Agent Device

Agent Device is an open source device automation tool from Callstack that lets AI coding agents operate and verify real apps on iOS, Android, TV and desktop targets. It applies the patterns that made browser automation practical for web agents to native mobile and TV apps, so an agent can open an app, read the screen, act on it and confirm that a change works. It is aimed at mobile developers and QA teams who use coding agents and want them to test their work on simulators, emulators and physical devices. A snapshot command exposes the accessibility tree with stable element references, keeping context far smaller than raw screenshots, and agents can act through those references or through semantic selectors that find elements by text, label or role. One workflow covers development simulators, QA devices and TV class targets, and the project also lists HarmonyOS, web, macOS and Linux support. Exploratory sessions can be saved as scripts and replayed deterministically, with an experimental mode that heals stale selectors. When something breaks it captures parsed logs, network dumps and screen recordings. It ships as a CLI installed from npm, an MCP server and a typed TypeScript client for custom or cloud agents, plus official agent skills, including an exploratory QA skill, and a proxy for agents running in cloud sandboxes. Agent Device is free under the MIT license, and no paid tier is published.