Android MCP Server

The Android diagnostic & forensic MCP — when an app crashes, leaks memory, drains battery, or behaves unexpectedly, this is what you point at the device.

76 tools across logcat / dumpsys / package internals / system properties / processes. 5 MCP Prompts (crash-investigation, memory-leak-detection, permission-audit, app-startup-profile, ui-element-locator) and a device-health aggregation. Pure ADB, no Appium / uiautomator2 dependency. 2-tier security (write + shell gates).

What it does that others don't

• Diagnostic depth — logcat search/crash extraction, dumpsys (mem/gfx/cpu), getprop, processes, package internals, app intents, port forwards. Cross-platform competitors hide this surface.
• MCP Prompts (5) — crash-investigation, memory-leak-detection, ui-element-locator, app-startup-profile, permission-audit. Workflow templates the model invokes directly.
• Aggregation tools — device-health (battery + memory + cpu + network in one call), analyze-app (package info + memory + activities).
• 2-tier security — ANDROID_MCP_ALLOW_WRITE (gates installs/taps/pushes) and ANDROID_MCP_ALLOW_SHELL (gates arbitrary adb shell) are separate flags. Distinct trust levels.
• Pure ADB — no Appium, no uiautomator2, no Python bridge. Just the official Android Debug Bridge wrapped over child_process.
• Token-efficient by design — 56 schema-trim sweep, ANDROID_TOOLS/ANDROID_DISABLE 9 categories, search-tools meta.
• Apps SDK card — device-health renders as a snapshot card on ChatGPT clients (battery, RAM, Wi-Fi + 4-section grid) via _meta["openai/outputTemplate"]. Claude clients receive the same JSON content.
• stdio + Streamable HTTP — defaults to stdio. Set MCP_TRANSPORT=http for ChatGPT Apps SDK or remote clients (Bearer auth via MCP_HTTP_TOKEN).

Try this — 5 prompts

Connect the server to Claude Desktop or Claude Code, then paste any of these:

1. Crash investigation — "My app com.us-all.api keeps crashing on this Pixel 6 emulator. Pull the last crash log, the offending stack frames, and any recent permission changes."
2. Memory leak detection — "Trace memory growth for com.us-all.api over the last 5 minutes. Show heap deltas, GC pressure, and the largest allocators."
3. Battery drain attribution — "What's draining battery on this device? Top 5 consumers, duration of each, and current battery health."
4. Permission audit — "Audit installed 3rd-party apps for dangerous permissions (location/camera/contacts/microphone). Flag any app that hasn't been used in the last 30 days but holds these."
5. App startup profile — "Profile the cold-start of com.us-all.api — measure activity launch time, identify the slowest fragment init, and suggest where to add tracing."

When to use this vs mobile-next/mobile-mcp

mobile-next/mobile-mcp (4.7K★) is the cross-platform action-oriented MCP. Different problem space:

Use both — they're complementary. mobile-mcp drives the device through your QA flows; this MCP tells you why it broke when something does. Especially:

• mobile-mcp finds the bug via UI exploration → this MCP captures the crash log + heap dump.
• mobile-mcp can't tell you why startup is slow → this MCP gives you dumpsys gfxinfo + activity launch timing.
• mobile-mcp can't reproduce a permission denial → this MCP shows the exact dumpsys package permission state and recent grants.

Install

Claude Desktop

{
  "mcpServers": {
    "android": {
      "command": "npx",
      "args": ["-y", "@us-all/android-mcp"],
      "env": {
        "ANDROID_MCP_ALLOW_WRITE": "true"
      }
    }
  }
}

Claude Code

claude mcp add android -s user \
  -e ANDROID_MCP_ALLOW_WRITE=true \
  -e ANDROID_MCP_ALLOW_SHELL=true \
  -- npx -y @us-all/android-mcp

Docker

docker run --rm \
  --device /dev/bus/usb \
  -e ANDROID_MCP_ALLOW_WRITE=true \
  ghcr.io/us-all/android-mcp-server:latest

Build from source

git clone https://github.com/us-all/android-mcp-server.git
cd android-mcp-server && pnpm install && pnpm build
node dist/index.js

Prerequisites

• ADB installed and on PATH (or set ADB_PATH)
• Android device or emulator with USB debugging enabled
• For multi-device setups: ANDROID_SERIAL=<serial> to target a specific one

Configuration

Categories (9): device, apps, ui, logcat, emulator, files, system, debug, shell (always-gated by ANDROID_MCP_ALLOW_SHELL), plus always-on meta.

When MCP_TRANSPORT=http: POST /mcp (Bearer-auth JSON-RPC) + GET /health (public liveness).

Token efficiency

Plus search-tools meta-tool (always enabled) for runtime tool discovery.

Read-only mode (default)

By default, only read operations are permitted. Write operations (tap, install-app, push-file, etc.) return an error unless ANDROID_MCP_ALLOW_WRITE=true. Shell command execution requires a separate ANDROID_MCP_ALLOW_SHELL=true for additional security — even with write enabled, raw shell stays blocked unless this is explicitly set.

MCP Prompts (5)

Workflow templates available via MCP prompts/list:

• crash-investigation — pull crash logs + stack frames + recent permission changes for a target package.
• memory-leak-detection — track heap delta over a window; cluster by allocator.
• ui-element-locator — find a UI element by visual + accessibility hints; return tap coordinates.
• app-startup-profile — cold-start profile: activity launch + fragment init + first frame.
• permission-audit — flag dangerous permissions held by under-used 3rd-party apps.

MCP Resources

URI-based read-only entities:

• android://devices — connected devices
• android://device/{serial} — device details (model/brand/version/display)
• android://app/{packageName}/activities — activities exposed by a package (exported/launchable flags)
• android://device/{serial}/processes — running processes

Tools (76)

9 categories. Use search-tools to discover at runtime; full list collapsed below.