opens a sandbox, navigates to a website, and screenshots the result.

Prompt 3

Help me set up Cua Driver on macOS so my AI agent can click around a native app in the background.

Prompt 4

Explain how to plug an LLM agent into CuaBot so it can do tasks inside a fresh browser sandbox.

Prompt 5

Show me how to run a Cua-Bench task and export the recording to fine-tune a computer-use model.

Frequently asked questions

What is cua?

Open-source toolkit for building and benchmarking AI agents that control full desktop operating systems, with sandboxed VMs, drivers, and a Python SDK.

What language is cua written in?

Mainly HTML. The stack also includes Python, QEMU, Docker.

How hard is cua to set up?

Setup difficulty is rated hard, with roughly 1h+ to a first successful run.

Who is cua for?

Mainly researcher.

Open on GitHub → Explain another repo

This repo across BitVibe Labs

Scan in gitsafehub Deploy in gitdeployhub trycua on gitmyhub

Verify against the repo before relying on details.