Skip to main content
Mobile Developmentbitwarden

interacting-with-android-device

Instructions for capturing UI state, comparing with mocks, and interacting with an Android device using MCP tools backed by ADB.

Stars
8,903
Source
bitwarden/android
Updated
2026-05-29
Slug
bitwarden--android--interacting-with-android-device
View on GitHubRaw SKILL.md

// install — copy + paste into any project

mkdir -p .claude/skills && curl -fsSL https://raw.githubusercontent.com/bitwarden/android/HEAD/.claude/skills/interacting-with-android-device/SKILL.md -o .claude/skills/interacting-with-android-device.md

Drops the SKILL.md into .claude/skills/interacting-with-android-device.md. Works with Claude Code, Cursor, and any agent that loads SKILL.md files from .claude/skills/.

Interacting with Android Device

Quick Start: MCP Tools

The android-device MCP server provides 6 tools for device interaction. These replace the previous shell scripts with proper XML parsing, structured dumpsys parsing, and native obstruction detection.

Available tools:

  • capture — Capture UI hierarchy XML and/or screenshot. Params: { xml?: boolean, screenshot?: boolean }. Default: both.
  • find_element — Find element by text or content-desc, return coordinates with obstruction detection. Params: { text: string }. Returns JSON with coordinates, bounds, and obstruction status.
  • tap_at — Tap at specific coordinates, wait, capture screenshot. Params: { x, y, waitSeconds? }.
  • tap_element — Find, tap, and capture in one call (recommended). Params: { text, waitSeconds? }. Auto-adjusts coordinates when obstructed.
  • navigate — Navigation actions: home, back, app-drawer. Params: { action, waitSeconds? }. Captures screenshot after action.
  • input_text — Type text into the focused field. Params: { text, clear? }. Set clear: true to erase existing content first.

Use these MCP tools instead of raw ADB commands to save tokens, get structured results, and benefit from automatic obstruction detection.

1. Capturing Current State

To understand what is currently on the device, use the capture tool:

  • It saves view.xml (UI hierarchy) and screen.png (screenshot) to the working directory
  • Read view.xml to find coordinates (bounds) and properties (like text or resource-id) of UI elements
  • Use screen.png for visual verification against design mocks

2. Interacting with the Device

Using MCP Tools (Recommended)

  • Find and tap an element by text — use tap_element: This finds the element, detects obstructions, taps (with adjusted coordinates if needed), and captures a screenshot — all in one call.

  • Tap at specific coordinates — use tap_at: When you already have coordinates from find_element or manual inspection.

  • Navigate (home, back, app-drawer) — use navigate: Performs the action and captures a screenshot.

  • Find element without tapping — use find_element: Returns coordinates and full element info. Useful when you need to inspect before acting.

  • Type text into a field — use input_text: Types text into the currently focused field. Set clear: true to erase existing content first. Tap the field before calling this if it isn't already focused.

Raw ADB Commands (When MCP Tools Aren't Sufficient)

  • Key Events:
    • Back: adb shell input keyevent 4
    • Home: adb shell input keyevent 3
    • Enter: adb shell input keyevent 66
  • Scrolling/Swiping: Use adb shell input swipe <x1> <y1> <x2> <y2> <duration_ms> where:
    • (x1, y1) = starting point
    • (x2, y2) = ending point
    • duration_ms = duration in milliseconds (1000ms is typical; adjust for speed/distance)
    • Note: For expanding containers/drawers, use large distances (e.g., 2400->300 for a 2992px tall screen)

3. Obstruction Detection

The find_element and tap_element tools automatically detect when another element would intercept the tap. This catches:

  • System overlays (Layer 1): TalkBack floating menu, PiP windows, accessibility services — detected via dumpsys window windows touchable regions
  • In-app elements (Layer 2): FABs, dialogs, bottom sheets, snackbars — detected by finding the topmost clickable element at the tap point in the UI hierarchy

When obstruction is detected:

  • Coordinates are auto-adjusted to the center of the largest unobstructed strip (top/bottom/left/right of the obstructor)
  • The response includes the obstructor identity, bounds, and visible region info
  • If fully obscured (no visible region), the original center is returned as best-effort
  • Compose parent wrapper pattern (identical bounds) is recognized as non-obstruction

4. Verification Workflow

Follow these steps for a complete UI test:

  1. Build and Install: Ensure the latest version of the app is running: ./gradlew installDebug.
  2. Inspect: Use capture to dump the UI hierarchy and take a screenshot.
  3. Compare: Check the current UI against any mock image files in the project.
  4. Interact: Use tap_element to tap a UI element by text. The tool handles coordinate calculation and obstruction detection automatically.
  5. Verify: Use capture again to confirm the UI has updated as expected (e.g., a new screen is shown, or a success message appeared).

5. Examples

Example: Navigate to Settings and Check for Updates

# Go to home screen
navigate({ action: "home" })

# Open app drawer
navigate({ action: "app-drawer" })

# Find and tap through settings
tap_element({ text: "Settings", waitSeconds: 2 })
tap_element({ text: "System", waitSeconds: 2 })
tap_element({ text: "Software updates", waitSeconds: 2 })
tap_element({ text: "Check for update", waitSeconds: 5 })

Example: Swiping

For swipe gestures not covered by the navigate tool, use raw ADB:

adb shell input swipe 672 2800 672 500 1000 && sleep 1 && adb shell screencap -p /sdcard/screen.png && adb pull /sdcard/screen.png .

6. Best Practices

Coordinate Calculation

  • Prefer find_element or tap_element over manual coordinate calculation — they handle bounds parsing, center computation, and obstruction detection automatically
  • When multiple instances of an element exist (e.g., in prediction row and full list), check the find_element response to verify you're targeting the correct one

Navigation and State Evaluation

  • Verify after each interaction: Don't assume an action succeeded — use capture after interactions to confirm the UI changed as expected
  • Check both visual and structural state: Use screenshot for visual verification, XML dump for structural confirmation (element presence, text content, state changes)
  • Identify navigation failures early: If a tap opened the wrong screen, use navigate({ action: "back" }) to recover immediately

Interaction Patterns

  • Scrolling before interaction: When looking for an element, check if it's visible on screen first. If not, scroll using swipe gestures to reveal it
  • Use consistent scroll direction: For vertical scrolling in lists/settings, use downward swipes (higher Y -> lower Y) to scroll down
  • Handle app crashes gracefully: Don't retry the same action — use back button and try an alternative approach
  • Check Accessibility: Use the content-desc and text properties in the UI hierarchy to ensure the UI is accessible for screen readers

7. Troubleshooting

Device Not Connected

If tools report ADB errors:

  • Check USB connection or emulator status
  • Enable USB debugging on the device (Settings > Developer Options > USB Debugging)
  • Accept the RSA key prompt on the device if asked
  • Restart the device or disconnect/reconnect the USB cable
  • Run adb devices to verify the device is visible

MCP Server Not Available

If tools are not listed in /mcp:

  • Ensure Node.js 18+ is installed
  • The server auto-builds on first use via .mcp.json at the project root
  • Check .claude/mcp/android-device-server/ exists with package.json
  • Try manual build: cd .claude/mcp/android-device-server && npm install && npm run build