Agent-S

Why we included this project

Agent S is an open framework for computer-use agents that operate real desktop applications through their own interfaces instead of relying on an API. A multimodal model reads the screen, while explicit planning and memory keep the agent on task and grounding anchors its actions to what is actually visible, so it can click, type, and navigate native GUIs on Windows, macOS, and Linux directly. That makes it a practical base for automating legacy or closed-source programs with no API, and for building general computer-use assistants. The project is backed by peer-reviewed research on GUI agents and ships evaluation tooling against benchmarks like OSWorld, so you can measure an agent's performance rather than just watch it run. Developers wanting a foundation for building or benchmarking desktop automation agents will find working code and a structured approach here, not a thin demo.