agent-device

agent-device gives a coding agent a live app to poke at. It drives iOS, Android, and HarmonyOS across simulators, emulators, and physical hardware, plus tvOS, Android TV, web, macOS, and Linux, through a CLI, a built-in MCP server, or a typed Node.js API. MIT, from Callstack, the consultancy behind a lot of the React Native ecosystem.

The design choice that matters is the accessibility tree. Rather than feeding an agent screenshots and hoping it reasons over pixels, agent-device hands it a text snapshot with element refs, so the agent presses the element it was handed, at-e2, instead of guessing at coordinates. Cheaper in tokens and far more reliable. Install is a global npm install on Node.js 22.12 or newer, and the bundled doctor command checks your setup before you hand the keys to an agent. It also arbitrates device ownership across parallel agent worktrees, which is the problem you hit the moment two agents want the same simulator.

Solo devs and small mobile teams get value immediately, and it plugs into Claude Code, Codex, Cursor, Windsurf, Cline, and Goose with no custom work. Bigger teams can point the same commands at BrowserStack, AWS Device Farm, or Limrun instead of local hardware. Free at every size.

The catch is that this is glue, not a device farm. It drives hardware you already own or a cloud you already pay for, so iOS still wants a Mac with Xcode and the cloud path still wants a cloud bill. Your app also has to be accessible: snapshots come from the accessibility tree, so if labels, roles, and test IDs are a mess, the agent is back to squinting at screenshots.