Agent S is Simular AI’s open framework for computer-use agents: AI systems that operate a real desktop through its graphical interface, seeing the screen and acting with mouse and keyboard the way a person does. It has earned about 12,700 stars on GitHub and is actively developed through 2026.
Most automation breaks when an application has no API. Agent S addresses the general case โ any software with a screen โ by giving a multimodal model a structured way to perceive and act. Its Agent-Computer Interface abstracts screenshots and UI grounding, so the model can locate buttons and fields reliably, while a planning layer with memory and retrieval breaks long tasks into steps and recalls how similar tasks were solved before. The framework is model-agnostic at the perception and reasoning layers, and versions of Agent S have posted leading results on the OSWorld benchmark for desktop task automation. Newer work adds in-context reinforcement learning so the agent improves from its own trajectories. Because it is a framework rather than a single product, researchers use it to test grounding and planning ideas, and developers use it as the skeleton for GUI automation products.
Agent S is free and open source under the Apache-2.0 license. Simular AI separately offers commercial agent products, but this framework costs nothing to use; you supply your own models and machines.
Agent S is for researchers and developers building GUI automation โ software testing, desktop workflows, legacy application control โ who need an open, extensible foundation for agents that see and operate real computer interfaces.