Cua is an MIT-licensed, open-source platform that gives AI agents computers to use, packaging desktop automation drivers, isolated cloud desktops, local macOS virtual machines, and benchmarks for training, evaluation, and data generation — it is aimed at developers and teams building computer-use agents who bring their own agent and model.

What it is

Cua lives in the computer-use agent ecosystem, alongside the agent hosts it connects to. The project supplies the computer and the automation tools while the user supplies the agent and the model. It bundles four distinct pieces: Cua Driver for inspecting and operating native desktop apps and browsers, Cua Fleets for isolated cloud desktops, Lume for local macOS and Linux virtual machines on Apple Silicon, and Cua Bench for creating and verifying simulated tasks. The README frames the target as Computer-Use 2.0, where an agent moves between code, APIs, and graphical interfaces within the same task rather than staying inside one modality.

The concrete problem is assembly. Without a platform of this kind, a team that wants an agent to click through LibreOffice Calc or drive an Inkscape canvas has to build its own desktop automation layer, its own sandbox lifecycle management, and its own evaluation harness before writing a single prompt. Cua replaces that ad hoc plumbing with a driver that speaks CLI, MCP, and typed SDKs, a Sandbox SDK that runs commands and captures screenshots inside a claimed desktop, and a benchmark surface for scoring the results. Local sandboxes and Fleets share the Sandbox SDK, so the same code shape carries from a developer laptop to pooled cloud capacity, though credentials, images, operations, and runtime requirements differ between the two.

Key capabilities

• Cua Fleets provisions isolated cloud desktops at run.cua.ai, where a Fleet maintains sandbox capacity and code claims a desktop from a pool.
• The Sandbox SDK runs commands, captures screenshots, and interacts with applications inside a claimed desktop, with a runtime support reference for choosing an environment.
• Cua Driver operates native desktop apps and browsers on macOS, Windows, and Linux, with background delivery that avoids moving the pointer or taking focus where the app and platform support it.
• Driver connection paths include the CLI, MCP, and typed SDKs, with documented integrations for Claude Code, Codex, Cursor, and OpenClaw.
• Lume creates and manages local macOS and Linux VMs on Apple Silicon using Apple's Virtualization.Framework, including a vanilla macOS Tahoe VM built from an Apple restore image.
• Cua Bench creates and verifies simulated tasks for evaluation work.
• Install scripts ship for each component: /bin/bash -c "$(curl -fsSL https://cua.ai/driver/install.sh)", PowerShell irm https://cua.ai/driver/install.ps1 | iex, and /bin/bash -c "$(curl -fsSL https://cua.ai/lume/install.sh)".

Who uses it and how

• Agent developers already working in Claude Code, Codex, Cursor, or OpenClaw connect their existing agent to Cua Driver and ask it to compute 6 × 7 in Calculator, then verify the app displays 42.
• Teams needing parallel machines claim desktops from a Fleet pool and drive them through the Sandbox SDK, following the tutorial that provisions a Linux desktop, runs uname -a, saves a screenshot, and deletes the cloud resources.
• Apple Silicon developers create a macOS Tahoe VM locally with the Lume CLI and connect over SSH, keeping the work off shared infrastructure.
• Evaluation and data-generation groups use Cua Bench to construct and verify simulated tasks rather than scoring agents against ad hoc scripts.
• Desktop-environment experiments run on cloud capacity, including the documented Omarchy on Fleet guide; the README demo shows two Cua Driver sessions selecting cells in LibreOffice Calc and objects in Inkscape while a terminal stays in the foreground.

Getting started

Install Cua Driver on macOS or Linux with /bin/bash -c "$(curl -fsSL https://cua.ai/driver/install.sh)" or on Windows with irm https://cua.ai/driver/install.ps1 | iex. Cua Fleets is available as a hosted service at run.cua.ai, and each path has a first-result tutorial covering credentials, permissions, dependencies, and cleanup.

How it compares

No list of paid products that Cua replaces appears in the supplied facts, so it stands alone in this registry as an open-source computer-use platform rather than a swap-in for named commercial tools. Its overlap is with the agent hosts it integrates with — Claude Code, Codex, Cursor, and OpenClaw — which supply reasoning and driving while Cua supplies the machine and the automation surface.

When to use it — and when not to

Self-hosters who stay on Lume and local sandboxes must run VMs on their own Apple Silicon hardware and manage the local sandbox lifecycle themselves, while Fleet users must handle credentials and delete cloud resources, because pools can retain paid capacity after a claim ends. Anyone expecting a single uniform runtime should note that local sandboxes and Fleets differ in credentials, images, operations, and runtime requirements despite sharing the Sandbox SDK. The repository also carries 926 open issues, which is a signal of a fast-moving project rather than a settled one.