agent-device

About

Mobile app automation and verification built for AI coding agents

agent-device closes the loop for agents that write mobile code: after a change, the agent opens the app on a simulator, emulator or real device, reads a token-efficient accessibility snapshot, acts on refs like @e2, and checks the diff that comes back after each step. It captures screenshots, video and logs only when needed, coordinates device access across parallel agent worktrees and can connect to remote device clouds. Use it as a CLI, as a stdio MCP server (`agent-device mcp`) or as a typed Node.js API under agents built with the AI SDK. Built by Callstack (the React Native consultancy), it crossed 5k GitHub stars and 1M monthly npm downloads, and powers mobile automation in T3 Code and TesterArmy's e2e. Repo: https://github.com/callstack/agent-device.

Features

• CLI, built-in MCP server and typed Node.js API
• iOS, Android, HarmonyOS, tvOS, Android TV, web, macOS and Linux
• Accessibility snapshots with refs instead of screenshot guessing
• Diff after every action so the agent sees what changed
• Screenshots, video and logs saved as review evidence
• Device locking for parallel agent worktrees and remote device clouds

Install