The Cua project equips AI agents with full desktop environments and automation tools across macOS, Windows, and Linux. Cua Spaces launches a local virtual machine—either a macOS VM built on Apple Silicon, a Linux image, or an Omarchy build—accessible via a menu‑bar app that lets users teleport signed‑in applications (e.g., Chrome, Slack) into the Space while keeping sessions encrypted in the Cua Keyvault. Multiple users can share the same desktop with independent cursors, handing control back and forth. Agent‑ready images include the cua‑spacesd daemon on port 3211, providing processes, file access, screenshots, input injection, and low‑latency video/audio streams. The Cua Driver enables background control of native apps; a tutorial shows an agent opening Calculator, computing 6 × 7, and confirming the display of 42. Lume creates and manages macOS Tahoe or Linux VMs using Apple’s Virtualization.Framework, with a CLI command that provisions a VM from a restore image and connects over SSH. A unified cua CLI and SDK (available in Python, TypeScript, Swift, and Kotlin) manage sandboxes—for example, cua sb create ubuntu --name dev starts a gVisor container, cua sb exec dev uname -a runs a command, and cua sb screenshot dev captures the screen. The CUA‑S1 family supplies small, specialized “System 1” models for rapid, bounded decisions such as field‑value selection, accompanied by synthetic data generation and evaluation code. Finally, Cua Bench offers a simulated task that can be run without VMs or API keys; executing the reference solution yields a verified reward of 1.0, allowing researchers to benchmark their own agents.