OpenGCUA: Toward a Fully Open Stack for Generalist Computer Use
Abstract
OpenGCUA introduces an open generalist computer-use agent that integrates visual and programmatic interfaces, harness-to-model decision transfer, and code execution to improve task performance and efficiency across models.
Computer use has reached a turning point: the focus is expanding from GUI-only interaction to systems that complete tasks on users' computers through visual and programmatic interfaces. We define a generalist computer-use agent (GCUA) as a system in which the model chooses how to observe, act, and organize its work across visual and programmatic interfaces. Yet frontier implementations, such as GPT-6 Astra's computer-use system, remain largely closed, limiting systematic study of their design. We introduce Open Generalist CUA (OpenGCUA), an ongoing project to develop and open the full stack for GCUAs with the community. This report presents the first release of OpenGCUA. Code execution is a key feature of our agent design: it lets intermediate results determine subsequent operations and enables models to apply their existing programming capabilities directly to computer interaction. We also propose harness-to-model decision transfer, moving selected observations and task-state decisions from fixed harness rules to the model. On OSWorld-v2, our full configuration improves task scores across all eight evaluated models, including GPT-6 Astra and Claude Opus 5, raising mean partial score by +19.6 points and reducing reported steps per task by an average of 53.0% relative to batched GUI-only execution. We release our implementation and prompts, evaluation trajectories from all eight models, and preliminary versions of a real-device harness and a cross-device SDK. The harness addresses real-device challenges not fully captured by benchmarks, while the SDK provides a common programming interface across computers, mobiles and browsers. Across four real macOS tasks under our real-device harness, selected Qwen3.8-Max runs use 19.4 times as many reported reasoning tokens as GPT-6 Astra, highlighting substantial differences between models in real-device execution efficiency.
Community
🚀 OpenGCUA First Release | Opening the Full Stack for Generalist Computer Use
GPT-6 made computer use the frontier. We're making it open.
The MAI-UI team at Alibaba introduces OpenGCUA (Open Generalist CUA), an ongoing project to develop and open the full stack for generalist computer-use agents (GCUAs), built together with the community.
This first release includes:
- the technical report
- our GCUA design and its harness
- evaluation trajectories from 8 frontier models
- a real-device harness for macOS
- a Cross-Device SDK
- an interactive lab you can try online
🔑 Five Highlights
1️⃣ GCUA: a new paradigm for computer use
Beyond "screenshot → click": the model decides for itself how to observe, how to act, and how to organize its work. It reads the accessibility tree, calls APIs, runs shell commands, and operates the GUI only when needed. GPT-6 Astra's computer-use system is a typical GCUA.
2️⃣ Agent design: code execution + harness-to-model decision transfer
GUI actions are written as code and combined with accessibility, API, and shell operations in a single program, where loops, branches, and intermediate results drive what happens next. Workflow decisions move from fixed harness rules to the model: the harness provides the mechanism, and the model decides when to use it, for example requesting a screenshot only when it needs one.
3️⃣ All 8 frontier models improve
On OSWorld-v2, 8 models, including GPT-6 Astra, Claude Opus 5, and Gemini 3.8 Flash, gain +19.6 points on average while taking 53% fewer steps. The gap between the strongest and weakest model shrinks by more than half. The OpenGCUA harness implementation, prompts, and all evaluation trajectories are open-sourced.
4️⃣ Real-device harness + Cross-Device SDK (Preview)
A macOS menu-bar app brings the agent onto the user's own computer, and we solve and document a range of problems that only appear on real devices. The Cross-Device SDK connects computers, browsers, and phones through one programming interface, so a single program can coordinate several devices within one task. In our real-device tests, Qwen3.8-Max uses 19.4× as many reasoning tokens as GPT-6 Astra; on real devices, the gap between models is even larger.
5️⃣ Interactive lab: every feature of this release, one click away
Seven labs, one for each of seven questions from the report:
- take a real Mac apart in 3D
- compare four ways to act side by side to see what code execution changes
- break down the challenges unique to real devices
- run one program across a phone, a browser, and a Mac
- replay real-device trajectories from five models
- …and more
Works on your phone too.
🔭 What's Next
A GCUA model that unifies computer use and browser use, systematic improvements to execution efficiency, and real-device evaluation environments and benchmarks… Come build it with us!
🔗 Learn More
🧪 Interactive lab: tongyi-mai.github.io/OpenGCUA
📄 Technical report: arxiv.org/abs/2610.22200
💻 GitHub: github.com/Tongyi-MAI/OpenGCUA
👋 Here's a 90-second tour of OpenGCUA.
Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper