pierridotite/mobile-agent-testers
Developer ToolsRun parallel Claude Code QA agents on iOS simulators and Android emulators to explore mobile apps and file bug reports.
This repo is an agent skill pack · 2 skillsSkill list and install →What it does
mobile-agent-testers turns mobile QA into a parallel exploration job. Each tester is a Claude Code agent with its own iOS Simulator or Android emulator on GitHub Actions, playing a persona such as an iPhone SE user with Swedish UI and the largest text size, a Pixel 7 at 200% font, or an iPad in German. The agent uses the app like a new user, then writes a bug report with numbered screenshots.
Besides free exploration, a brief can drive agents through one feature end to end, reproduce a bug report on every device, or produce an improvement report. A compare mode runs the same personas against two builds for a pre-release diff. The supported frameworks include Flutter, React Native/Expo, native Kotlin/Java, Swift/Objective-C, Kotlin Multiplatform, Capacitor/Ionic and .NET MAUI (beta); because agents read screenshots and the accessibility tree, other UI toolkits can be wired through custom build inputs.
Quick start
You need the gh CLI and a Claude subscription or API key; a Mac is not required. Create a repository from the template, run the setup script to generate the key pair and store the secrets, then trigger the demo workflow:
gh repo create my-testers --public --template pierridotite/mobile-agent-testers --clone
cd my-testers && ./setup.sh # Windows: .\setup.ps1
gh workflow run agent-testers.yml # 2 Android + 2 iOS testers, ~20 min
bash scripts/collect.sh <run id> # decrypts reports into runs/<id>/SUMMARY.md
To test your own app, dispatch the workflow with the source repository, app id and the persona ids from personas.json:
gh workflow run agent-testers.yml -f source_repo=you/app -f app_id=com.you.app \
-f android_personas=anna,malik -f ios_personas=clara,jonas
The remaining inputs cover the app directory, a brief (default exploration, improvement, or your own feature script), a bug report to reproduce, the model and provider, a forced framework, custom build commands or prebuilt app paths, blocked analytics hosts, and per-device video recording. Results come back decrypted only on your machine.
Practical limits
Runs are not deterministic: two runs take different paths, so expect the same kinds of bugs rather than the same list. Scripted suites such as Maestro or Appium remain the tool for regressions. Each agent explores for roughly 1 to 18 minutes, which covers main paths, not every corner of a large app, and screen misreads happen — LocalSend's run had 3 of 20 agents mistake the simulator keyboard for app UI — so check the screenshots before filing.
Free GitHub runners allow about 20 concurrent jobs for public repos but only 5 macOS jobs, and iOS work consumes macOS minutes. Plans with a 5-hour usage limit can be exhausted by 20 agents in around 15 minutes, so batches of 6–10 are advised. Hosted runners are intended for the repository's own project; closed-source apps should run the workflow in a private repo or on your own Mac. Results, screenshots and videos are encrypted with your key because public-repo logs are public, and analytics hosts are blocked inside the phones; still use test accounts. Model usage was cited at $0.15–$0.55 per agent in the published runs. The project is MIT licensed.
HotRepo's take
The repository showed up on HotRepo's new-project list on 2026-11-22, gaining 36 stars in the recorded one-hour window to reach 140 stars and 12 forks. That single snapshot supports only the fact of an early burst on release day; no daily all-language rank was recorded, so there is no sustained trend to read from the data yet. The technical substance that stands out is the distributed-session design: a workflow builds the app once, then fans out one emulator plus one agent per persona, with encrypted reports grouped back on the user's machine. That is more than a single agent attached to one simulator, as in agent-device, and it targets a real gap left by scripted UI testing. The documented caveats — nondeterminism, short exploration windows and occasional screenshot misreads — are the right lens for judging whether it fits a release checklist rather than replacing a regression suite.
Written by AI from the project's public information and HotRepo's records, updated . The project's own docs are the source of truth.
You might also like
- callstack/agent-deviceA CLI, MCP server, and typed Node.js API that let AI coding agents inspect, control, and verify apps on mobile, TV, web, and desktop devices.
- jo-duchan/tapflowSelf-hosted relay and macOS agent that stream iOS simulators and Android emulators to a browser for team QA.
- mobile-next/mobile-mcpAn MCP server that lets AI agents inspect and control iOS and Android apps on simulators, emulators and real devices.
- hridaya423/phonedriverMCP server and CLI that let AI agents read and control Android phones and iPhones with a confirmation step before payments.
- erkamyaman/unslopovichAn agent skill that reviews mobile and web UIs for AI-generated layout, UX and accessibility problems, then proposes and applies fixes.