phone-eye

Your AI agent writes code and runs tests, but it still cannot touch a real phone screen. Game canvases, popups, and dynamic images are invisible to the accessibility tree, while pure-screenshot approaches keep mis-hitting coordinates. phone-eye solves this by fusing a vision model with the Android UI tree, so the agent understands what is on screen and also knows exactly where each element lives.

Core capabilities include: capturing and interpreting the live screen, tapping and swiping at precise coordinates, ASCII text input, pressing hardware keys (wake, back, home, recents), launching arbitrary pages via Android intents, and saving screenshots to disk. After a single USB connection the tool switches to Wi-Fi adb automatically, freeing you from the cable for good. Multiple phones can run in parallel, each in its own process.

Built for developers using MCP clients such as Claude Code, Codex, or Cursor who need mobile web QA, form regression checks, OEM setup-wizard automation, or unattended night-watch monitoring. No app to install on the phone, no root, no extra server. Bring any vision-capable model (OpenAI, GLM, local Qwen-VL, and more), spend sixty seconds granting developer options, and hand the rest to your agent.