Computer use · agent infrastructure component
Cua: review, pricing, security and alternatives
Cua is an open-source agent infrastructure component published by Cua, in the computer use family. Cua provides drivers, isolated environments, and benchmarks for developing and evaluating computer-use agents. Its repository has 30,856 stars and its last commit is from October 24, 2026.
What Cua does
Cua belongs to the computer use family: agents that see the screen and drive the keyboard and mouse of a computer or phone. On our grid it ranks #2 of the 17 projects of this family listed in the directory.
It is not an agent in itself: it is a component other agents use (memory, isolated execution, web access, model gateway or monitoring). You pick it based on the agent or framework it has to equip.
You use it from: developer SDK and API. It can run on your own machine or server.
Who is Cua for?
It is for people who want to automate desktop software with no API, by driving the screen as a person would. Typical tasks people give it:
- chain actions in legacy business software
- copy data from one application to another
- test a desktop application
It is a good candidate if your data must stay on your machines.
Access and human control
According to its documentation, Cua can:
- control the screen, keyboard and mouse
We could not establish whether it asks for human approval before sensitive actions: test it first on an unimportant account or folder.
Cua pricing
The code is free, under the MIT licence.
Getting started with Cua
- Install it: the project ships as a compiled Rust binary (direct download or
cargo install). The exact command is in the repository README. - Start with the minimal example from the docs, then add one tool at a time and check what the agent does with it.
- Give it a short, checkable task, review the result and its action log, then raise the difficulty step by step.
Security: risk level
An agent that controls the computer has the same rights as you: run it in a dedicated virtual machine.
Points found for Cua:
- Sensitive access: it can control the screen, keyboard and mouse, with no established default human approval.
Three habits: least privilege (minimal rights, dedicated keys), human approval for irreversible actions, a sandbox for code execution. The AI agent security check reviews your setup in a few minutes.
Project activity and releases
Created on January 31, 2025, the trycua/cua repository has 30,856 stars, 2,250 forks and 667 open issues, with 5,863 commits in total. It received changes this month.
Over the last 52 weeks: 3,886 commits, an average of 74.7 per week.
801 releases in total; 21 in the last twelve months. Latest release: cua-spacesd-v0.8.0, on October 24, 2026.
Cua vs other computer use
GitHub stars of the best-rated open-source projects in the computer use family. Stars measure attention, not quality: a heavily promoted project collects them fast, a specialised tool will have few.
Cua and other computer use at a glance
The 7 best-ranked agents in the same family, 7 of them open source. The score out of 100 only applies to open-source projects; for proprietary services we show the published entry price. A dash marks information the documentation does not state.
| Agent | Model | Score | Stars or entry price | Self-hostable | Human approval |
|---|---|---|---|---|---|
| Cua | open source | 80 | 30.9K ★ | yes | – |
| Midscene | open source | 81 | 15.4K ★ | yes | – |
| Mobilerun | open source | 79 | 9.8K ★ | yes | – |
| GenericAgent | open source | 73 | 14.3K ★ | yes | – |
| OpenBot | open source | 72 | 7K ★ | yes | – |
| UI-TARS Desktop | open source | 71 | 39.4K ★ | yes | – |
| OpenMausBot | open source | 71 | 5.5K ★ | yes | – |
| Observer | open source | 70 | 1.7K ★ | yes | – |
How we rated it
Score: 80/100, computed only from public, verifiable data, with no subjective testing or partnership:
- Adoption 30/40: 30,856 stars on a log scale (40 points from about 200,000).
- Activity 30/30: last commit less than a month ago, 21 releases in six months, 3,886 commits in a year.
- Framework 20/30: licence, default human approval, self-hosting, model choice and documentation site.
The score does not say whether the agent will succeed at your tasks: it measures how solid the project is. The full grid is on the method page.
Verdict
Cua is one of the most solid projects in its family: followed, active and well framed.
Strengths
- Widely followed: 30.9K stars
- Very active development
- Frequent releases (21 in a year)
- Can run on your side: data under control
- Permissive licence (MIT)
Limitations
- Broad access without default approval
Avoid if…
- You cannot isolate it (container, virtual machine)
Alternatives
- Midscene15.4K ★
- Mobilerun9.8K ★
- GenericAgent14.3K ★
- OpenBot7K ★
- UI-TARS Desktop39.4K ★
- OpenMausBot5.5K ★
Learn more
Frequently asked questions
Is Cua free?
The software is free (MIT licence). You still pay for the model calls, unless you use a local model.
Is Cua safe?
Our risk level: high. Points found: sensitive access. Give it minimum rights and approve irreversible actions.
Which models does Cua use?
Its documentation does not specify the supported models.
Can Cua be installed on your own server?
Yes, from the trycua/cua repository. Data then stays with you, but security and updates are your job.
What are the alternatives to Cua?
In the same family: Midscene, Mobilerun, GenericAgent, and OpenBot.
Cua or Midscene: what is the difference?
Both belong to the computer use family. Cua has 30.9K stars against 15.4K for Midscene. The right choice mostly depends on the access you are willing to give it.
Related articles
- Open source or proprietary AI agent: how to decideOctober 25, 2026
- AI agent vs. chatbot: what's the difference?October 24, 2026
- Autonomous AI agents: how much should you let them do?October 12, 2026
- AI agent memory: how agents remember (and forget)October 11, 2026
Neutrality. No affiliate links. No vendor pays to be reviewed, rated or ranked. Reviews are directed by François Chollet, neutral and factual, and every data point is dated and sourced. How we score