A browser for AI agents, but make it mobile

What it does

agent-device is a CLI that lets AI agents open, inspect, and interact with real mobile and desktop apps. It returns compact accessibility snapshots with semantic refs like @e3, handles taps and gestures, and captures evidence—screenshots, logs, traces, network traffic—only when needed. Think of it as Vercel’s agent-browser, but for the app store instead of the web.

The interesting bit

The project is explicitly not Appium or Detox. It is optimized for token efficiency: agents get sparse, actionable UI refs rather than dumping full view hierarchies, and they can record exploratory sessions into replayable .ad scripts for CI. The README also notes a typed Node client and public subpaths like agent-device/android-adb for bridge integrations, suggesting the CLI is a thin orchestration layer over platform-specific backends.

Key highlights

• Supports iOS/tvOS (XCTest), Android (ADB), macOS desktop, and Linux (AT-SPI) targets
• Works with native apps, React Native, Expo, and Flutter
• Evidence capture is on-demand, not a firehose
• Records sessions to .ad replay scripts for CI
• Used by Callstack, Expensify, Shopify, and others

Caveats

• Requires platform-specific toolchain setup (Xcode, Android SDK, macOS Accessibility permissions)
• GitHub Actions template is listed as “coming soon”; only an EAS workflow template exists now
• Cloud/remote execution routes through Callstack’s commercial offering

Verdict

Worth a look if you’re building agentic workflows that need to verify actual app behavior on devices. Skip it if your testing stays in headless browsers or you need a mature, framework-agnostic community ecosystem today.