Apps · PROJECT NOTES
AgentNet (OpenCUA)
Cross-platform computer-use demonstrations from the OpenCUA project, accompanied by recording tools and an offline action-matching evaluator.
What can you do with it?
Train GUI agents on screenshot/action trajectories from Windows, macOS, Ubuntu and web interfaces. Use AgentNetBench to compare predicted low-level actions with annotated actions, including position, typing, key and scroll matching.
What should you know before using it?
This entry covers data and evaluation tools, not the OpenCUA model variants or a new live desktop environment. Human demonstrations are supplemented with generated descriptions, reasoning and quality labels. AgentNetBench scores recorded actions rather than live task completion and expects its own trajectory schema; the entire training corpus is not a verified held-out test split. Training code remains a release TODO. Code and data carry MIT licenses, with separate use/privacy/IP notices in the card. Platform tags describe recorded interfaces.
What is verified here?
Runtime: Static JSONL plus multipart screenshot archives. The public offline evaluator under evaluation/agentnetbench documents Python 3.9+, Pillow and an OpenAI-compatible model endpoint; recording tools are a separate AgentNetTool submodule.
License note: OpenCUA code and AgentNet LICENSE.txt use MIT. The dataset card separately states prohibited-use, privacy and third-party intellectual-property notices.
This entry was reviewed against original project material. We have not independently run the software or reproduced its results. AI-development claims are author-reported; check upstream for current versions, setup requirements, and licenses.
- Resource type
- Dataset + benchmark
- Provenance
- Development provenance not established
- Reviewed
- Platforms
- Windows, macOS, Linux, Web