ADB Claw — Android Device Control

Your eyes, hands, and ears on Android. See what's on screen, tap any element, scroll through content, open deep links, wait for UI changes, capture system audio, manage apps, and more — all through a single CLI with structured JSON output.

Why ADB Claw

Superpowers — What You Can't Get Elsewhere:

• - Live stream intelligence — monitor connects to Android's accessibility framework, reading all UI text in real-time — even during video playback and live streams where uiautomator dump hangs. Chat messages, captions, dynamic overlays — data no other tool exposes to agents.
• System audio capture — audio capture records device audio via REMOTE_SUBMIX (Android 11+); streams WAV to stdout for piping to ASR tools. Combined with monitor, you get full sensory coverage: visual text + audio.

Core Strengths:

• - Observe → Act → Verify loop — observe returns screenshot + indexed UI tree in one call; use element indices to target precisely across any screen size
• Deep links bypass CJK limits — adb input text can't type Chinese/Japanese/Korean; adb-claw open 'app://search?keyword=中文' can
• Wait, don't poll — wait --text "Done" blocks until the UI element appears, replacing fragile sleep/observe loops
• Smart scroll — auto-calculates swipe coordinates from screen size; supports direction, page count, and scrolling within specific elements
• App Profiles — pre-built knowledge (deep links, layouts, known issues) for popular apps like Douyin; load once, skip trial-and-error
• Full app lifecycle — install, launch, stop, uninstall, clear data — no raw adb needed
• Agent-optimized JSON — every command returns {ok, command, data, error, duration_ms} with actionable suggestion on errors
• Minimal device footprint — nearly all operations are pure ADB commands; only monitor and audio capture push temporary ~7KB helpers that auto-exit

Actively Evolving — new capabilities ship regularly. Each release expands what you can perceive and control on Android devices.

Getting Started

Claude Code

Install the plugin, then just talk to Claude — no slash commands needed:

CODEBLOCK0

The plugin auto-downloads the adb-claw binary on first session. Make sure adb is installed and a device is connected via USB with debugging enabled.

Then simply ask Claude to interact with your Android device:

CODEBLOCK1

Claude reads the Triggers list below and automatically activates this skill when your message matches — no explicit invocation required.

OpenClaw

Install from ClawHub:

CODEBLOCK2

Same natural-language triggers apply. Ask your agent to control an Android device and it will invoke adb-claw commands.

Triggers

These patterns tell the agent when to activate this skill:

• - User asks to control, interact with, or automate an Android device
• User asks to test a mobile app or UI on Android
• User mentions tapping, swiping, scrolling, screenshots, or app management on Android
• User wants to open a URL, deep link, or specific app screen on a connected device
• User wants to wait for UI elements to appear/disappear on Android
• User wants to manage screen state (on/off/unlock/rotation) on Android
• User wants to push/pull files to/from an Android device
• User wants to monitor live stream chat or read UI text during video playback on Android
• User wants to capture or record audio from an Android device
• User wants to grab shopping cart products from a Douyin live stream
• User wants to run shell commands on an Android device

Binary

The adb-claw binary is located at ${CLAUDE_PLUGIN_ROOT}/bin/adb-claw.

The binary is installed automatically via the SessionStart hook. If adb-claw is not available, inform the user that the plugin needs to be reinstalled — do not attempt to download or install it yourself.

Setup

Requires two binaries:

1. 1. adb-claw — the control CLI
2. adb — Android Debug Bridge (from Android SDK Platform-Tools)

Install adb-claw

Installed automatically by the plugin. For manual installation, see GitHub Releases.

Install adb

CODEBLOCK3

Connect device

The Android device must have USB debugging enabled and be connected via USB. This is the most common blocker — most users haven't turned it on. When a user first asks to control their phone, always check connection first (adb-claw doctor) and if it fails, walk them through the setup steps below before attempting any other commands.

How to enable USB debugging (guide the user through this)

1. 1. Open Settings on the Android phone
2. Go to About phone (some phones: Settings → My device)
3. Tap Build number (or MIUI version) 7 times — a toast confirming Developer Mode is enabled will appear
4. Go back to Settings → Additional settings → Developer options (path varies by brand):

- Xiaomi/Redmi: Settings → Additional settings → Developer options - Samsung: Settings → Developer options - Pixel/Stock: Settings → System → Developer options - OPPO/Vivo: Settings → System management → Developer options

1. 5. Enable USB debugging toggle
2. Connect phone to computer via USB cable
3. A dialog "Allow USB debugging?" will appear on the phone — tap Allow (check "Always allow from this computer" for convenience)

Verify connection

CODEBLOCK4

If doctor reports no device, ask the user to:

• - Check the USB cable (some cables are charge-only, no data)
• Try a different USB port
• Re-authorize USB debugging on the phone (revoke and re-allow)
• On some phones, change USB mode from "Charging" to "File Transfer" in the notification shade

Quick Start

The core loop is observe → decide → act → observe:

CODEBLOCK5

For CJK apps, use deep links to bypass text input limits:

CODEBLOCK6

App Profiles

App Profiles are pre-built knowledge bases for specific apps — deep links, UI layouts, device-specific behavior, and known issues. They dramatically reduce the trial-and-error needed to automate an app.

Available Profiles: skills/apps/ directory

App	File	Key Content
Douyin (抖音)	INLINECODE19	Search/user/live deep links, feed/search/profile layouts, Phone vs Pad differences, live stream chat monitoring
Meituan (美团)

meituan.md | Search/waimai deep links, homepage/menu/search layouts, WebView workarounds, popup chain handling |

Usage:

1. 1. adb-claw app current → get foreground app package name
2. Check skills/apps/ for a matching Profile
3. Has Profile → use deep links and known layouts (fast path)
4. No Profile → observe + explore (slow path)
5. Check device form factor: adb-claw device info → short edge < 1200px = Phone, >= 1200px = Pad

Profiles are plain Markdown files. New app support = drop a .md file into skills/apps/.

Global Flags

Flag	Short	Description	Default
INLINECODE27	INLINECODE28	Target device serial (when multiple devices connected)	auto-detect
INLINECODE29

-o | Output format: json, text, quiet | json | | --timeout | | Command timeout in milliseconds | 30000 | | --verbose | | Enable debug output to stderr | false |

Commands

observe — Screenshot + UI Tree (Primary Command)

Captures screenshot and UI element tree in one call. Always use this before and after actions.

CODEBLOCK7

Returns: base64 PNG screenshot, indexed UI elements with text/id/bounds/center coordinates.

screenshot — Capture Screen

CODEBLOCK8

tap — Tap UI Element

Tap by element index (preferred), resource ID, text, or coordinates:

CODEBLOCK9

Always prefer --index over coordinates. Index values come from observe output.

long-press — Long Press

CODEBLOCK10

swipe — Swipe Gesture

CODEBLOCK11

type — Input Text (ASCII only)

CODEBLOCK12

Important: Only ASCII text is supported. For CJK/emoji input, use open with deep links (e.g., adb-claw open 'myapp://search?keyword=中文').

key — Press System Key

CODEBLOCK13

clear-field — Clear Input Field

Clear text in the currently focused input field. Optionally tap an element first to focus it.

CODEBLOCK14

Uses Ctrl+A+DEL on SDK 31+, falls back to repeated DEL on older devices.

open — Open URI (Deep Link)

Open any URI using Android's ACTION_VIEW intent. The key to CJK text input — pass Chinese/Japanese/Korean text as URL parameters in deep links.

CODEBLOCK15

scroll — Smart Scroll

Scroll the screen or a specific scrollable element. Auto-calculates swipe coordinates from screen size — no manual coordinate math needed.

CODEBLOCK16

Always prefer scroll over manual swipe for page navigation.

wait — Wait for UI Condition

Wait for a UI element or activity to appear or disappear. Replaces fragile sleep+observe polling loops with a single blocking call.

CODEBLOCK17

Default timeout: 10s. Default poll interval: 800ms.

screen — Screen Management

CODEBLOCK18

app — App Management

CODEBLOCK19

monitor — Continuous UI Text Monitoring

Monitor UI text by connecting directly to the Android accessibility framework. Unlike ui tree which uses uiautomator dump, this command skips video surface nodes and works reliably during live streams and video playback.

CODEBLOCK20

Bounded mode (default): runs for --duration ms, returns all unique text entries in a JSON envelope.

Streaming mode (--stream): outputs each new text as a JSON line in real time.

Default duration: 10s. Default poll interval: 2s.

Note: This command pushes a small (~7KB) DEX helper to the device on first use. The helper runs temporarily via app_process and exits when monitoring completes.

audio capture — System Audio Capture (Android 11+)

Capture system audio via REMOTE_SUBMIX. Streams WAV (16kHz mono 16-bit PCM) to stdout for piping to external tools, or saves to a file.

WARNING: Device speakers are muted while capturing.

CODEBLOCK21

Stream mode (default): outputs WAV to stdout — designed for piping:

CODEBLOCK22

File mode (--file): saves to local file and returns JSON envelope with file path, byte count, and duration.

Default duration: 10s. Default sample rate: 16000 Hz.

Requires Android 11+ (API 30). Like monitor, this command pushes a small DEX helper to the device on first use.

When to use audio capture vs monitor:

• - monitor — read UI text (chat messages, labels, captions) as structured data
• INLINECODE53 — record what's being heard (speech, music, sound effects) as audio

For live streams, they complement each other: monitor captures on-screen chat text while audio capture captures the streamer's voice.

live cart — Douyin Shopping Cart Capture (抖音小黄车)

Grab product information from a Douyin live stream shopping cart. Captures the currently explaining product (without opening the cart) and the first N products in the cart list.

This command is Douyin-specific — it relies on Douyin's UI patterns and accessibility node structure.

CODEBLOCK23

How it works:

1. 1. Reads the "讲解中" floating card via accessibility (no cart open needed)
2. Taps the shopping cart button to open the panel
3. Scrolls slowly through products with continuous accessibility polling
4. Stops when products 1..N are all captured (consecutive, no gaps)
5. Closes the cart and outputs structured JSON

Returns: product number, title, price, sold count, shop name, tags for each product, plus the currently explaining product if any.

Must be used while viewing a Douyin live stream with a shopping cart.

shell — Run Raw Shell Command

Escape hatch for anything adb-claw doesn't have a dedicated command for.

CODEBLOCK24

Returns stdout, stderr, and exit_code in JSON envelope.

file — File Transfer

CODEBLOCK25

device — Device Info

CODEBLOCK26

ui — UI Element Inspection

CODEBLOCK27

Workflow Patterns

Always Observe First

Before any action, run observe to see the screen. After every action, observe again to verify.

CODEBLOCK28

Prefer Index-Based Targeting

Use --index N over coordinates. Indices from observe are stable across screen sizes.

Type After Focus

Always tap an input field first, then type:

CODEBLOCK29

Scroll Pattern

CODEBLOCK30

Always prefer scroll over manual swipe. After scrolling, always observe.

CJK Text Input

INLINECODE64 only supports ASCII. For Chinese/Japanese/Korean input:

1. 1. Preferred: Use adb-claw open with deep links (e.g., adb-claw open 'myapp://search?keyword=中文')
2. Clear before input: adb-claw clear-field --index 7 to clear existing text first

Wait for Page Load

Instead of repeated observe polling, use wait:

CODEBLOCK31

Device Form Factor Detection

Use adb-claw device info to get screen size, then determine form factor:

• - Short edge < 1200px → Phone (portrait-first)
• Short edge >= 1200px → Pad/Fold (landscape-first)

Swipe coordinates and UI layouts differ between Phone and Pad. App Profiles document these differences.

Audio Capture

First check if the device supports it:

CODEBLOCK32

To record audio to a file for the user:

CODEBLOCK33

To transcribe live audio (requires asrclaw installed separately):

CODEBLOCK34

Important: Device speakers are muted during capture. Inform the user before starting. For live streams, combine with monitor for complete context (audio + chat text).

Douyin Live Cart

Grab shopping cart products from a Douyin live stream:

CODEBLOCK35

The command handles everything automatically: opens cart, scrolls, captures, closes. No manual tapping needed.

Error Recovery

If an action fails or produces unexpected results:

1. 1. Run observe to see the current state
2. Check if the screen changed unexpectedly (dialog, permission prompt)
3. Adjust and retry

Output Format

All commands return JSON:

CODEBLOCK36

On error:

CODEBLOCK37

Security & Trust

What adb-claw is: A pure CLI wrapper around standard adb commands. It translates high-level instructions (e.g., adb-claw tap --index 3) into adb shell input tap ... calls. That's it.

What adb-claw does NOT do:

• - Does not install APKs or persistent services on the device — monitor and audio capture each push a temporary ~7KB DEX that runs briefly and auto-exits
• Does not collect or transmit data — no telemetry, no analytics, no network requests
• Does not request credentials or environment variables
• Does not modify your host system beyond placing the binary

Source code is fully open: github.com/llm-net/adb-claw. If you don't trust pre-built binaries, you can audit the source code and build from source — see the README for instructions. Every release also includes SHA256 checksums for binary verification.

Device sensitivity: adb-claw can capture screenshots, record system audio, and control apps on the connected device — this is the core purpose of the tool. Only connect devices you trust, and disable USB debugging when not actively using adb-claw.

Agent scope: adb-claw commands are the only commands this skill should execute. Do not run install scripts, download binaries, or modify the host system. If adb-claw or adb is not available, inform the user.

Troubleshooting

Problem	Solution
No devices found	Walk the user through: Settings → About phone → tap Build number 7x → Developer options → enable USB debugging → reconnect USB → approve the "Allow USB debugging" dialog on phone. See Setup section for full guide
adb not found

brew install android-platform-tools (macOS) | | Tap hits wrong element | Use --index instead of coordinates; re-run observe | | type doesn't work | Tap input field first to focus; ASCII only | | CJK text needed | Use adb-claw open with deep link containing the text as URL parameter | | UI dump fails | Pause animations (tap to pause video), wait 1s, retry | | UI dump fails on search pages | Search results may auto-play video previews; use screenshot instead or monitor to read text | | UI dump fails during live stream | Use monitor command — it bypasses video surfaces via accessibility framework | | Command timeout | Increase with --timeout 60000 | | Permission dialog | Use observe to see it, tap the allow/skip button | | Screen is off | adb-claw screen on or adb-claw screen unlock | | Audio capture fails | Requires Android 11+ (API 30); check with adb-claw device info | | No audio captured | REMOTE_SUBMIX may not be available on all devices/ROMs | | Speakers muted during capture | Expected behavior — REMOTE_SUBMIX redirects audio output |