OpenGCUAOpen Generalist CUA
← Back to OpenGCUA

BUILD WITH US

Contributing to OpenGCUA

Open Generalist CUA is a community project, and we welcome contributions from everyone. Whether you are a developer, a researcher, or simply interested in computer-use agents, there are many ways to get involved.

01

Feedback and suggestions

  • Usage feedback on the harness and the SDK, and bug reports. Please attach the trajectory when reporting a failure, it makes the report actionable.
  • Suggested research questions worth exploring for GCUA. Open an issue with the research-question label. A good issue states the question in one sentence and why an open stack is needed to answer it.

Questions we are already tracking, join these or add your own:

  1. Training harness-to-model decision transfer: when are todo, intent, and screenshot worth using?
  2. Letting the model decide what to keep in context, instead of fixed harness policies.
  3. Efficiency on real devices: where the cost comes from, and how the model and the harness can reduce it.
  4. Using accessibility trees in programs: how AX access changes the programs models write.
  5. When should an agent ask the user instead of acting?
02

Code contributions

  • The harness and the SDK themselves.
  • Platform adapters for Windows and Linux, alongside the existing macOS and Android adapters.
  • SDK bindings in languages other than Python and JavaScript.
  • Driver improvements, for example element identity on macOS and multi-display support.
  • Integrations with other agent frameworks. The SDK is model-agnostic by design, so running another model through it is itself a contribution.

Open a PR. Small, reviewable changes land fastest. For anything that touches the driver or the confirmation gate, include a trajectory demonstrating the behavior.

03

Evaluation

  • New tasks for the real-device benchmark, especially on applications we have not covered.
  • Replication of reported numbers under different configurations.
  • Studies based on the released trajectories.
  • Real-device runs on diverse hardware and macOS versions, and permission-flow reports.

Community results are featured in each numbered update.

04

Ground rules

  • Every reported number must ship with the harness configuration and step budget that produced it.
  • No data distilled from closed models in any contribution intended for training.
  • Label contributions honestly as released or planned, like we do.
GitHub