cua

Quick Verdict

cua is best understood as infrastructure for computer-use agents, not one desktop assistant. Cua Driver lets agents operate native applications in the background; the Sandbox SDK creates local or cloud environments across Linux, Windows, macOS, and Android; Cua Bench evaluates agents and exports trajectories; Lume manages macOS and Linux virtual machines on Apple Silicon; and Cua Fleets scales training, evaluation, and data-generation workloads across parallel machines.

As of July 21, 2026, the official repository had same-day activity, and an @trycua/fleet maintenance release shipped on July 15. The core repository is MIT-licensed, but optional components can carry separate obligations. The official license notes identify OmniParser as CC-BY-4.0 and the optional cua-agent[omni] dependency ultralytics as AGPL-3.0. The website offers hosted, BYOC, and on-premises fleets, but does not publish per-minute machine prices. A cloud API or access request should not be described as free compute without a current commercial quote.

Best For

• AI infrastructure teams training or evaluating computer-use models across several operating systems.
• Developers connecting Claude Code, Cursor, Codex, or custom agents to native desktop applications.
• QA teams automating legacy applications, mixed browser-and-desktop flows, or mobile interfaces.
• Researchers who need reproducible snapshots, trajectories, benchmarks, and parallel rollout capacity.
• It is not ideal for a browser-only workflow, a team without virtualization operations, or an organization that cannot audit and approve high-impact GUI actions.

Key Features

• Cua Driver: A shared CLI and MCP stdio surface for native applications on macOS, Windows, and Linux, with background input designed not to take over the user’s cursor and focus.
• Sandbox SDK: One Python API for Linux containers and VMs, macOS, Windows, and Android, including shell, screenshot, mouse, keyboard, and touch operations.
• Local and cloud backends: Docker, QEMU, and Apple Virtualization Framework support local execution, while Cua Cloud supplies hosted cross-OS environments.
• Cua Fleets: Warm pools and snapshot-based machines support parallel rollouts, evaluation, reinforcement learning, and data generation.
• Cua Bench: OSWorld, ScreenSpot, Windows Arena, and custom tasks can be run with trajectory export for evaluation or training.
• Lume: Near-native macOS and Linux virtualization on Apple Silicon, subject to host-platform and operating-system licensing constraints.
• Failure reproduction: Images, snapshots, and fixed tasks make a failed episode easier to replay than an action performed on an employee’s live desktop.

Use Cases

Practical uses include testing enterprise software on isolated Windows machines, replaying Xcode tasks in macOS VMs, running OSWorld evaluations in parallel, collecting screenshot-and-action trajectories, and letting a coding agent control a test application without stealing the user’s cursor. A credible benchmark should pin the image, application version, resolution, language, seed, and evaluator. Otherwise, environment drift and model changes become impossible to separate.

Computer use exposes more than a browser session. Screens, clipboards, file dialogs, notifications, and logged-in applications can reveal sensitive information, while an app or page may contain prompt injection. Use disposable VMs, network and domain allowlists, temporary credentials, dedicated test tenants, read-only inputs, and controlled output directories. Deletion, software installation, settings changes, messages, transfers, orders, and payments must pause for confirmation. A VM or container narrows some impact, but it does not eliminate hypervisor escape, incorrect mounts, credential leakage, or misuse of legitimate permissions.

Pricing

The open Driver, Sandbox, Bench, and Lume components can run on infrastructure you control under their applicable licenses. The public website offers hosted fleets, BYOC, and on-premises deployment, but on July 21, 2026 it only presented “request access” and dedicated fleets by request, not a rate card by operating system, CPU, memory, or minute. Buyers should request a quote covering image storage, snapshots, network traffic, concurrency, warm capacity, logs, support, and any minimum commitment.

Cost controls must cover maximum concurrency, episode duration, machine size, snapshot retention, model tokens, and a daily budget. Retries should start from a clean snapshot. External writes need idempotency keys and backend state checks so a replay does not submit a form or payment twice. For local deployments, “free software” still means paying for hosts, licensed operating systems, storage, networking, and engineering time.

Pros

• A single API spans containers, desktop VMs, and Android for cross-OS computer-use development.
• Driver, Sandbox, Fleets, Bench, and Lume cover operation, scaling, evaluation, and reproduction.
• Snapshots and warm pools can improve benchmark throughput and failure analysis.
• The core MIT license is permissive, and repository maintenance was active on the review date.
• CLI, MCP, and SDK interfaces support several coding agents and custom model loops.
• Local backends let teams validate workflows before committing to hosted fleet capacity.

Cons

• Cua Cloud has no public unit prices, making comparison difficult without a sales quote.
• Cross-OS support does not mean identical behavior; Wayland, macOS setup, Windows, and mobile backends have platform-specific limits.
• GUI automation remains sensitive to layout, resolution, popups, application updates, and timing.
• Optional vision dependencies have different licenses, so the repository’s MIT badge is not a complete compliance review.
• Virtual environments are not an absolute security boundary and still need egress, credential, approval, recording, and audit controls.
• Fleet-scale failures can multiply compute and model costs quickly if retries are unbounded.

Alternatives

FAQ

Is cua a desktop agent?

It is more accurately a computer-use infrastructure suite. Teams can build agents with Driver, Sandbox, and the Agent SDK, then evaluate or train them with Bench and Fleets.

Does Cua Cloud have a free allowance or public unit price?

No public cloud-machine rate card was available on July 21, 2026. The site offered access requests and dedicated fleet discussions, so obtain a current quote instead of assuming an API key includes free capacity.

Is every component covered only by MIT?

No. The core repository is MIT, but third-party and optional components retain their own licenses. The official notes call out CC-BY-4.0 for OmniParser and AGPL-3.0 for optional ultralytics.

Does a VM completely isolate dangerous actions?

No. It reduces host exposure, but shared folders, clipboards, egress, temporary credentials, and virtualization vulnerabilities remain. Combine isolation with least privilege and approval gates.

How can a GUI agent avoid duplicate submission or payment?

Require confirmation immediately before the final action, use test accounts and transaction caps, attach an idempotency key, and verify backend transaction state before any retry.

Is cua appropriate for browser automation?

Yes, but Browser Use or Browser Harness is usually lighter for browser-only work. cua is differentiated by complete desktops and cross-OS infrastructure.

Bottom Line

cua moves computer use from a single-machine demo toward reproducible cross-OS environments, benchmarks, and fleets. That is valuable for model research, desktop software testing, and trajectory generation, but can be excessive for an ordinary web workflow. Start with a fixed-image benchmark and measure success rate, human takeover, cost per episode, and reproducibility before buying hosted capacity. Local or cloud, the production design still needs domain allowlists, credential isolation, idempotent retries, budget caps, and human confirmation for high-impact actions.