mobile-use-mcp

mobile-use-mcp is a local stdio MCP server that lets coding agents such as Codex and Claude Code observe and operate Android devices through ADB and uiautomator2.

The server contains no LLM and requires no model API key. The host agent interprets screenshots and UI elements, plans actions, and calls MCP tools directly.

Capabilities

• Discover, select, and disconnect Android physical devices and emulators

• Return screenshots as native MCP image content

• Encode original-resolution screenshots as PNG or quality-controlled JPEG

• Return compact or full UIAutomator accessibility hierarchies with query and pagination

• Tap or long press using bounds, resource ID, or text fallbacks

• Swipe using validated screen coordinates

• Type and clear text in the focused field

• Press a bounded set of Android keys

• List, launch, and terminate apps

• Diagnose App launch attempts and foreground blockers

• Open HTTP and HTTPS URLs

• Record bounded Android screen videos with automatic segment rollover

• Detect unauthorized, offline, disconnected, and ambiguous device states

Related MCP server: Android Mobile MCP

Non-goals and safety boundaries

The initial release intentionally excludes:

• iOS, IDB, and WebDriverAgent

• Embedded LLMs or model provider configuration

• Arbitrary ADB shell access

• APK installation or app uninstall

• Clearing app data, rebooting, shutting down, or factory reset

• Cloud devices, BrowserStack, Limrun, and telemetry

The server never automatically uninstalls software from the connected device. Screenshots and UI text can contain sensitive information and are returned to the host agent, so use a trusted agent and model environment.

The MCP initialize response includes service instructions that tell compatible host agents to use these tools for natural-language requests involving a physical Android phone or mobile App, such as opening an App, browsing content, reading comments, collecting visible data, tapping, swiping, typing, screenshots, and screen recording. Users normally should not need to name mobile-use or ADB explicitly, although tool selection ultimately remains under the host agent's control.

Requirements

• macOS, Linux, or Windows with Python 3.12+

• uv

• Android SDK platform-tools with adb on PATH

• A physical Android device with USB debugging enabled, or an Android emulator

• Optional: ffmpeg on PATH to merge recordings longer than one Android segment

Confirm ADB connectivity before starting:

adb devices -l

Expected online state:

List of devices attached
ABC123 device product:... model:... transport_id:1

If the state is unauthorized, unlock the phone and accept the USB debugging prompt. If it is offline, reconnect the device or restart the ADB server.

Installation

Clone the repository and install the locked environment:

git clone https://github.com/NeoAgentman/mobile-use-mcp.git
cd mobile-use-mcp
uv sync

The MCP server entry point is:

uv run mobile-use-mcp

It uses stdio and is intended to be started by an MCP client, not used as an interactive CLI.

Codex configuration

Add the local server with the Codex CLI, replacing the path with your clone location:

codex mcp add mobile-use -- \
  uv --directory /absolute/path/to/mobile-use-mcp run mobile-use-mcp

Equivalent ~/.codex/config.toml configuration:

[mcp_servers.mobile-use]
command = "uv"
args = [
  "--directory",
  "/absolute/path/to/mobile-use-mcp",
  "run",
  "mobile-use-mcp",
]

Start a new Codex task after adding the server so the tool list is refreshed.

Claude Code configuration

Add the same stdio server to Claude Code:

claude mcp add --scope user mobile-use -- \
  uv --directory /absolute/path/to/mobile-use-mcp run mobile-use-mcp

Use --scope project instead if the configuration should only apply to one project.

Recommended agent workflow

1. Call android_list_devices.

2. Call android_connect, specifying a serial when more than one device is online.

3. Call android_snapshot to receive the screenshot and current UI elements.

4. Prefer a target containing bounds plus resource ID or text.

5. Perform one action.

6. Call android_snapshot again to verify the resulting state.

7. If an action fails, use the returned selector attempts and refresh the snapshot.

android_wait is only a fixed delay. It does not wait for a condition or inspect whether an element appeared, and it invalidates the cached snapshot. Always observe again afterward.

Example instruction for the host agent:

Use the mobile-use MCP tools to open Android Settings, navigate to About phone,
and report the device model. Observe the screen again after every action.

Tools

Device lifecycle

Observation

android_snapshot defaults to detail_level="compact", returns at most 200 elements, and reports snapshot_id, total_elements, returned_elements, truncated, and next_offset; no truncation is silent. The default max_text_length=500 is applied separately to each node's text and content description, including in full mode. Increase it up to 2000 when reading long-form content.

When truncated=true, first query or page the same snapshot with android_get_ui_elements; use detail_level="full" only as an explicit large-output fallback. interactive_only=true keeps only clickable, focusable, or scrollable nodes, so leave it false when reading static content such as articles, product descriptions, prices, tables, or comments.

android_snapshot and android_screenshot support image_format and image_quality. Screenshots always keep the original device resolution. The default is JPEG quality 60; use image_format="png" when small text, tables, icons, or other visual details require the original lossless image. Every JPEG response includes a quality_notice and machine-readable lossless_fallback with the exact PNG retry arguments. Screenshot base64 is not duplicated in structured JSON. Use android_snapshot when UI text or targets are needed; use android_screenshot for visual-only inspection without a pageable hierarchy.

Use android_get_ui_elements with the returned snapshot_id to page the exact same hierarchy. Its default page size is 100. It supports offset, limit, a case-insensitive query across text, content description, resource ID, class, and package, plus exact package and interactive-only filters. Filters are applied before pagination; reset offset to zero when changing a filter:

{
  "snapshot_id": "s-...",
  "query": "校招",
  "package": "com.taobao.idlefish",
  "offset": 0,
  "limit": 50
}

Only the latest snapshot is cached. Actions, waiting, reconnecting, and disconnecting invalidate it; an expired ID returns SNAPSHOT_NOT_FOUND and instructs the agent to observe again.

Actions

android_launch_app reports every launch attempt, polling count, and the last foreground App. A permission controller, chooser, or other foreground blocker is reported separately from an App that remains in an indeterminate loading state.

Recordings are limited to 5–1800 seconds. Device-side temporary files use randomized names and are removed after transfer. Android's per-recording time limit is handled with 170-second segments. If ffmpeg is available, multiple segments are losslessly concatenated; otherwise the tool returns the segment paths and a warning. Returned paths are local to the MCP host and may contain sensitive screen content.

Target selection

Tap and long-press tools accept a target object:

{
  "bounds": {"x": 20, "y": 100, "width": 280, "height": 80},
  "resource_id": "com.example:id/continue",
  "resource_id_index": 0,
  "text": "Continue",
  "text_index": 0
}

The server tries selectors in this order:

1. Valid on-screen bounds

2. Resource ID and occurrence index

3. Exact case-insensitive text/content description and occurrence index

Target does not accept content_description, class_name, or package fields. To act on a snapshot node whose accessible label is in content_description, pass that value through the target's text field. class_name and package are available as query filters through android_get_ui_elements, not as action selectors. Bounds use original device pixels and become stale after the UI changes.

Failure responses include every attempted selector and recommend taking a fresh snapshot.

Error model

Expected operational failures return structured data instead of internal tracebacks:

{
  "success": false,
  "error_code": "DEVICE_DISCONNECTED",
  "message": "Android device 'ABC123' is no longer connected and ready.",
  "suggestion": "Reconnect the device, verify `adb devices -l`, then call android_connect again.",
  "data": {}
}

Stable error codes include:

• ADB_UNAVAILABLE

• DEVICE_NOT_FOUND

• DEVICE_UNAUTHORIZED

• DEVICE_OFFLINE

• DEVICE_DISCONNECTED

• MULTIPLE_DEVICES

• NOT_CONNECTED

• INVALID_TARGET

• ELEMENT_NOT_FOUND

• INVALID_COORDINATES

• OPERATION_FAILED

• TIMEOUT

• UNSUPPORTED

• SNAPSHOT_NOT_FOUND

Development and verification

Install development dependencies:

uv sync --dev

Run all non-device checks:

uv run ruff check .
uv run ruff format --check .
uv run pyright
uv run pytest

Inspect the live stdio MCP schema:

npx -y @modelcontextprotocol/inspector --cli \
  uv --directory "$PWD" run mobile-use-mcp \
  --method tools/list

Tests marked android require a connected device or emulator:

MOBILE_USE_ANDROID_SERIAL=ABC123 uv run pytest -m android

Device tests are excluded from the default test run and require an explicit serial so the test suite never selects and operates a connected phone accidentally.

See COMPATIBILITY.md for current Codex, Claude Code, MCP client, and real device verification results.

Known limitations

• Custom-rendered Canvas/game interfaces may not expose useful accessibility elements; coordinate actions can still be selected from the screenshot.

• Screen recording availability and supported resolution depend on the Android device's built-in screenrecord implementation.

• UI elements are snapshots, not stable DOM nodes. Observe again after navigation, animation, or scrolling. Only the latest snapshot_id can be paged, and state-changing operations invalidate it deliberately.

• detail_level="full" can produce a large tool result. Prefer pagination and query first.

• max_text_length is a per-node limit, not a total response budget; full mode does not disable it.

• interactive_only=true intentionally hides most static text nodes and should not be used for long-form reading or content extraction.

• Screenshots always use the phone's original resolution. JPEG is lossy by default; retry with PNG when the returned quality notice indicates that visual detail may be insufficient.

• android_type_text and android_clear_text accept an optional target. Without one, they operate on the currently focused field.

• App discovery currently returns package names rather than localized display names.

• Screenshot tools return visible pixels as MCP image content; they do not extract or download an App's original remote media file.

• Physical-device and emulator compatibility still requires validation across Android versions, OEM ROMs, and input methods.

Attribution

Android controller and selector behavior is derived from minitap-ai/mobile-use, licensed under the Apache License 2.0. See LICENSE and NOTICE for attribution and license details.