Strands ADB

Give your agent a phone.

@tool decorated Android control for Strands Agents & DevDuck — drive any adb-connected Android device (phone, tablet, emulator) from an LLM.

The Pitch¶

Your agent can already read files, call APIs, and run shell commands. Now it can also:

See the screen (screenshots come back as Converse API image blocks, not paths)
Tap & swipe — real UI automation
Read notifications, battery, sensors, thermals
Launch apps, open URLs, send SMS drafts, make calls
Take physical photos (yes, through the actual camera)
Stream logcat into your agent's event bus
Mutate settings — brightness, ringer, airplane mode, bluetooth

One tool. One verb: action=.... Works over USB or wireless adb.

2-Minute Quickstart¶

pip install strands-adb
brew install android-platform-tools   # adb on PATH
adb devices                           # plug in phone, accept USB debugging

from strands import Agent
from strands_adb import adb

agent = Agent(tools=[adb])
agent("take a screenshot of my phone and describe what's on screen")

→ Full Quickstart | Installation | Connect a Device

👁️ Agent Can SEE the Screen¶

screenshot returns a proper Converse API image block — the same format as strands_tools.image_reader. The agent doesn't just get a file path, it actually receives the pixels and reasons over them.

agent("take a screenshot and tell me what app is open")
# → adb(action="screenshot") returns PNG bytes in Converse image block
# → vision model reads it → "You're on the WhatsApp chat with Mom..."

graph LR
    A["🗣️ Agent"] -->|action=screenshot| B["📱 adb exec-out screencap"]
    B -->|PNG bytes| C["🖼️ Converse image block"]
    C -->|in context| D["👁️ Vision model"]
    D -->|"WhatsApp, chat w/ Mom"| A

Capabilities¶

👁️ Vision

Screenshots as image blocks. Agent literally sees the screen.

→ Vision guide
🎯 Smart Tap

Find UI elements by text/content-desc, tap by semantic meaning.

→ Smart tap
📸 Physical Camera

Drive GoogleCamera via intent + shutter — real photos, real video.

→ Camera guide
🔔 Notifications & Logcat

Parse notifications, stream logcat events to DevDuck's event bus.

→ Logcat streaming
🌡️ Sensors & Thermals

Accelerometer, gyro, light, CPU temps — the phone as sensor platform.

→ Sensors
⚙️ Settings Mutation

Ringer mode, brightness, airplane, bluetooth — programmatic device state.

→ Settings
♿ Accessibility

Read screen via accessibility service, magnification, captions.

→ Accessibility
🤖 DevDuck Native

DEVDUCK_TOOLS="strands_adb:adb" — zero-config integration.

→ DevDuck integration

DevDuck in 1 Line¶

export DEVDUCK_TOOLS="strands_adb:adb;strands_tools:shell"
devduck "open whatsapp, read the last message from mom, reply 'on my way'"

→ DevDuck guide

90+ Actions, One Tool¶

adb(action="list_devices")
adb(action="screenshot")
adb(action="smart_tap", text="Send")
adb(action="launch", package="com.whatsapp")
adb(action="camera_photo", facing="front")
adb(action="notifications_parsed")
adb(action="sensors")
adb(action="set_brightness", value=128)
adb(action="log_stream_start", filter="WhatsApp")
adb(action="setting_put", namespace="global", key="airplane_mode_on", value="1")

→ Full actions overview | API reference

Architecture¶

graph TD
    AGENT["🗣️ Strands Agent"] -->|@tool call| ADB["strands_adb.adb"]
    ADB -->|dispatch action| ROUTER{Action Router}

    ROUTER -->|screenshot| SCREEN["📸 screencap → PNG bytes"]
    ROUTER -->|tap/swipe| INPUT["🎯 input tap/swipe"]
    ROUTER -->|launch| AM["📱 am start"]
    ROUTER -->|logcat| LC["📜 logcat -v"]
    ROUTER -->|sensors| DS["🌡️ dumpsys sensorservice"]

    SCREEN --> PHONE["📱 Android Device"]
    INPUT --> PHONE
    AM --> PHONE
    LC --> PHONE
    DS --> PHONE

    SCREEN -->|Converse image block| AGENT

→ Full architecture