What it is
Lets a coding agent hand off device testing in one instruction.
What it does: Drives a real Android device, then reports with a verifier check.
Why it belongs in the showcase: Clear split of planner, selector and verifier roles.
Caveats: Data leaves the device to two APIs.
Evidence
Public MIT repo created 2026-10-18, installable as a Claude Code plugin. README shows about 3–4 s per step, a pass/fail report plus one report image of every step, exit statuses such as `passed` (verifier p ≥ 0.7), `passed_weak` and `unverified`, and a planner cost of about $0.01–0.04 for a 30-step run. It needs adb, an Android 8.0+ device, ANTHROPIC_API_KEY and TYPESAFE_API_KEY.
Caveats
- Screenshots and screen text are sent to both APIs during a run
- Needs two paid API keys and a real device
- Box on screenshots relies on a Jev preview feature
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.