ADB MCP Server
Control Android devices from an LLM
An MCP server that lets Claude, Cursor and other MCP-compatible IDEs inspect, control and automate Android devices and emulators through natural language.
- TypeScript
- Node.js
- MCP
- ADB
- Android
The problem
Debugging an Android app means a lot of small commands. Install the build. Clear the logs. Take a screenshot. Check a setting. Read a database row. I know most of them, and I still look up the flags every time.
An AI assistant can read my code, but it could not see my phone. So I was the one copying logs and screenshots back and forth.
How it works
It is an MCP server that sits between an AI client and adb, the tool that talks to Android devices. The model asks for something in plain words, the server runs the right command, and the result goes back to the model.
The tools cover devices and Wi-Fi connections, installing and launching apps, moving files, logs and screenshots, taps and typing, and reading an app's database and preferences.
Setting it up is a few lines in the AI client's config.
What I learned
For a model, the tool description is the interface. A person reads the docs once. A model reads the tool list every time. Clear names and clear descriptions decided whether it picked the right tool.
Return what the model can use. A screenshot comes back as an image the model can look at, not a file path. The screen layout comes back as a tree, so it can find a button and tap it.
Make the safe way the only way. Commands are passed as a list of arguments, never as one string for a shell to parse. So text from the model cannot turn into a second command.
Keep an escape hatch. I could not wrap every adb feature. A raw command tool covers the rest, and it uses the same safe argument list.