Skip to main content
Glama

admcp

admcp is a local-first Model Context Protocol (MCP) server for Android development. It connects any MCP stdio client to a real device or emulator and exposes 51 typed tools for UI inspection, screenshots, logcat, Java/Kotlin JDI debugging, native LLDB, Flutter/Dart DAP, Compose inspection, Perfetto traces, heap dumps, artifacts and evidence-based diagnostics.

No root. No cloud account required.

MCP Android JDK Tools License

Quick start | What it can do | Example prompts | Why admcp | Documentation

Status: v1, 51 tools, live-tested on a OnePlus CPH2747 (API 36, arm64-v8a). Full evidence and remaining limits: implementation status.

Quick start

  1. Install JDK 21 and Android SDK platform-tools (adb).

  2. Download admcp-<version>.zip from GitHub Releases and unzip it. MCPB-compatible clients can install admcp-<version>.mcpb from the same release instead.

  3. Point your MCP client at bin/admcp.bat on Windows or bin/admcp on Linux and macOS.

Any MCP stdio client can use it. The JSON below uses the common mcpServers shape; adapt the command and arguments to your client (Claude Code, Cursor, VS Code, Codex, Gemini CLI, and similar).

{
  "mcpServers": {
    "admcp": {
      "command": "C:\\Tools\\admcp\\admcp-<version>\\bin\\admcp.bat",
      "args": [
        "--platform-tools", "C:\\Android\\Sdk\\platform-tools",
        "--artifacts", "C:\\temp\\admcp-artifacts"
      ]
    }
  }
}

On Linux and macOS, use /opt/admcp/admcp-<version>/bin/admcp as command.

Verify the install:

C:\Tools\admcp\admcp-<version>\bin\admcp.bat --selftest

On Linux and macOS:

/opt/admcp/admcp-<version>/bin/admcp --selftest

You should see catalog=51 registered=51 version=<version> on stderr. The server then speaks MCP over stdio; stdout is reserved for protocol traffic.

Related MCP server: adb-mcp

What it can do

Area

Tools

Device and workspace

device_list, device_inspect, toolchain_probe, workspace_open, workspace_describe

App lifecycle

app_build, app_install, app_launch

Sessions

session_start, session_status, session_stop

UI and screenshots

ui_inspect, ui_find, ui_act, ui_screenshot, compose_inspect

Logs

logcat_start, logcat_query, logcat_stop, logcat_statistics, logcat_health

Java/Kotlin debug

debug_attach, breakpoints, debug_step, debug_stack, debug_variables, debug_evaluate

Native debug

LLDB attach, breakpoints, stack, registers, memory reads and modules

Flutter debug

DART_DAP attach, breakpoints, stack, variables, step, pause and detach

Performance

Perfetto system traces, heap dumps and performance_trace_analyze

Artifacts and evidence

artifact_get, analyze_artifact, previews, diagnostics_* and signals_crashes

The full 51-tool contract is in docs/architecture/mcp-tool-catalog-v1.md. The catalog is validated by contract tests.

Example prompts

Once admcp is connected, ask your agent things like:

Inspect the current screen, find the INCREMENT button, tap it, and verify the counter changed.
Start logcat, reproduce the crash, then summarize the stack trace and the last 50 lines before it.
Launch the app, attach JDI, set a breakpoint in MainActivity, and show locals when it hits.
Inspect this Compose screen and tell me which accessibility node is the DELETE COUNTER button.
Record a 5-second Perfetto trace and summarize what happened.
Attach LLDB to the native library, break in fixture_increment, and show the registers.

Why admcp

Many Android MCP tools are thin command wrappers. admcp keeps device, workspace, process and session state together, then exposes debugging primitives with typed schemas, artifacts and safety gates.

It is for:

  • Android developers who want Claude Code, Cursor or another agent to inspect and debug a real device.

  • Agent builders who need stable, typed Android actions instead of parsing adb strings.

  • QA and automation workflows that need reproducible evidence from a device.

adb and shell scripts

Android Studio

admcp

Built for AI agents

No

No

Yes

Typed MCP tool schemas

No

No

51 tools

Session, process and backend state

No

Partial

Yes

JDI, LLDB and Flutter over MCP

No

Human UI

Yes

Evidence, artifacts and diagnostics

No

Partial

Yes

Local-first, no cloud account

Yes

Yes

Yes

Android Studio is the better tool for a human sitting at the IDE. admcp is for the agent that needs to inspect, act and debug without a human driving every click.

How it works

flowchart LR
    Client["MCP client / agent"] -->|stdio| Server["admcp (MCP server)"]
    Server --> ADB["ADB / Android CLI"]
    Server --> JDI["JDI (JDWP client)"]
    Server --> LLDB["lldb-dap"]
    Server --> DAP["Flutter DAP"]
    ADB --> Device["Android device / emulator"]
    JDI -.->|via adb forward| ADB
    LLDB -.->|via adb forward / lldb-server| ADB
    DAP -.->|via adb forward| ADB

The server runs on your machine next to the Android toolchain. It does not upload your project or device data anywhere.

Verified

  • 186 tests with 0 failures on the verified host.

  • The catalog registers all 51 tools over MCP stdio.

  • Live device: OnePlus CPH2747, API 36, arm64-v8a.

  • JDI: attach, breakpoints, step, stack, locals and read-only evaluation.

  • LLDB: attach, breakpoints, stack, registers, bounded memory reads and modules.

  • Flutter: DART_DAP breakpoints, stack, variables, step, pause and detach.

  • Compose: accessibility-tree inspection and verified UI actions.

  • Performance: Perfetto capture/analysis and HPROF heap-dump capture.

  • Evidence: logcat, crash buffer, artifacts, session timeline and diagnostics.

Device-dependent integration tests skip when no device is attached. See implementation status for the full acceptance evidence.

Requirements

  • JDK 21.

  • Android SDK platform-tools (adb).

  • Windows, Linux or macOS.

  • Optional: Android CLI, Flutter SDK, NDK lldb/lldb-server, lldb-dap and a connected device or emulator.

Only debuggable developer builds are supported. Root and Android security bypasses are out of scope.

Build from source

The Gradle wrapper is checked in; no local Gradle install is required.

Windows:

gradlew.bat clean test
gradlew.bat :app:admcp:installDist

Linux and macOS:

./gradlew clean test
./gradlew :app:admcp:installDist

The installable distribution is written to app/admcp/build/install/admcp/. The release archive is built with :app:admcp:distZip and written to app/admcp/build/distributions/admcp-<version>.zip.

Releases are automated with semantic-release from Conventional Commits on pushes to main or master: fix: is a patch release, feat: is a minor release and BREAKING CHANGE: is a major release. No release PR or changelog is generated.

Configuration

Option

Purpose

--platform-tools <dir>

Directory containing adb

--workspace <dir>

Default workspace root

--android-cli <path>

Android CLI executable

--flutter-root <dir>

Flutter SDK root

--artifacts <dir>

Artifact store root

--trust-workspace <dir>

Operator-approved workspace root; repeatable

--selftest

Print catalog/registration counts and exit

--version

Print the version and exit

workspace_trust only succeeds for a canonical root passed to the server at startup with --trust-workspace. A tool argument cannot self-assert approval.

Known limitations

  • The verified device is a OnePlus CPH2747 on API 36; other OEM, API and transport combinations are not claimed.

  • Simultaneous JDI and LLDB inspection is not reliable on the verified device. Sequential switching works with debug_attach(switchBackend=true).

  • Compose TREE uses the accessibility hierarchy. Internal semantics, slot table and recomposition data require app instrumentation.

  • No host trace_processor_shell is bundled, so trace analysis returns a bounded summary unless Perfetto SQL tooling is installed.

  • Crashlytics issue clusters require android auth login; the local crash buffer remains the default evidence source.

  • Native LLDB needs a separately installed lldb-dap and an ABI-matched lldb-server.

  • The host confirmation dialog requires an interactive desktop. Headless servers fail closed to a one-shot token.

See implementation status for details and exact error behavior.

Security

  • Only debuggable developer builds are supported.

  • Project-controlled code runs only after an explicit workspace trust grant.

  • Destructive UI actions require host confirmation or a one-shot token bound to the session and action.

  • Artifact reads are confined to the store root, capped and redacted by default.

  • The server is local-first: no cloud account is required.

Documentation

Contributing

Issues and pull requests are welcome. The public tool catalog is versioned; open an issue before adding or renaming a public tool. Run gradlew test before sending a PR.

If admcp saves your agent time, a star helps more Android developers find it.

License

Apache License 2.0. See LICENSE.

Related MCP Connectors

Related MCP Servers

  • A
    license
    B
    quality
    C
    maintenance
    An MCP server that enables AI assistants to build, test, and debug Android applications by interacting directly with the Android development environment. It provides tools for managing emulators, executing Gradle tasks, running ADB commands, and performing UI automation via accessibility trees.
    4
    14
    66 npm
    18
    MIT
  • A
    license
    B
    quality
    C
    maintenance
    A comprehensive MCP server for Android Debug Bridge, enabling AI agents to control Android devices through structured tools like launching apps, UI automation, file operations, and log capture.
    90
    5
    MIT
  • F
    license
    Not graded
    quality
    A
    maintenance
    MCP server that provides AI coding agents real Android development tools—Gradle, adb, logcat, lint, crash triage—through a local, permissioned interface. Enables agents to inspect projects, run safe Gradle tasks, capture logs/screenshots, and triage crashes.
    1
    -
  • A
    license
    A
    quality
    D
    maintenance
    An MCP server that gives coding agents hands and eyes on an Android device, providing 19 tools for building, testing, installing, UI inspection, input, logs, and toolchain diagnosis. It returns concise, structured results to make the build-debug loop efficient.
    19
    1
    MIT