callstack agent-device
COMMUNITYLABSCO SUMMARY
The agent-device skill teaches an agent to drive that CLI against real apps: opening iOS, Android, tvOS, macOS, Android TV, Linux desktop, and Amazon Vega OS TV targets, inspecting the current screen through structured accessibility snapshots with interactive refs like @e3, and acting on it — tapping, typing, scrolling, gestures, alerts. It also covers evidence capture (screenshots, video, logs, network traffic, audio-level probes, performance samples, crash context, React profiles) and replay, either recording .ad scripts for CI or exporting to Maestro YAML. The second skill, dogfood, sits directly on top of the first: it drives the same CLI to systematically explore an iOS or Android app looking for bugs and UX issues, rather than running a scripted test path.
This is for a team already shipping a native iOS/Android, Expo, Flutter, or React Native app who wants a coding agent to verify a change on a real simulator, emulator, or device instead of only reading the diff. The README frames it explicitly as the mobile/TV/desktop counterpart to Vercel's agent-browser, down to reusing agent-browser for the CLI's own minimal --platform web mode rather than reimplementing browser automation.
READ THE FULL ANALYSIS
What it needs before either skill does anything. The underlying CLI requires Node.js 22+ (24+ for web), Xcode for iOS/tvOS/macOS targets, the Android SDK plus ADB for Android, and a macOS Accessibility permission grant for desktop automation — none of that is optional scaffolding, it's what XCTest, ADB, and AT-SPI need to drive a real device. Both skills carry an unverified dependency label in our data, so treat that requirement list as read from the README rather than confirmed against the skill files themselves.
Two more rows exist for this repository but are dead. One is a marketplace-listing entry with no skill content of its own; the other, react-devtools, has no description in the repository at all. Neither is counted or described above.
The CLI itself is MIT-licensed and free to run locally; Callstack also sells a separate "Agent Device Cloud" for remote execution on managed devices, which is a paid product distinct from the open-source skills.
WHAT'S INSIDE
2 showing · 2 totalagent-device
A real app on a phone, TV or Mac, driven from typed commands instead of by hand — with each step checked against what actually appeared on screen before the next one runs.
dogfood
Hunting for bugs in a mobile app by actually using it — walking through its screens on a real device and judging it by what it does there, not by what its code says.
HOW TO GET IT
npx skills add callstack/agent-devicenpx skills add callstack/agent-device --skill <name> --full-depthPick the skill name from the Skills tab — each entry there installs independently.