Claude Mobile

MCP server for mobile, desktop, and browser automation — Android (ADB), iOS Simulator (simctl + WDA), Desktop (any macOS app), Aurora OS (audb), and Browser (CDP). Like Claude in Chrome but for devices, apps, and browsers.

Control your Android phone, emulator, iOS Simulator, desktop app, Aurora device, or headless browser with natural language through Claude.

Table of Contents

• Quick Start
• Features at a Glance
• Quality Engineering
• Installation
  • Homebrew (macOS)
  • One-liner (any client)
  • Claude Code
  • OpenCode
  • Other Agents (Pi, Qwen, Gemini, Codex, Cursor)
  • From npm / source
  • Windows
• Platform Guides
  • Android
  • iOS
  • Desktop
  • Browser
  • Aurora OS
• Tools Reference
  • Core Meta-Tools
  • Optional Modules
  • Flow Tools
• Native CLI
• Architecture
• License

Quick Start

# Install via Homebrew (macOS)
brew tap AlexGladkov/claude-in-mobile https://github.com/AlexGladkov/claude-in-mobile
brew install claude-in-mobile

# Verify dependencies
claude-in-mobile doctor

# Add to Claude Code
claude mcp add --scope user --transport stdio mobile -- npx claude-in-mobile@latest

Then talk to Claude naturally:

"Take a screenshot of the Android emulator"
"Tap on the Login button"
"Type hello in the search field"
"Switch to iOS simulator"

Features at a Glance

| Feature | Description | |---------|-------------| | Unified API | Same 8 meta-tools work across Android, iOS, Desktop, Aurora, and Browser | | Token-optimized | 8 meta-tools + 3 optional modules instead of 81 tools (~85% token reduction) | | Dynamic modules | Browser, Desktop, Store load on demand — default tool list stays lean | | Smart screenshots | Auto-compressed for optimal LLM processing | | Annotated screenshots | Colored bounding boxes + numbered element labels | | Security hardened | Shell injection protection, URL validation, path traversal blocking | | Structured errors | Typed error codes with auto-recovery hints | | Multi-device parallel | Run actions on multiple devices simultaneously | | Flow engine | Batch, conditional loops, and fan-out flows | | Permission management | Grant/revoke/reset app permissions (Android + iOS) | | Store publishing | Google Play, Huawei AppGallery, RuStore | | Telemetry | Per-tool call metrics via system(action:'metrics') | | Doctor command | claude-in-mobile doctor — checks all dependencies at once |

Quality Engineering

Advanced testing and monitoring built into Claude Mobile:

| Feature | What it does | How to use | |---------|-------------|------------| | Accessibility Auditing | WCAG 2.2 checks: missing labels, touch targets < 48px, focus order, duplicates | accessibility(action:'audit') | | Visual Regression | Baseline screenshots + pixel-level diff detection | visual(action:'baseline_save'), visual(action:'compare') | | Test Recorder | Record taps/swipes/input, replay without code | recorder(action:'start'), recorder(action:'play') | | Multi-Device Sync | Barrier-based coordination for parallel testing | sync(action:'create'), sync(action:'barrier') | | App Autopilot | Autonomous BFS/DFS exploration with self-healing locators | autopilot(action:'explore') | | Performance Monitor | Real-time memory, CPU, FPS tracking with snapshots | performance(action:'start'), performance(action:'snapshot') |

Installation

Homebrew (macOS)

brew tap AlexGladkov/claude-in-mobile https://github.com/AlexGladkov/claude-in-mobile
brew install claude-in-mobile

Verify setup:

claude-in-mobile doctor

One-liner (any client)

Auto-detects installed clients via add-mcp:

npx add-mcp claude-in-mobile -y

Target a specific client:

npx add-mcp claude-in-mobile -a claude-code -y
npx add-mcp claude-in-mobile -a opencode -y
npx add-mcp claude-in-mobile -a cursor -y

Claude Code

# Project-local
claude mcp add --transport stdio mobile -- npx claude-in-mobile@latest

# Global (all projects)
claude mcp add --scope user --transport stdio mobile -- npx claude-in-mobile@latest

Claude Code Plugin

claude plugin marketplace add AlexGladkov/claude-in-mobile
claude plugin install claude-in-mobile@claude-in-mobile

OpenCode

Two modes:

A) MCP server (Node.js):

opencode mcp add
# Choose local MCP → npx -y claude-in-mobile

Or in opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "mobile": {
      "type": "local",
      "command": ["npx", "-y", "claude-in-mobile"],
      "enabled": true
    }
  }
}

B) Native CLI + Skill (no Node.js needed):

claude-in-mobile setup opencode            # project-local
claude-in-mobile setup opencode --global   # user-wide

Other Agents

Native CLI skill works with any agent that supports Agent Skills:

claude-in-mobile setup pi --global       # Pi
claude-in-mobile setup qwen --global     # Qwen Code
claude-in-mobile setup gemini --global   # Gemini CLI
claude-in-mobile setup codex --global    # Codex
claude-in-mobile setup cursor --global   # Cursor

Drop --global for project-local install. Restart the agent after setup.

Qwen Code — .qwen/settings.json or ~/.qwen/settings.json:

{ "mcpServers": { "mobile": { "command": "npx", "args": ["-y", "claude-in-mobile"] } } }

Gemini CLI — .gemini/settings.json or ~/.gemini/settings.json:

{ "mcpServers": { "mobile": { "command": "npx", "args": ["-y", "claude-in-mobile"] } } }

Codex:

codex mcp add mobile -- npx -y claude-in-mobile

Cursor — .cursor/mcp.json:

{ "mcpServers": { "mobile": { "command": "npx", "args": ["-y", "claude-in-mobile"] } } }

From npm / source

# npm (no install)
npx claude-in-mobile

# From source
git clone https://github.com/AlexGladkov/claude-in-mobile.git
cd claude-in-mobile
npm install
npm run build:all

Using a local build with any MCP client:

{ "mcpServers": { "mobile": { "command": "node", "args": ["/path/to/claude-in-mobile/dist/index.js"] } } }

Windows

claude mcp add --transport stdio mobile -- cmd /c npx claude-in-mobile@latest

Platform Guides

Android

Requirements:

• ADB installed (auto-discovered or set ADB_PATH)
• USB debugging enabled on device, or running emulator

ADB discovery order:

| Priority | Location | |----------|----------| | 1 | ADB_PATH env var | | 2 | $ANDROID_HOME/platform-tools/adb | | 3 | $ANDROID_SDK_ROOT/platform-tools/adb | | 4 | OS default: ~/Library/Android/sdk (macOS), %LOCALAPPDATA%\Android\Sdk (Windows), ~/Android/Sdk (Linux) | | 5 | adb from PATH |

If none found → [ADB_NOT_INSTALLED] error with probed paths.

Examples:

"Show connected devices"
"Take a screenshot on Android"
"Tap on Settings"
"Swipe down to scroll"
"Type 'hello' in the search field"
"Press the back button"
"Grant camera permission to com.example.app"
"Launch com.example.app"

CLI:

claude-in-mobile screenshot android
claude-in-mobile tap android 540 960
claude-in-mobile input android "hello world"
claude-in-mobile ui-dump android | grep "Login"

Coordinate space (raw x/y in tap / swipe / long_press)

When you call an input tool with raw x/y (or x1/y1/x2/y2 for swipe), the values are interpreted in the most recent screenshot's pixel space and auto-scaled to device coordinates before dispatch. The scale comes from the last screen_capture call: e.g., capture at preset='low' (270×480) on a 1080×2400 device sets a 4× factor, so tap(135, 240) becomes tap(540, 960) on the device.

This is convenient for the common flow screen_capture → reason about pixel → tap, but has two gotchas worth knowing:

• Coordinates from ui_find / ui_tree are device coordinates, not screenshot coordinates. They come from uiautomator which always reports in device space. If the most recent screenshot was at a low preset, passing those device coords as raw x/y will over-scale them. Prefer index, text, or resourceId for ui-sourced taps to avoid the issue entirely.
• No screenshot taken yet? Then there's no scale stored, and raw x/y are passed through 1:1 as device coords.

The cleanest mental model: raw coords match whatever pixel space you're looking at on screen (your last screenshot). For everything else, use the resolver fields (index, text, resourceId, label).

iOS

Requirements:

• macOS with Xcode
• iOS Simulator (no physical device support yet)
• WebDriverAgent for full UI inspection (optional but recommended)

WebDriverAgent setup:

# Automatic (via Appium)
npm install -g appium
appium driver install xcuitest

# Or set custom path
export WDA_PATH=/path/to/WebDriverAgent

On first use, WDA is auto-built (~2 min one-time), launched on simulator, and connected on port 8100+.

What WDA enables:

• ui(action:'tree') — full accessibility tree
• ui(action:'find') — element discovery by label/text
• input(action:'tap', label:'...') — element-based tapping
• Improved swipe and gesture simulation

Troubleshooting:

# Install Xcode CLI tools
xcode-select --install

# Accept license
sudo xcodebuild -license accept

# Check simulator is booted
xcrun simctl list | grep Booted

# Check port
lsof -i :8100

cd ~/.appium/node_modules/appium-xcuitest-driver/node_modules/appium-webdriveragent
xcodebuild test -project WebDriverAgent.xcodeproj \
  -scheme WebDriverAgentRunner \
  -destination 'platform=iOS Simulator,id=<DEVICE_UDID>'

Examples:

"Take a screenshot on iOS"
"Open Safari on iOS"
"Tap on the Login button"
"Type my email in the text field"
"Swipe left on the card"
"Reset all permissions for com.apple.Maps"

Desktop

Requirements:

• macOS (Windows/Linux planned)
• Accessibility permissions: System Settings → Privacy & Security → Accessibility
• JDK 17+ (for building Desktop companion)

Supported apps: Any macOS application — SwiftUI, AppKit, Electron, Compose Desktop.

Launch modes:

| Mode | Example | |------|---------| | By bundleId | desktop(action:'launch', bundleId:'com.apple.Calculator') | | By .app path | desktop(action:'launch', appPath:'/Applications/Slack.app') | | Attach by PID | desktop(action:'launch', pid:12345) |

Enable the module first:

"Enable desktop module"

Or it auto-enables on first desktop(...) call.

Examples:

"Launch Calculator"
"Take a screenshot of the desktop app"
"Get window list"
"Resize window to 1280x720"
"Tap at 100, 200 on desktop"
"Get clipboard content"
"Get performance metrics"
"Stop the desktop app"

Full API documentation: docs/SPEC_DESKTOP.md

Browser

Requirements:

• Chrome or Chromium installed (or set CHROME_PATH)

Browser automation via Chrome DevTools Protocol (CDP). The browser module loads on demand.

Examples:

"Open https://example.com in the browser"
"Click the Sign In button"
"Fill the email field with [email protected]"
"Take a browser screenshot"
"Execute JS: document.title"
"Wait for the loading spinner to disappear"

Available actions:

| Action | Description | |--------|-------------| | open | Open URL in new session | | navigate | Go to URL in existing session | | click | Click element by ref | | fill | Type into input field | | fill_form | Fill multiple fields at once | | press_key | Keyboard input | | snapshot | DOM snapshot with element refs | | screenshot | Visual screenshot | | evaluate | Run JavaScript | | wait_for_selector | Wait for element to appear | | close | Close session | | list_sessions | Show active sessions | | clear_session | Reset cookies/storage |

Aurora OS

Requirements:

• audb CLI: cargo install audb-client
• SSH-enabled Aurora OS device
• Python on device for tap/swipe: devel-su pkcon install python

Examples:

"List Aurora devices"
"Take a screenshot on Aurora"
"Tap at 100, 200 on Aurora"
"Launch ru.example.app on Aurora"
"List installed apps on Aurora"
"Get logs from Aurora device"
"Push file.txt to /home/defaultuser/"

Tools Reference

v3.8.0 provides 8 core meta-tools + 3 optional modules. Each meta-tool uses an action parameter.

Core Meta-Tools

| Meta-Tool | Actions | Description | |-----------|---------|-------------| | device | list, set, set_target, get_target, enable_module, disable_module, list_modules | Device management, module control | | input | tap, double_tap, long_press, swipe, text, key | Touch and keyboard input | | screen | capture, annotate | Screenshots and visual annotation | | ui | tree, find, find_tap, tap_text, analyze, wait, assert_visible, assert_gone | UI hierarchy, element interaction | | app | launch, stop, install, list | App lifecycle | | system | activity, shell, wait, open_url, logs, clear_logs, info, webview, clipboard_*, permission_*, file_*, metrics, reset_metrics | System ops, clipboard, permissions, files, telemetry | | flow_batch | — | Execute multiple commands in one round-trip (max 50) | | flow_run | — | Multi-step automation with conditionals and loops (max 20 steps) |

Optional Modules

Load on demand via device(action:'enable_module', module:'<name>') or auto-enable on first call.

| Module | Actions | Description | |--------|---------|-------------| | browser | open, close, list_sessions, navigate, click, fill, fill_form, press_key, snapshot, screenshot, evaluate, wait_for_selector, clear_session | Chrome/Chromium via CDP | | desktop | launch, stop, windows, focus, resize, clipboard_get, clipboard_set, performance, monitors | Any macOS app | | store | upload, set_notes, submit, get_releases, discard, promote, halt_rollout, get_versions | Google Play, Huawei AppGallery, RuStore |

Flow Tools

| Tool | Description | |------|-------------| | flow_batch | Sequential execution, one round-trip (max 50 commands) | | flow_run | Multi-step flows with if_not_found, repeat, on_error (max 20 steps) | | flow_parallel | Same action on multiple devices via Promise.allSettled (max 10) |

Backward Compatibility

All v3.0/v3.1 tool names work as aliases: tap → input(action:'tap'), screenshot → screen(action:'capture'), launch_app → app(action:'launch'), etc.

Native CLI

2 MB Rust binary. No Node.js, no dependencies.

Install

brew tap AlexGladkov/claude-in-mobile
brew install claude-in-mobile

Or download from Releases.

Why use the CLI

| | CLI | MCP Server | |---|---|---| | Install | brew install or copy binary | npx / npm | | Dependencies | None | Node.js | | Startup | ~5ms | ~500ms | | Use from terminal | Direct commands | Needs MCP client | | CI/CD | Exit codes, stdout/stderr | Not designed for CI | | Token cost | Skill loads on demand | Schema always present |

Test script example

#!/bin/bash
claude-in-mobile launch android com.example.app
claude-in-mobile wait 2000
claude-in-mobile tap android 0 0 --text "Login"
claude-in-mobile input android "[email protected]"
claude-in-mobile screenshot android -o result.png
claude-in-mobile ui-dump android | grep "Welcome" && echo "PASS" || echo "FAIL"

Store management (CLI)

claude-in-mobile store upload --package com.example.app --file app.aab
claude-in-mobile huawei upload --package com.example.app --file app.aab
claude-in-mobile rustore upload --package com.example.app --file app.apk

Doctor

Check all dependencies at once:

claude-in-mobile doctor

Checks: ADB, ANDROID_HOME, Xcode, simctl, Appium, WDA, JDK, audb-client, Chrome. Color-coded output with fix suggestions.

Architecture

┌─────────────┐     ┌──────────────────┐     ┌─────────────────┐
│ Claude Code │────▶│                  │────▶│  Android (ADB)  │
├─────────────┤     │  Claude Mobile   │     ├─────────────────┤
│  OpenCode   │────▶│   MCP Server     │────▶│ iOS (simctl+WDA)│
├─────────────┤     │                  │     ├─────────────────┤
│   Cursor    │────▶│  8 meta-tools    │────▶│ Desktop (macOS) │
├─────────────┤     │  + 3 modules     │     ├─────────────────┤
│ Qwen/Gemini │────▶│                  │────▶│ Aurora (audb)   │
├─────────────┤     │  Auto-detects    │     ├─────────────────┤
│  Any MCP    │────▶│  platform        │────▶│ Browser (CDP)   │
└─────────────┘     └──────────────────┘     └─────────────────┘

1. Client sends commands via MCP protocol (8 meta-tools + 3 optional modules)
2. Server routes to platform adapter (ADB, simctl+WDA, Desktop, audb, CDP)
3. Commands execute on device/app/browser
4. Results (screenshots, UI trees, metrics) return to client
5. Modules auto-enable on first call — no manual setup needed

License

MIT