androir-mcp

Drive real Android devices from any AI agent — see the screen, tap, swipe, type, over adb.

An independent, standalone Model Context Protocol server that gives an AI agent a clean, safe automation surface for physical Android devices and emulators: capture the screen, read the live UI tree, and drive input — backed entirely by adb + uiautomator, no native code and no device-side app.

CI License Node TypeScript MCP PRs welcome Stars

[!NOTE] Requires adb (Android platform-tools) and an Android device with USB debugging enabled (or an emulator). The server runs locally, holds no credentials, and talks only to the device(s) adb already sees.

✨ What is androir-mcp?

androir-mcp is a standalone MCP server for automating Android over adb. It is its own project — not a copy or port of anything — exposing a small, conventional screen-read + input tool set that drops into any MCP-based agent loop.

Everything is backed by adb and uiautomator: there is no native code and nothing to install on the device. The agent shells out to adb (always as an argv array — never a shell string) to capture screenshots, dump the UI hierarchy, and send input events.

The key advantage on Android is that describe_screen returns the real UI tree — exact element bounds and text straight from uiautomator — rather than guessing from OCR. Tap coordinates come back as element centers and feed straight into tap, so an agent can read a screen and act on it deterministically.

back to top ↑

🚀 Features

• 📸 screenshot — capture the device screen as a PNG (signature-validated).
• 🌳 describe_screen — parse the uiautomator UI tree into a flat list of labels with center tap coordinates (text / content-desc / class, entity-decoded).
• 👆 tap / swipe / long_press — coordinate input in device pixels.
• ⌨️ type_text — type arbitrary text safely (spaces, metacharacters, and unicode all type verbatim).
• 🔘 press_key / press_home / press_back — named key events.
• 📱 launch_app — launch by package name or a friendly name (resolved from the installed package list, cached per device).
• 🔗 open_url — open an http(s) URL in the default browser.
• 🔎 list_targets / status — enumerate connected devices and read device state, properties, and battery.

Safety by construction:

• 🛡️ argv-only adb — every command is an argument array, never a shell string, and any value handed to the device shell is single-quoted for it, so there is no shell-injection surface.
• ✅ strict serial validation — serials are checked against [A-Za-z0-9.:_-] (≤ 128 chars) before reaching any subprocess.
• ⏱️ per-call timeout (default 30 s) with process-group kill on timeout.
• 🤫 scrubbed errors — concise messages only; no raw adb stderr (which can leak serials/paths) and no host stack traces reach the model.

back to top ↑

🛠️ Tools

All tools take an optional serial; it defaults to the single connected device and errors if the choice is ambiguous. Coordinates are in device pixels, so describe_screen tap points feed straight into tap with no translation.

back to top ↑

🏁 Getting started

Prerequisites

• adb (Android platform-tools) on your PATH.

• An Android device with USB debugging enabled, or a running emulator. Confirm it's visible:

  adb devices
  

Build

npm install
npm run build

Attach to an MCP client

Add the built server to your MCP client config:

{
  "mcpServers": {
    "androir": {
      "command": "node",
      "args": ["/path/to/androir-mcp/dist/index.js"]
    }
  }
}

Verify

With a device connected, run the self-check (it lists targets, takes a screenshot, and dumps the UI tree):

npm run selfcheck

It prints PASS when the three core tools work end-to-end against a real device.

back to top ↑

🤖 Use with an AI agent

Once androir is attached to your MCP client, give the agent a goal and let it read the screen and act:

You: Open the Settings app, go to Wi-Fi, and tell me which network is connected.

Agent: calls launch_app("settings") → describe_screen() (reads the labels + tap coordinates) → tap(x, y) on "Wi-Fi" → describe_screen() again → reports the connected network.

Because describe_screen returns exact element bounds and text, the agent taps real coordinates rather than guessing from a screenshot.

back to top ↑

🗺️ Status & roadmap

• Core tools — screenshot, describe_screen, tap, swipe, long_press, type_text, press_key/home/back, launch_app, open_url, list_targets, status
• uiautomator XML parsing → labels + center tap coordinates (entity-decoded, malformed-XML tolerant)
• Device-shell-safe input escaping (no shell injection) + strict serial validation
• Unit tests for the XML parser and the input-escaping logic
• Pre-push secret-scan hook
• Recording / replay skills
• Multi-device parallel control
• Emulator-specific paths
• Published npm package / npx bin

This README and roadmap fill in as the project progresses.

back to top ↑

📄 License

Apache-2.0.

back to top ↑