I build a mobile app on my own. Unit tests pass, the app works on my phone, and then someone on an iPhone SE with large text finds a button hidden under the keyboard. I wanted something between unit tests and real beta testers, so I built this:
One GitHub Actions workflow that builds your app, starts up to 20 iOS Simulators and Android emulators, and gives each one to an AI agent with a persona. Each agent uses the app like a new user and writes a bug report with numbered screenshots.
It's open source (MIT): github.com/pierridotite/mobile-agent-testers
A real bug, found twice
To check it on an app I didn't write, I pointed two agents at the example app of Flutter's official animations package. One ran on a Pixel 7 with Android 14, the other on an iPhone SE in French with large text. Neither knew about the other.
Both reported the same typo: the screen title says "Streamling your courses" instead of "Streamlining". It's still there on line 110.
A typo is small, but nobody writes a test for one. The same run also noted things a scripted test would not look for: the app ignores the system dark mode (the Android phone was in dark mode), and some light grey text is hard to read on a small screen.
It also got one wrong: the iPhone agent flagged a narrow settings sheet as cut off when it was not. That is why every issue points to numbered screenshots, so you can check it in a few seconds.
The run took 19 minutes, builds included, and $0.47 of model usage for both agents.
How it works
- Personas set the device. Each persona is a few lines of JSON: phone model, OS version, font size, dark mode, language, location, and who the user is ("retired teacher, 68, low patience with jargon"). The repo ships four.
- Each agent has three tools: see the screen (a screenshot plus the accessibility tree with tap coordinates), gestures, and typing. Agents read pixels and the accessibility tree, so any UI toolkit works.
- The build is automatic. Flutter, React Native, Expo, native Kotlin and Swift, Kotlin Multiplatform, Capacitor and .NET MAUI are detected. For anything else, give your own build command or a prebuilt app.
- Results are encrypted. Logs of a public repo are public, so reports, screenshots and optional screen recordings are encrypted with your key, then decrypted on your machine.
- Analytics are blocked inside the phones, so test sessions never show up in your real stats.
What it costs
GitHub Actions runners are free for public repos: about 20 jobs at once, 5 of them on macOS for the iPhones. The only real cost is the model. A Claude subscription works if you run batches of 6 to 10 agents, and an API key works too.
What it is not
It does not replace Maestro, Detox or Appium. Those replay the steps you wrote, the same way every time, which is what you want for regressions. These agents explore, so two runs never find exactly the same things. I use scripted flows for regressions and the agents before a release.
Try it
gh repo create my-testers --public --template pierridotite/mobile-agent-testers --clone cd my-testers && ./setup.sh # key pair, secrets, Claude token gh workflow run agent-testers.yml # demo app: 2 Android + 2 iOS testers, about 20 minutes Then point it at your own app with -f source_repo=you/app -f app_id=com.you.app. Or install the Claude Code plugin and let Claude set it up:
/plugin marketplace add pierridotite/mobile-agent-testers /plugin install mobile-agent-testers@mobile-agent-testers If you run it on your app, I'd like to know what it finds, good or bad. Issues and feedback on the persona format are welcome.





