Agent Droid Bridge

What Neverlow512/agent-droid-bridge MCP server does

Neverlow512/agent-droid-bridge MCP server connects an MCP-compatible AI client to Android devices and emulators through the Android Debug Bridge. It exposes structured operations for reading the current interface, sending input, launching applications, inspecting device state, and executing ADB commands.

The server is suited to mobile UI automation, app testing, development diagnostics, and dynamic analysis. An agent can inspect visible controls before acting, enter text into the focused field, send Android keycodes, capture screenshots, and check whether an action changed the interface. The optional app_manager tool pack adds package management, app lifecycle, APK extraction, permissions, and intent-related operations.

How it works

The server runs over standard input/output and communicates with the connected Android target through ADB. With one device connected, it detects that device automatically. If several devices are visible to ADB, the server supports choosing the target.

For content inspection, get_ui_hierarchy returns the current XML hierarchy, while get_screen_elements provides shaped element data and get_screen_text returns visible text in reading order. take_screenshot returns a PNG as base64 data along with the screen dimensions.

For efficient state checks, an agent can call snapshot_ui before an action and pass its token to detect_ui_change afterward. This avoids requesting the full hierarchy when the agent only needs to know whether the screen changed. detect_ui_change can also return the hierarchy when needed.

Setup and configuration

Neverlow512/agent-droid-bridge requires Python 3.11 or later and ADB installed on the host. The documented launcher is:

uvx agent-droid-bridge

Connect a physical Android device or start an emulator, then add the command and the agent-droid-bridge argument to the MCP client configuration. ADB_PATH can point to an ADB executable that is not on the system path. Set ADB_EXTRA_TOOL_PACKS to app_manager to load the additional application-management tools.

The server supports ADB_EXECUTION_MODE values of unrestricted and restricted. In restricted mode, shell commands must be permitted by the configured allowlist. ADB_ALLOW_SHELL=false disables shell commands regardless of execution mode. Optional JSONL session recording is enabled with MCP_LOG_ENABLED=true and requires a writable MCP_LOG_DIR.

Tools and capabilities

The core tool set includes:

• UI hierarchy, screen element, and visible-text inspection
• Taps, swipes, text entry, and Android key events
• PNG screenshots with dimensions and encoded image data
• Device listing and hardware, identity, and security capability checks
• App launching by package/activity component
• Parsed ADB command execution
• Lightweight UI snapshots and change polling

The execution layer parses commands with shlex rather than passing them directly to a system shell. Tool availability is reported in the server’s startup instructions.

Limitations and notes

The host must have ADB available and an Android device or emulator connected. The server does not provide a hosted device environment; it operates on targets visible to the local ADB installation. UI coordinates and app component names must correspond to the connected device and application. Arbitrary ADB access depends on the selected execution mode and shell settings, while the optional app-management features are not loaded unless explicitly enabled.