tongyi-mai/opengcua, published by Alibaba Tongyi-Mai, is a general-purpose computer use agent stack. The latest release on the GitHub repository is v0.1.0. It does not stop at directly manipulating the GUI: the tasks of reading accessibility trees and calling app APIs and shell commands are written as code and composed inside a single program. It is released under the Apache-2.0 license.
The aim of this approach is to handle several kinds of computer operation in one execution flow. The project reports that on the offline OSWorld-v2 set, the average partial score of 8 frontier models rose from 45.3 to 64.9, up 19.6 points, while the average number of steps per task fell by 53%. The gains were larger for weaker models: GLM-5.3-Flash gained 35.2 points and DeepSeek-V4.1-Flash gained 29.8 points. In a comparison that fixed the GUI and shell access conditions, the code-based way of handling the GUI raised the score by 6.1 points and cut the step count by 14.8%.
The catch is that the performance figures exist only as result SVG images, not as text tables, which makes them hard to verify against the text. In four tasks on real macOS devices, Qwen3.8-Max reportedly used 19.4 times the reasoning tokens of GPT-6 Astra yet completed 3 tasks. The macOS harness and the cross-device SDK are labeled as preliminary steps, and the GCUA model itself has not been released either. So the current public material is closer to a research preview that includes evaluation data and some tooling than to a finished product.