20 Claude Code testers run mobile QA on GitHub runners
An open-source project puts a Claude Code agent on each of 20 simulators and emulators to hunt for broken flows, layout and translation bugs without anyone writing a test script.
What happened
On 2026-11-16 a developer publishing as @pierridotite created mobile-agent-testers. Six days later, on 2026-11-22, the repository had 329 stars and 31 forks. It is a public, MIT-licensed Python project built around AI mobile app testing agents: 20 Claude Code testers on 20 different phones, in parallel, on GitHub's free runners. The phones are iOS Simulators and Android emulators rather than physical hardware, and there are no test scripts to write.
That is the difference from Maestro, Appium or Detox, which run flows a person wrote. Here each agent plays a persona and looks for what nobody thought to test.
How a run works
Each tester is a Claude Code agent with its own simulator or emulator on GitHub Actions. A persona fixes the device and the user: an iPhone SE in Swedish at the largest text size, a Pixel 7 at 200% text, an iPad in German, an Arabic right-to-left phone. The agent gets three tools, seeing the screen as a screenshot with tap coordinates, gestures, and typing, plus the accessibility tree. On every screen it checks for controls with no name and touch targets under 48 dp on Android or 44 pt on iOS, then confirms each hit on the screenshot.
Runs are configured through workflow inputs: app repo and branch, app folder and bundle id, which personas to use, a brief, an issue or set of steps to reproduce, the model, and whether to record video. Flutter, React Native and Expo, native Kotlin and Java, Swift and Objective-C, Kotlin Multiplatform, Capacitor and Ionic, and .NET MAUI in beta are detected automatically. Anything else can point at its own build command and output path. Reports, screenshots and recordings are encrypted with your key before leaving the runner, and analytics hosts are blocked inside the phones so test runs do not touch real stats.
The README lists six uses: explore as a new user, push every agent through one feature from your own brief, reproduce a bug report on each phone, collect an improvement report, compare two branches, or switch to a cheaper model. Claude Haiku is supported, and so is any Anthropic-compatible API, with OpenRouter, DeepSeek, Kimi, GLM, MiniMax and Qwen named.
What the sample runs found
The project's README on GitHub reports three runs on open-source apps. LocalSend, a Flutter app, across 20 phones (10 Android, 10 iOS, 14 languages): 46 distinct bugs, including a blocker where an iPhone SE in Swedish at the largest text size could not send a text message at all. The run took 69 minutes and $4.62 of model usage. The 103 reports collapsed to 46 bugs, 29 of which a single agent found, and 34 had nothing to do with text size or accessibility.
AnkiDroid on 10 Android phones produced 43 distinct issues, among them answer buttons cut in half at 200% font and "Get started" left in English in three languages, in under 10 minutes and $4.69. The bundled demo app ships with 11 planted bugs; four testers found 9 of them, plus a button hidden under the Android nav bar and data lost after a relaunch.
Cost, limits and open questions
GitHub's runners are free for public repos, macOS included, so the bill is the model. Documented usage was $0.15 to $0.55 per phone, about $5 for 20 phones, or it draws on a Claude subscription instead. Roughly 20 jobs run at once, 5 of them on macOS. On plans with a five-hour usage limit, 20 agents can use it up in about 15 minutes, so the README recommends batches of 6 to 10. Private repos spend Actions minutes, and macOS minutes count ten times.
Two runs will not produce the same list, because each agent explores rather than replays. The README keeps scripted tools for regressions and records a known failure: 3 of LocalSend's 20 agents flagged the Simulator's English keyboard as part of the app. Every finding arrives with numbered screenshots so it can be checked before filing.
What this suggests
For a small team without a mobile QA process, this likely lowers the price of a first pass over a release. A few dollars and an hour buys coverage across languages, text sizes and tablets that would otherwise mean borrowing devices and people. It looks less like a replacement for a regression suite than a pre-release sweep, which is roughly how the README frames it.
What it means for building apps with AI
The pattern here, an agent given a device, a persona and a goal instead of a script, is the same shift showing up in coding tools, and it depends on the agent reading screens rather than following instructions. The caveat is maturity: this is a six-day-old repository with 329 stars whose evidence comes from the author's own runs.
If you build with Draftship, the output is a hosted web app rather than an iOS or Android build, so this workflow does not apply to it. Publishing is a separate step in the editor, covered in Draft and publish.
Primary sources
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.