Cua open-sources computer-use agent stack: cloud desktops, driver, System 1 models and benchmarks

An all-in-one toolkit for agents that use computers

The open-source project Cua, published by Cua AI, Inc. at trycua/cua and recently featured on Hacker News's front page, packages everything an AI agent needs to operate a computer: isolated desktops, automation tooling, specialist decision models and an evaluation harness. The core source is MIT-licensed, with model weights and some third-party components handled separately.

According to the project's repository, Cua deliberately positions itself as the infrastructure layer: you bring your own agent and model, and Cua supplies the computer and the automation tools. The project frames its target as "Computer-Use 2.0" — an agent that moves between code, APIs and graphical interfaces within a single task, rather than only clicking through a screen.

Cua Driver for operating real applications

Cua Driver gives an agent tools to inspect and operate native desktop applications and browsers on macOS, Windows and Linux. The repository lists three integration paths — a CLI, MCP and typed SDKs — and points to integrations for agent tools such as Claude Code, Codex, Cursor and OpenClaw. A background delivery mode lets agents act without moving the user's pointer or stealing window focus, where the app and platform support it. The project's demonstration shows two driver sessions selecting cells in LibreOffice Calc and objects in Inkscape on an Omarchy Linux desktop while a terminal stays usable in the foreground.

CUA-S1, small models for bounded decisions

Alongside the plumbing, Cua ships its own family of small, specialized models called CUA-S1. The "System 1" label is an engineering analogy, the repository stresses: these models handle fast, bounded decisions — such as which value belongs in a form field, or whether to leave an element alone — and are not a strict architectural classification or a replacement for a general-purpose agent's planning and reasoning.

The first research profile, CUA-S1-FORMS, targets forms. Instead of generating a response token by token, it scores decisions drawn from structured interface elements and document values; application code orders the actions, and the optional Cua Driver integration executes them with explicit action boundaries. The release covers Python model code, synthetic-data generation, training and evaluation. The GitHub component is an early, source-only research release, with model weights and a dataset hosted separately on Hugging Face, and each artifact carries its own license and documented limitations.

Sandboxing through Fleets and Lume

Two components provide the computers themselves. Cua Fleets provisions isolated Linux desktops in the cloud via run.cua.ai: a Fleet maintains sandbox capacity, code claims a desktop from a pool, and the Sandbox SDK runs commands, captures screenshots and interacts with apps inside it. The documentation warns that pools can retain paid capacity after a claim ends, so following the cleanup steps matters.

For local work, Lume creates and manages macOS and Linux virtual machines on Apple Silicon using Apple's Virtualization.Framework — the quickstart builds a macOS Tahoe VM from an Apple restore image and connects over SSH. Local sandboxes and Fleets share the Sandbox SDK, though credentials, images, operations and runtime requirements differ between the two environments.

Cua Bench for evaluation and training data

Cua Bench rounds out the stack with tooling to define computer-use tasks, evaluate agents and export trajectories for training. The quickstart is deliberately light: with Python 3.12 or 3.13 and uv installed, a simulated task runs with no VM, no Docker and no model API key, and verifies that the evaluator reports a reward of 1.0.

Licensing notes

The main project is MIT-licensed, but the repository flags third-party components under their own terms: Kasm is MIT, OmniParser is CC-BY-4.0, and the optional OmniParser integration pulls in ultralytics under AGPL-3.0 — a distinction commercial adopters should check. The project also notes it is not affiliated with or endorsed by Apple, Microsoft or Canonical despite supporting their platforms.

Why it matters

Most computer-use efforts to date have been assembled from pieces: one vendor's model, a screenshot loop, ad hoc virtual machines and hand-rolled evaluations. Cua covers the entire pipeline — sandboxed infrastructure, cross-OS actuation, specialist decision models and reproducible benchmarks — in a single open-source project that is explicitly agent-agnostic. That lowers the barrier to experimenting with agents that mix code, APIs and GUIs in one task, and Cua Bench's trajectory export points at the real bottleneck: generating and evaluating training data for computer use is as hard as running the agents. The early, source-only state of CUA-S1 and the mixed licenses on some components mean production adoption needs diligence, but as a research and integration platform the stack is unusually complete.