adbridge

Android Bridge for AI-Assisted Development. A CLI tool and MCP server that gives AI coding assistants direct access to your Android device: screenshots, OCR, logcat, input control, and device state inspection.

No more manually screenshotting, copy-pasting logs, or describing what's on screen. Just ask your AI assistant to look at the device.

What it does

adbridge runs as a standalone CLI or as an MCP server that exposes your connected Android device as structured, queryable tools. Any MCP-compatible AI tool (Claude Code, Cursor, Cline, etc.) can then:

• Capture screenshots (auto-compressed JPEG for token efficiency) and extract text via OCR
• Parse interactive UI elements with tap coordinates from the view hierarchy
• Read filtered logcat entries
• Send taps, swipes, keystrokes, and text to the device
• Inspect the current activity, fragment backstack, and memory stats
• List connected devices with model and version info
• Pull crash reports with full context

              ADB protocol
 Android ◄──────────────────► adbridge
 Device                          │
                          ┌──────┴──────┐
                          │             │
                        CLI        MCP Server
                      (human)     (AI tools)

Install

Prerequisites

• Rust toolchain (1.75+): https://rustup.rs
• ADB server running (adb start-server)
• Tesseract OCR (for --ocr feature):

  # Arch
  sudo pacman -S tesseract tesseract-data-eng
  
  # Ubuntu/Debian
  sudo apt install tesseract-ocr libtesseract-dev libleptonica-dev
  
  # macOS
  brew install tesseract
  

From crates.io

cargo install adbridge

From source

git clone https://github.com/Slush97/adbridge.git
cd adbridge
cargo install --path .

Verify

adbridge --version
adbridge --help

CLI Usage

Screenshot + OCR

# Capture screenshot, save to file
adbridge screen --output screenshot.png

# Capture with OCR text extraction
adbridge screen --ocr

# Full context: screenshot + OCR + view hierarchy as JSON
adbridge screen --ocr --hierarchy --json

# Parsed interactive UI elements with tap coordinates
adbridge screen --elements

Logcat

# Recent errors
adbridge log --level error --lines 20

# Filter by app
adbridge log --app com.myapp --level warn

# Filter by tag, JSON output
adbridge log --tag NetworkManager --json

Input

# Type text on device
adbridge input text "hello world"

# Tap coordinates
adbridge input tap 540 1200

# Swipe (scroll down)
adbridge input swipe 540 1500 540 500 --duration 300

# Hardware keys
adbridge input key home
adbridge input key back

# Set clipboard
adbridge input clip "copied text"

Device State

# Current activity, fragments, display info
adbridge state

# Include memory stats, as JSON
adbridge state --memory --json

Devices

# List connected devices
adbridge devices

# As JSON
adbridge devices --json

Crash Report

# Full crash context: stacktrace + recent errors + screenshot
adbridge crash --json

# Pipe into an AI for analysis
adbridge crash --json | claude "what caused this crash?"

MCP Server

adbridge exposes 7 tools over MCP's stdio transport:

Claude Code

Add to ~/.mcp.json:

{
  "mcpServers": {
    "adbridge": {
      "command": "adbridge",
      "args": ["serve"]
    }
  }
}

Restart Claude Code. You can now say things like:

"What's on the phone screen right now?"

"Check the logcat for errors in my app"

"Tap the login button and tell me what happens"

"The app just crashed, what went wrong?"

Other MCP clients

Any client supporting MCP stdio transport can use adbridge. The server starts with:

adbridge serve

How it works

• ADB communication via adb_client, native Rust ADB protocol with no adb binary dependency (ADB server still required)
• OCR via leptess, FFI bindings to Tesseract/Leptonica
• Image processing via image for JPEG compression and downscaling
• MCP server via rmcp, the official Rust MCP SDK
• CLI via clap

All device commands go through adb shell under the hood. The tool structures the raw output into JSON that AI assistants can reason about.

Token optimization

The MCP server is designed to minimize token usage when communicating with AI assistants:

• Screenshots are downscaled to 720px width and compressed to JPEG (80% quality) instead of full-resolution PNG
• View hierarchy XML is stripped of default/false attributes (checkable="false", enabled="true", empty strings, etc.), typically reducing size by 50%+
• OCR output is post-processed to remove noise lines (garbage from non-text screens like wallpapers)
• UI elements are auto-included by default as a compact, structured alternative to raw hierarchy XML

Project Structure

src/
├── main.rs            Entry point, CLI dispatch
├── cli.rs             Clap argument definitions
├── adb/
│   ├── mod.rs         Core ADB shell commands
│   └── connection.rs  Device discovery and info
├── screen/
│   ├── mod.rs         Screenshot, OCR, hierarchy stripping, image compression
│   └── elements.rs    UI element parser (view hierarchy to compact format)
├── logcat/mod.rs      Log parsing and filtering
├── input/mod.rs       Text, tap, swipe, key, clipboard
├── state/mod.rs       Activity state, memory, crash reports
└── mcp/mod.rs         MCP server with 7 tools (stdio)

License

MIT