Using Claude to Automate Mobile App QA with Capacitor WebViews
A developer documented how they taught Claude to perform automated quality assurance for a mobile app built with Capacitor. The app uses React wrapped in native shells (WebView on Android, WKWebView on iOS) with server-driven UI architecture, allowing one codebase to run on web, iOS, and Android platforms.
Testing Challenge and Solution
Capacitor apps exist in a testing gap: Playwright can't access the native shell, while XCTest and Espresso can't interact with HTML inside WebViews. The developer created a Python script that uses Claude to drive both mobile platforms, take screenshots, analyze them for issues, and file bug reports automatically.
Android Implementation Details
Android setup took 90 minutes. Key steps:
- Connectivity fix:
adb reverse tcp:3000 tcp:3000andadb reverse tcp:8080 tcp:8080(needs re-running after emulator restart) - WebView DevTools access: Find socket with
adb shell "cat /proc/net/unix" | grep webview_devtools_remote - Forward to local port:
adb forward tcp:9223 localabstract:$WV_SOCKET - Full Chrome DevTools Protocol access via
curl http://localhost:9223/json
The script sweeps all 25 app screens in about 90 seconds using CDP for navigation and authentication (injecting JWT into localStorage) and adb shell screencap for screenshots.
Analysis and Bug Reporting
Screenshots are analyzed for visual issues: broken layouts, error messages, missing images, blank screens, and status bar overlap. When issues are found, the system:
- Authenticates as zabriskie_bot
- Uploads screenshots to S3
- Files bug reports to the production forum with format:
[Android QA] Shows Hub: RSVP button overlaps venue text
The system knows expected states: "Forbidden" responses for non-members on crew pages aren't bugs, empty avatar circles aren't bugs, and "Preview" text in profile settings is a known cosmetic issue.
iOS Implementation
iOS setup took over six hours, highlighting differences in mobile automation tooling. The article notes this contrast but provides fewer specific technical details about the iOS implementation compared to Android.
Deployment
The entire QA system runs as a scheduled task every morning at 8:47 AM.
📖 Read the full source: HN AI Agents
👀 See Also
Deterministic Compiler Architecture for Multi-Step LLM Workflows Shows Strong Benchmark Results
A deterministic compilation architecture for structured LLM workflows uses typed node registries, parameter contracts, and static validation to compile workflow graphs ahead of time. Benchmarks show it outperforms GPT-4.1 and Claude Sonnet 4.6 across workflow depths from 3-12+ nodes.
codebase-md: Tool auto-generates CLAUDE.md with git hook maintenance
codebase-md v0.1.0 scans projects to generate CLAUDE.md files with architecture detection, dependency health checks, and git insights. It includes git hooks to keep documentation fresh and supports other AI coding tools with additional config files.
Lisp Development with AI Agents: High Costs and Technical Challenges
A DevOps engineer found AI agents struggle with Lisp development, costing $10-$20 in minutes for subpar code, while Python and Go work efficiently. He created tmux-repl-mcp to improve REPL interaction but still faced high token costs and tooling issues.
Codev: AI agent workflow for 106 PRs in 14 days
Codev is an open-source system that coordinates multiple AI agents through a strict Spec→Plan→Implement→Review→PR workflow, catching 20 bugs before shipping and producing code rated 1.2 points better on a 10-point scale.