mrkeyoor.comAsk

trycua/cuatested 27 Sept 2026commit e1c44bd

cua review

Cua deserves a component-level trial when an agent needs desktop controls, local VMs, or computer-use evaluation tools.

  • 2 packages
14 sto install
    AI Toolsevaluationupdated 31 Oct 2026
    +297stars / 7d

    Choose Cua if your agent needs native desktop controls or computers to run on, and you can validate the specific component on your target platform. Cua Driver provides tools to inspect and operate native desktop apps and browsers on macOS, Windows, and Linux. Our run covered a root install, not desktop automation or VM operation.

    Verdict

    Cua deserves a component-level trial when an agent needs desktop controls, local VMs, or computer-use evaluation tools. Our successful root install does not establish that those products work on your target machine. Choose the relevant quickstart, pin the component or commit, verify a real task, and inspect its license before committing to it.

    We ran it

    Lab card: what happened when we ran cua
    Lab card from our sandbox run
    Screenshot of cua (cua.ai)
    cua.ai
    Install✓ · 14s2 packages · 9 MB
    Buildn/ano build script
    Testsn/ano test script
    Repo5266 files~959,948 lines of source · 320 MB · 132 CI workflows · tests dir

    Answers from our run

    Does cua build from source?

    Dependencies installed in 14 seconds (2 packages), and the project has no separate build step. We cloned commit e1c44bd into a clean Debian container with 3 CPUs and no project-specific setup.

    Does cua have tests you can run?

    Not through a standard command: the project exposes no test script or target that our harness could run.

    Who should not use cua?

    Buyers who require this review to demonstrate successful desktop tasks, VM startup, or agent task completion: our run tested none of those workloads.

    What are the alternatives to cua?

    Browser Use, E2B. Cua deserves a component-level trial when an agent needs desktop controls, local VMs, or computer-use evaluation tools.

    Setup2/5On the visible setup rubric, a successful install with product setup unverified earns 2. Our root install completed, but build and test were skipped and no desktop or VM workload ran.
    Docs5/5On the visible documentation rubric, component quickstarts, platform requirements, explicit support boundaries, and licensing guidance earn 5. The supplied README and platform reference contain these; their operational accuracy was not tested.
    Community5/5On the visible participation-route rubric, multiple contact routes, contribution guidance, and private security reporting earn 5. The README documents these routes; adoption, response times, and satisfaction remain unassessed.
    Maturity3/5On the visible maturity rubric, documented usable paths with material experimental boundaries earn 3. The platform reference identifies experimental routes and remaining evidence gaps; our run adds no operational validation.

    Discussed on

    1. hnApple Silicon and macOS VMs: Faster LLM Inference with llama.cpp307 points
    2. hnShow HN: Drive any macOS app in the background without stealing the cursor192 points
    3. hnLaunch HN: Cua (YC X25) – Open-Source Docker Container for Computer-Use Agents172 points
    4. hnShow HN: Lumier – Run macOS VMs in a Docker159 points
    5. hnShow HN: CUA-S1 – A System One Model for Computer Use95 points

    Who it’s for

    Developers connecting an existing agent to desktop applications through Driver's CLI, MCP connection, or SDKs.
    Apple Silicon users evaluating Lume for local macOS or Linux VMs.
    Teams building computer-use evaluation tasks with Cua Bench.
    Teams prepared to test their actual application, operating system, and permissions before deployment.

    Who it’s NOT for

    Buyers who require this review to demonstrate successful desktop tasks, VM startup, or agent task completion: our run tested none of those workloads.
    Teams expecting uniform background input across browsers and Linux compositors: the current platform reference documents refusals and experimental routes.
    Users installing the current Cua Spaces macOS app on a system older than macOS 26.
    Providers requiring every selected component to permit a competing hosted service under MIT today: current Spaces components and Cua Driver Bench have different terms.

    Setup reality

    In our run on 2026-09-27 at commit e1c44bd, the harness recorded a successful 14-second root install. The matching sandbox record identifies a Node 22 image with 3 CPUs and 8 GB of memory. Build and test were separately marked skipped with the reason "no script"; neither was a failed command.

    The harness stores command outcomes separately: install success, elapsed time, exit code and timeout status, followed by build and test outcomes. Here the install exit code was 0 and it did not time out. Those results establish command completion, not a working Cua deployment.

    The attached dependency summary records 2 packages and 9 MB. The retained install tail instead says "Packages: +1" and "Done in 1s". These are different recorded views of the step; we preserve the stored harness values rather than replace them with the tail. No complete-log reconciliation or new execution was performed.

    Select a component quickstart before installing. The current README distinguishes Driver permissions, Spaces requirements, Lume's Apple Silicon runtime, and Bench's Python setup. Those current instructions do not describe what our September install exercised.

    What our run establishes

    In our run on 2026-09-27 at commit e1c44bd, the harness recorded a successful 14-second root install. Build and test were skipped because the detected root scripts contained neither target. The matching raw record also detected a tests directory and Python tests, so the skipped root test step does not establish that Cua lacks tests.

    The sandbox used the lab-node:22 image with 3 CPUs and 8 GB of memory. Our harness records each step separately, including success, elapsed time, exit code, timeout status, and skip reason where present. The root install returned exit code 0 without timing out. We did not test Driver actions, VM startup, cloud provisioning, model inference, or agent task completion.

    The stored dependency summary reports 2 packages and 9 MB, while the retained command tail reports one package added and its own shorter duration. We have not reconciled those observations against a complete log. The stored checkout measurements are 5,266 files, 959,948 source lines, and 320 MB; they describe repository size, not installed product coverage. The audit result is null, so this run supplies neither an advisory result nor known scan coverage.

    The MrKeyoor testing methodology explains how command results reach our reviews. It is provenance context, not evidence that its current procedure was used unchanged for this older run.

    Pick the component that answers your task

    Cua Driver provides tools to inspect and operate native desktop apps and browsers on macOS, Windows, and Linux. It connects through a CLI, MCP, or typed SDKs. Bring your own agent and model; Cua's documented role is to supply computers and automation tools.

    Lume manages local macOS and Linux VMs on Apple Silicon. Cua Bench builds evaluation tasks and exports trajectories. Its documented simulated starting task requires no VM, Docker, or model API key. These are separate entry points, so completing a root package install does not validate any of them.

    The README at our measured commit presents Cua Fleets for isolated cloud desktops. The current README instead leads with Cua Spaces and also documents the SDK and CLI. Current Spaces requirements and licensing should therefore guide a current evaluation, without being attributed to our September run.

    Check background control against your actual platform

    The current Driver platform reference makes support conditional on the application and delivery route. macOS needs Accessibility and Screen Recording permissions. Windows elevated targets can refuse input. Linux support depends on the display session, compositor, and application backend.

    For Wayland, the reference distinguishes Sway coverage from experimental KDE and Hyprland routes. Semantic background actions can work where applications expose them, while raw input to an occluded window generally cannot. A broad operating-system label is insufficient to choose a deployment target.

    Use the documented Calculator exercise as a first check: ask the agent to compute 6 × 7 and verify that Calculator displays 42. That is a proposed next step from the README, not an exercise completed in our sandbox. Follow it with a representative task in your own application, including the background behavior you need.

    Compare the layer you need

    Browser Use offers a Python library for running its browser agent locally with a choice of model and a local or cloud browser. Consider it when the task stays in a browser and you want that agent layer. Cua Driver instead supplies desktop controls to an agent you bring. The two may serve different parts of a system; browser automation alone does not establish native desktop coverage.

    E2B supplies cloud sandboxes for AI-generated code and a separate Desktop SDK for mouse, keyboard, screenshots, applications, and desktop streaming. Consider E2B when evaluating cloud execution or desktop infrastructure. It overlaps with Cua's sandbox and desktop needs, so dismissing it as code execution alone would miss a documented capability. These alternatives were not installed or compared in our run.

    Check the current component license

    The current README places the SDK, CLI, Driver, and Lume under MIT unless a directory states otherwise. Spaces components and Cua Driver Bench use FSL-1.1-MIT, with restrictions on competing hosted services and conversion of each release to MIT after two years. Optional perception artifacts carry additional terms. Check the exact component and artifact you intend to ship; the root license signal in our stored record is insufficient for that decision.

    How to read our ratings

    These are editorial ratings, not benchmark results. Setup ease uses 1 for a blocked install, 2 for an install that leaves product setup unverified, 3 for a verified component setup, 4 for a documented repeatable setup, and 5 for verified setup across the evaluated targets. Cua receives 2 because our root install completed while product setup remained untested.

    Documentation uses 1 for missing entry instructions, 2 for basic installation notes, 3 for component quickstarts, 4 for quickstarts plus platform requirements, and 5 for those materials plus explicit support boundaries and licensing guidance. The supplied current README and platform reference meet the fifth category; this does not mean we executed their instructions.

    Community uses 1 for no documented participation route, 2 for one route, 3 for multiple routes, 4 for those routes plus contribution guidance, and 5 for all of those plus private security reporting. The supplied README meets the fifth category through its Discord, issues, contribution, and security links. This rates the documented participation routes, not adoption, responsiveness, or user satisfaction.

    Maturity uses 1 for a concept without a usable path, 2 for an early implementation, 3 for documented usable paths with material experimental boundaries, 4 for demonstrated coverage across the evaluated scope, and 5 for that coverage plus operational validation. Cua receives 3 because the current documentation describes usable paths alongside experimental routes and open evidence gaps. Our run provides no production validation.

    Alternatives

    ProjectWhat it isPick it when
    Browser Use gh↗A browser agent with a local Python library and hosted options.Choose it for browser tasks when you want an agent implementation with model choice and local or cloud browsers. It can complement a broader desktop system; this evidence does not establish it as a replacement for native desktop controls.
    E2BCloud sandbox infrastructure with code execution and a separate Desktop SDK.Evaluate it for cloud code execution or desktop workloads. Its Desktop SDK makes it a relevant overlapping option, rather than merely a complement for code execution; our run did not compare deployment behavior.

    What people are saying

    1. [velocity-scout] shhivv/arc-cua
    2. [github-trending] trycua/cua

    Sources

    1. Cua repository and README
    2. Cua Spaces v0.6.0 release
    3. Cua Driver platform support
    4. Wayland-native window discovery issue 2145
    5. Windows password-field metadata issue 3576

    More ai tools reviews