Ukrainian MacPaw has taken another step towards «independent» artificial intelligence — the team introduced the GUIrilla platform, which allows AI to interact with macOS interfaces as if it were a human. The development was presented at the DATA-FM workshop within the framework of ICLR 2026 — one of the key global conferences in the field of machine learning.
Ukrainian MacPaw has taken another step towards «independent» artificial intelligence — the team introduced the GUIrilla platform, which allows AI to interact with macOS interfaces as if it were a human. The development was presented at the DATA-FM workshop within the framework of ICLR 2026 — one of the key global conferences in the field of machine learning.
GUIrilla essentially teaches a model to «understand» interfaces: find buttons, fill out forms, navigate between elements, and perform actions without user intervention. This approach could become the basis for a new generation of desktop agents.
According to DOU Research Manager at MacPaw Maria Girna, the key problem is the lack of high-quality open data for training. Most existing datasets, she says, barely cover macOS. Girna notes that in the largest open dataset for desktop AI — OS-ATLAS — the share of macOS is only 0,06%.
«Large companies may have their own proprietary data collected manually, but these resources are not available to the academic community and independent researchers. Collecting everything manually is expensive and practically impossible to scale,» the researcher explains.
This is exactly the problem the team set out to solve. GUIrilla automates the process: it independently installs applications, interacts with them, takes screenshots, and collects structural interface data.
It’s not just a «picture» of the screen. According to Girna, the system generates a text description of the interface — what elements are present, what they do, and how they are related to each other. This is essentially the same data used in Apple’s accessibility features, but now it’s used to train AI.
As a result, the MacPaw Research team collected the GUIrilla-Task dataset — over 27,000 interface interaction scenarios in over 1,100 macOS applications.
In parallel, the company opened the macapptree library, which allows developers to obtain structural information about any Mac application with literally a few lines of code.
«This many macOS interfaces have never been made publicly available before,» adds Girna.
MacPaw does not hide that GUIrilla is part of a broader strategy. As Director of AI and Research Serhiy Kryvoblotsky notes, developing our own research and creating complex technological IP is the key to competing in the global AI market.
In fact, the company is investing in the foundation — data and tools, without which it is impossible to create full-fledged AI agents for working with the desktop.
