Mobile PR Reviewer

Part of Mobile DevTools — open-source tools for mobile engineering teams.

Claude reviews mobile PRs by booting a cloud simulator, tapping through the app, and posting a link to the recording.

That's a real PR review. Claude read the diff, started a Pixel 7 simulator on Revyl, navigated to the changed screen, hit a bug, and posted a comment with the recording link. ~90 seconds. No test code, no YAML, no maintenance.

Two recordings from the bundled Bug Bazaar sample app — the same flow, different platforms in parallel:

• Android — session 76a6daae · Pixel 7 · 1:36
• iOS — session 2960f466 · iPhone 17 Pro Max · 1:34

A full sample PR comment is in examples/sample-pr-comment.md.

Setup

1. Fork this repo. The bundled sample-app/ is Bug Bazaar — it builds out of the box and has a real bug to demo against.

2. Add 5 secrets in Settings → Secrets and variables → Actions:

   Secret	Where to get it
   ANTHROPIC_API_KEY	console.anthropic.com
   REVYL_API_KEY	app.revyl.ai → Settings → API Keys
   REVYL_APP_ID_ANDROID	revyl app create --name "myapp-android" --platform android --json
   REVYL_APP_ID_IOS	revyl app create --name "myapp-ios" --platform ios --json
   EXPO_TOKEN	Only if your build uses Expo (the sample app does)

3. Open a PR. The bot reviews on PR open and on /review comments.

If you only support one platform, set just that platform's REVYL_APP_ID_* and drop the other from strategy.matrix.platform in .github/workflows/review.yml.

How it works

For each PR, two parallel jobs (Android + iOS) start a Revyl cloud simulator, hand control to Claude, and let it loop:

1. revyl device screenshot     → Read the PNG
2. Decide the next action from what's on screen
3. revyl device tap --target "..."   (or swipe / type)
4. Repeat

When Claude is done — or sees a bug — it posts a PR comment with the recording link. The recording has the video and every action Claude took. Claude is the planner; there's no test plan written ahead of time, no DSL, no YAML.

Three files the bot reads: CLAUDE.md (the reactive loop), CLAUDE-test.md (optional /test mode), .github/workflows/review.yml (the trigger). That's the whole product.

Optional: trigger an existing Revyl test

If your team already has Revyl tests or workflows, the bot can trigger one against the PR build instead of doing a free-form review. Add one of these as a repo secret:

• REVYL_TEST_NAME — name of an existing Revyl test
• REVYL_WORKFLOW_NAME — name of an existing Revyl workflow

Then comment /test on a PR. Claude runs revyl test run "$REVYL_TEST_NAME" --json --verbose, parses report_link, posts it. No test design, no YAML — trigger and report. Leave both secrets unset to disable. See CLAUDE-test.md.

Customize for your app

Replace sample-app/ with your source, or point the Build app step in .github/workflows/review.yml at your existing CI artifact:

- name: Build app
  run: |
    if [ "${{ matrix.platform }}" = "android" ]; then
      ./gradlew assembleDebug
    else
      xcodebuild -scheme MyApp -sdk iphonesimulator -configuration Debug
    fi

Then edit the bottom of CLAUDE.md (the "Demo hint" section) to describe your screens — bottom tabs, top-level routes, login flow, anything Claude needs to know to navigate confidently.

To swap models, edit claude_args: --model claude-sonnet-4-5 in the workflow. claude-opus-4-6 for harder reviews, claude-haiku-4-5 for cheaper smoke tests.

License

MIT