Control an Android phone from any skill-capable AI agent (Claude Code first): plug the phone in over USB, install this skill, and the agent can see the screen and drive the UI through adb — semantic UI-tree first, screenshots as fallback, CJK input included, optional live mirroring so you can watch every tap.
Status: v0.1 — working. Field-tested end to end on a real device against Taobao (search, product research, report generation). See PLAN.md for architecture and roadmap, TECH-REVIEW.md for the technology-selection review and field-test results.
git clone https://github.com/ReScienceLab/android-use.git
cd android-use && ./install.sh # symlinks into ~/.claude/skillsRequires adb on PATH (macOS: brew install android-platform-tools) and python3.
Optional: scrcpy for the mirror command. On the phone: enable USB debugging
(see skills/android-use/references/setup.md).
One stdlib-only CLI (skills/android-use/scripts/cli.py) behind a ~100-token skill:
| Command | Effect |
|---|---|
snapshot [--settle] |
ref-annotated semantic UI tree of the current screen |
tap <ref> / tap <x> <y> |
tap an element or coordinate (--long to long-press) |
type "text" [--clear] |
type into the focused field; CJK/emoji via bundled yadb |
scroll / swipe / key |
gestures and keyevents |
screenshot [out.png] |
PNG of the screen for visual fallback / report evidence |
app <pkg> [--stop], app --list F |
launch (verified) / find packages |
mirror [--off] |
live scrcpy window with visible touch indicator |
devices, selftest |
plumbing |
Design notes: no daemon (adb server is the daemon), no on-device agent to install,
state between calls is one refmap JSON under ~/.cache/android-use.
Apache-2.0. The bundled bin/yadb dex is from YADB
(LGPL-3.0) — see NOTICE.
The banner image includes the Android robot, which is reproduced or modified from work created and shared by Google and used according to terms described in the Creative Commons 3.0 Attribution License.