TuriX-CUA: five branches for one desktop agent
This is the official website for TuriX Computer-use-Agent
At a glance
- What is it?
- An open-source computer-use agent that drives a desktop through clicks rather than app-specific APIs, with the model policy swappable from a config file. What is worth reading closely is the branch structure and a dependency list that does not match the privacy claim.
- Who is it for?
- TuriX fits someone who needs an agent to work across desktop applications that expose no API, and who is willing to grant accessibility permissions on macOS. Two things to weigh first.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 62 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- TurixAI/TuriX-CUA3,177 ★ · +19
| Repository | 2026-09-10 | 2026-11-15 |
|---|---|---|
| TurixAI/TuriX-CUA | 3.2k | 3.2k |
19 Stars in the last 66 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
Five branches, each a different configuration
The quick start opens with instructions to switch branches before anything else, and the list is the clearest description of the project there is. Windows users are told to check out the multi-agent-windows branch for Windows-specific setup and installation. Linux users check out multi-agent-linux.
git checkout multi-agent-linuxTwo more branches exist for people who want the previous setups: windows_legacy for the earlier Windows configuration and mac_legacy for the earlier single-model macOS one. The default branch is the third option, a macOS-focused setup that also carries the newest lightweight skill package. So the same repository holds five supported arrangements, and the choice is made by which branch you check out rather than by a configuration flag. Anyone automating a clone needs to make that branch selection explicit, since the default is only the right answer for recent macOS hardware.
The benchmark claim has three separate parts
The performance section states more than a single number, and the parts do not measure the same thing. The headline figure is 64.2%, written as 229.88 out of 358, on the full OSWorld benchmark, described as third overall among submitted agents at a fifty-step limit. A fractional numerator in a task count is unusual, which suggests the published score is normalised rather than a raw tally of solved tasks. Second, the project reports more than 80% on a Mac benchmark it hosts itself, described as OSWorld-style, and pairs that with the claim that the agent is built and optimised for macOS. Third, it states that zero Linux training data was used while still finishing in the top three on OSWorld's Linux-based environment. The README points to a technical report for conditions and limitations, which is the place to check what those three numbers share.
The dependency list complicates the privacy claim
The README states plainly that no data is ever collected, and the requirements file lists a client for a product analytics service. Alongside it are other packages that describe a network-instrumented application rather than an offline tool. Both the old and the new name of the same DuckDuckGo search package appear, one pinned and one not. A tracing package with LangChain extras is also present, which is normally used to record model calls and their costs. The rest of the list explains what the agent does rather than how it reports: screen and input automation libraries, an image library, fuzzy matching with two overlapping implementations and a Levenshtein helper, HTML extraction with BeautifulSoup and a separate main-content extractor, HTTP clients, Playwright for browser work, a macOS bridge, and a Gradio interface. Seven separate LangChain provider integrations cover OpenAI, Anthropic, Ollama, Fireworks, AWS, and Google, which is the code behind the promise to swap models.
No app-specific APIs is the whole design
The feature table states the central constraint in the plainest possible terms: if a human can click it, the agent can too, with WhatsApp, Excel, Outlook, and in-house tools given as examples. Everything else follows from that. Hot-swappable brains means the vision-language policy can be replaced by editing a config file, with no code change, which is why the README opens its setup section with an invitation to change that file and go. Model Context Protocol support lets the agent be called from another tool rather than only from its own interface, and the demo shows a different assistant searching for news, calling this agent, and writing the result into a document. Skills are markdown playbooks with a deliberate split of labour: a planner selects relevant guides by name and description, and only the brain reads the full instructions while planning each step. That keeps the context small without reducing the playbook.
The same skill ships in two forms and on two branches
The agent has an OpenClaw integration in two layers. A published skill is available from a public skill registry under a different account name from the organisation, and local copies live in a directory inside the repository. On the default branch that directory holds the macOS package, a skill document plus a shell script that launches the agent. On the Windows branch the package has three files instead of two: the same document, a PowerShell launcher rather than a shell one, and an agent configuration file. The README says installation and permissions are documented in that directory's own readme, which means the per-platform instructions live outside the main file. A March update records what changed in the Windows package, adding direct dispatch, safer pre-flight checks, and that agent configuration file, which is the sort of change worth reading if you are pinning a branch.
A 3.0 alpha split the product in two
The news section shows a project that branched into two directions. An April announcement introduces a 3.0 alpha described as an all-in-one productivity application that combines the computer-use agent and a command-line interface in one workflow, adding two named capabilities: one for everyday office execution and task orchestration, and one for coding, automation, and engineering tasks. The stated idea is that you can execute with command-line precision and close the loop through graphical actions in a single flow, which is a different pitch from clicking on behalf of a person. Earlier entries record a 0.3 release with a search integration, a local model runtime, recoverable memory compression, and skills, followed by a move to a multi-model core architecture and added vision-language model support including one Chinese open-weights family. Version 0.4 was tagged in March 2026.
The repository describes itself as a website
One inconsistency is worth naming because it shapes first impressions. The repository description says this is the official website for the agent, while the tree holds what is plainly a code project: a source directory, a skills directory, a documentation directory, an examples directory, a requirements file, and the OpenClaw skill packages. There is a website, and the homepage is a different domain from the download page used in the news items, so a reader following both ends up on two properties. The examples directory holds a config file and a single entry script, and it also contains a macOS metadata file that was committed by accident, which is the kind of detail that tells you the repository is maintained by people working quickly. Requirements name Python 3.12 in their header, the quick start requires macOS 15 or newer, and the permission steps are accessibility plus Safari automation, which is the pair an automation tool needs on that platform.
Editorial conclusion
TuriX fits someone who needs an agent to work across desktop applications that expose no API, and who is willing to grant accessibility permissions on macOS. Two things to weigh first. Choose the branch deliberately, since five of them carry different platform and model configurations and picking the wrong one costs more than reading a paragraph of documentation. And check the telemetry question yourself, because the README states that no data is collected while the requirements file includes an analytics package.
Frequently asked questions
What is TuriX-CUA?
An open-source computer-use agent that lets AI models take hands-on actions on a desktop. It works without app-specific APIs, so anything a person can click can be driven, and it is MIT licensed and free for personal and research use.
How do I run TuriX-CUA on Linux or Windows?
Switch branches first. Linux users check out the multi-agent-linux branch and Windows users the multi-agent-windows branch, each of which carries platform-specific setup instructions. Branches named windows_legacy and mac_legacy hold the previous setups for those platforms.
Can I use a different model with TuriX-CUA?
Yes. The model policy is described as hot-swappable, and the change is made in config.json without touching code. The requirements include LangChain integrations for several providers, including OpenAI, Anthropic, a local Ollama runtime, Fireworks, AWS, and Google.
What permissions does TuriX-CUA need on macOS?
The quick start is written for macOS 15 or newer and covers two permissions: accessibility, and Safari automation. The README also states that no data is collected, though the requirements file includes an analytics client, so check that yourself before deploying.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/turixai-turix-cua)Star history chart for your README
[](https://hysenlabs.com/projects/turixai-turix-cua)Was this useful?