App Walk: open phone agents run exploratory QA on indie Android apps

This space is crowded, and the exact pitch is already taken. Google ships a Gemini-powered app testing agent inside Firebase App Distribution. It takes an uploaded APK plus natural-language goals and reports results without an SDK, and many indie Android developers already use Firebase. On top of that, TestPilot, Stora, Autosana and FlyTrap all sell upload-a-build agent exploration that reports crashes with replays. Open-source tools like FinalRun and DroidRun let a technical solo developer do the same with their own API keys at the cost of model tokens. Paying for wrapped open agents alone is not a defensible gap. A newcomer's only realistic opening is narrow. One option is to be much cheaper and simpler than the YC-backed tools for very small studios: flat per-run pricing, no CI setup, and a clear video and step log for each goal. Another is to go deeper on the parts Google and the general tools treat lightly. That means reliably detecting dead ends and UX traps, not just crashes; handling real-world blockers like OTP sign-ups, paywalls and test purchases through Play Billing; and turning findings into ready-to-file reproduction steps. A third is to target niches where generic agents struggle, such as Unity and other game engines, Flutter canvas rendering, or apps in languages other than English. Even with one of those angles, Google could add features or drop prices, and indie developers are notoriously unwilling to pay. That makes this a hard business unless distribution comes cheaply, for example through an existing indie developer community or a marketplace listing.

• Firebase App Testing agent (Google, Gemini in Firebase)

  This is the core App Walk workflow, built into a platform indie Android developers already use. You upload a signed APK or AAB to App Distribution and write test goals in natural language. Gemini then works through the app, taps through each step and reports results. No SDK is needed. I could not confirm from the results whether it is in preview or what it costs.

• Firebase Test Lab Robo test

  Google's older crawler. It explores an APK on its own without scripts, accepts test login details, and reports crashes. Indie developers already get a cheap or free baseline from it.

• TestPilot (Minitap)

  You drop in an APK or IPA, AI agents explore the app, and bugs are logged with stack traces, screenshots and session replay. This is almost the same pitch as App Walk.

• Stora

  Agents install the build on a fresh simulator and rank what they find: crashes as P0, ANRs as P1, visual regressions as P2. Critical failures can automatically block a release.

• Autosana

  A YC-backed startup doing mobile QA with natural-language tests. It accepts Android APK and iOS simulator builds and plugs into CI/CD. It markets crash testing with AI agents directly.

• FlyTrap

  A smaller exploratory testing tool for mobile apps. It scrolls and taps like a human would and sends bugs to Jira, Linear and GitHub.

• GPT Driver (MobileBoost)

  An AI-native platform for mobile test automation with natural-language tests that cope with UI changes and pop-ups. It sells to larger app teams such as Noom.

• Revyl

  An AI testing platform for mobile and web whose agent reacts to the UI as it runs, pitched as a replacement for Appium-style scripts.

• FinalRun Agent (open source)

  An open-source CLI. You write tests as plain-English YAML, and a vision model drives an emulator using your own Google, OpenAI or Anthropic API keys. It is a free do-it-yourself version of the idea.

• DroidRun (open source)

  An open-source framework for agents that control Android devices over adb using the accessibility tree and vision. It is positioned for automation and testing in CI. This is one of the open phone agents App Walk would wrap.

• Academic tools (CrashScope, DroidAgent, LLMDroid, LELANTE)

  CrashScope already offered upload-an-APK crash discovery on an emulator. Newer LLM agents like DroidAgent and LLMDroid set their own goals and push coverage up. LLMDroid reports running for roughly $0.18 to $4.77 an hour, so the underlying method is well understood and cheap.

Checked 2026-10-25