droidlume
Enables AI agents to control local Android virtual devices managed by DroidLume, including device lifecycle, app management, UI input, screenshots, and diagnostics through structured MCP tools.
README
DroidLume Agent Control
npm-distributed CLI and MCP server plus a standard AI Agent Skill for DroidLume.
DroidLume Agent Control lets Codex, Claude Code, Cursor, Gemini, and other AI agents control local Android virtual devices managed by DroidLume. It combines a structured CLI, a local MCP server, macOS App Intents, and an open Agent Skill without relying on fragile screen-coordinate automation.
What it can do
- Create, clone, configure, start, stop, restart, repair, show, and delete Android devices
- Install APK files through a streamed sandbox upload
- List, launch, stop, and uninstall Android apps
- Type text, tap coordinates, swipe, and send Android key events
- Capture Android screenshots as PNG or native MCP image content
- Export redacted DroidLume diagnostics
- Return stable device IDs, typed JSON, command schemas, timeouts, and structured errors
Requirements
- Apple silicon Mac
- macOS 14 or later
- DroidLume 1.1 or later
- Node.js 20 or later
Install
Install the CLI and MCP server from npm, then install the Skill for your agent:
npm install --global droidlume-agent-control
droidlume agent install codex
Or run the maintained installer, which performs the same two steps:
curl -fsSL https://raw.githubusercontent.com/tageecc/droidlume-agent-control/main/install.sh | zsh -s -- codex
Targets:
all | codex | claude | cursor | agents
Verify the installation:
droidlume version
droidlume status --format pretty
droidlume device list --format pretty
MCP setup
Generate an MCP configuration snippet:
droidlume agent config
Recommended configuration without a manually downloaded ZIP:
{
"mcpServers": {
"droidlume": {
"command": "npx",
"args": ["-y", "droidlume-agent-control", "mcp"]
}
}
}
The MCP server uses stdio and exposes 23 typed tools, three resources, and an Android app inspection prompt. Screenshots are returned as native image/png MCP content.
CLI examples
Start a device and open its window:
droidlume device start DEVICE_ID --show --timeout 180
Install and launch an APK:
droidlume app install DEVICE_ID /absolute/path/application.apk --timeout 300
droidlume app launch DEVICE_ID com.example.app
Inspect the Android screen:
droidlume device screenshot DEVICE_ID --output screen.png
droidlume input tap DEVICE_ID 540 1200
droidlume device screenshot DEVICE_ID --output screen-after.png
Read the machine contract before automation:
droidlume schema
droidlume schema device start
See the complete command catalog, MCP guide, state and safety rules, and error reference.
Architecture
Codex / Claude / Gemini / Cursor / CI
│
Agent Skill / npm MCP / npm CLI
│
DroidLume Control API v1.0
http://127.0.0.1:55777
│
DroidLume RuntimeController
│
Managed Android devices
The public CLI and MCP server communicate with the DroidLume host app over loopback. Device lifecycle, app installation, screenshots, input, and diagnostics still run through the same RuntimeController used by the native macOS interface.
Why DroidLume Agent Control
Generic Android automation often exposes raw ADB serials or depends on screenshot coordinates. DroidLume provides a product-level control contract instead:
- stable device UUIDs instead of transient emulator serials;
- lifecycle-aware start, stop, restart, repair, and Quick Boot handling;
- streamed APK upload into the DroidLume sandbox;
- structured command discovery and errors;
- one control plane shared by the native app, CLI, MCP, and Shortcuts.
FAQ
Is this an Android emulator?
DroidLume is the Android virtual device application. This repository provides the npm CLI/MCP implementation and Agent Skill used to automate it; it is not a second macOS application.
Does it use raw ADB as the public interface?
No. AI clients call the versioned DroidLume control API. The host app owns its private ADB namespace and managed-device lifecycle.
Which AI tools are supported?
The Agent Skill follows the open SKILL.md format. The CLI works from any shell, and the MCP server works with MCP-capable clients including Codex, Claude Code, Cursor, and Gemini tooling.
Is the tool local?
Yes. The DroidLume host control endpoint listens on 127.0.0.1, and droidlume-mcp uses local stdio transport.
Links
License
The Agent Skill and documentation in this repository are available under the MIT License. DroidLume application binaries and trademarks remain subject to their respective product terms.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。