Open Source · MITAndroid + BrowserSelf-Hostable

Computer-Use Testing for Android Apps and Browser Extensions

Test Android apps and browser extensions with a computer-use agent — it drives the real UI like a user and checks the result. Open source and MIT licensed — no vendor lock-in, no agent infra in the middle, runs entirely on your own CI.

View on GitHub Read the Blog

pip install agentprobeagentprobe run --target android|browser --case <path>

How it works

Two backends. One CLI.

agentprobe ships an Android backend driven by adb and a browser backend driven by Chrome CDP. Both feed screenshots to a vision model and translate its decisions into real UI actions.

Android backend

adb screencap → base64 PNG
       ↓
  vision model (Azure / OpenAI / Gemini / xAI)
       ↓
  adb shell input (tap / type / swipe / key)
       ↓
  ui_dump assertions + NL verification

Browser backend

Chrome CDP → scrot screenshot
       ↓
  vision model
       ↓
  xdotool click / type
       ↓
  CDP assertion + NL verification

Features

Everything you need to run real-device UI tests in CI.

Drives the real UI

No mocks, no test doubles. Interacts like a human user via adb input or xdotool. Catches regressions test doubles miss.

BYOK, no vendor lock-in

Supports Azure, OpenAI, Gemini, xAI. MIT licensed, self-hosted on your CI, no call home.

GIF artifacts + CI reports

Every run records a screen GIF via ffmpeg and emits JUnit XML. Pass the GIF URL in the PR comment.

CI Usage

GitHub Actions integration

Android tests run inside android-emulator-runner (API 28). Browser tests run in a containerized Chrome instance. Both targets emit JUnit XML and a screen GIF.

.github/workflows/android-cua.yml

# .github/workflows/android-cua.yml
- uses: reactivecircus/android-emulator-runner@v2
  with:
    api-level: 28
    script: agentprobe run --target android --case cases/onboarding.yaml

# For browser extensions:
- run: |
    agentprobe run --target browser \
      --extension ./my-extension.crx \
      --case cases/sidepanel.yaml

Test case schema

Minimal YAML — no code required

Test cases are plain YAML files. Describe the goal and success criteria in natural language; the vision model handles the rest.

cases/onboarding.yaml

name: onboarding-flow
instruction: "Open the app and complete onboarding until you see the dashboard"
successCriteria: "Dashboard screen is visible"
failureCriteria: "App crashes or shows error dialog"
maxSteps: 30
verification:
  prompt: "Is the dashboard visible and functional?"

instruction

Natural language goal for the agent

successCriteria

NL condition checked after completion

failureCriteria

Abort condition detected mid-run

maxSteps

Hard cap on agent action steps

verification.prompt

Final NL assertion sent to vision model

Open source

MIT licensed. No agent infra in the middle.

agentprobe is MIT licensed and entirely self-hostable. It runs on your own GitHub Actions runner — there is no Agent Labs infrastructure between your test runner and your app. Your screenshots, your API keys, and your source code stay on your machines.

View on GitHub

pip install agentprobe