Image: GitHub
Why it mattersAn AI agent can now use a real phone through MCP. The phone connects out, so there is no open port and no USB cable.
Automating an Android phone usually means a USB cable, developer options and ADB. Droidline uses none of that.
Droidline is an open-source tool that drives a real Android phone from code on a PC. The project was created on 7 October by a developer who posts as KnifeLemon. It had 431 stars and 46 forks on 31 October, with the current release, 0.1.12, published a day earlier. The code is MIT.
How it reaches the phone
The user installs one app on the phone. The user then turns on its accessibility service, keyboard and battery permissions. Pairing is a QR code or a six-digit code. After pairing, the phone opens the connection outward to the PC on port 8780.
The project says the handshake uses a P-256 key exchange. Every line after it is encrypted with AES-256-GCM. Because the phone connects out, the setup works on the same Wi-Fi. It also works through a cloudflared or ngrok tunnel. It works through a relay the user runs on Cloudflare Workers or a VPS. There is no open port on the phone and no USB cable.
What a script can do
The server exposes the same commands to Python, Node.js, C#, the shell, HTTP and MCP. The project's docs say touch is touch in every one of them. Elements are selected by text, id or description. Pixel coordinates are not needed. The touch, input and wait commands wait up to ten seconds for the element by default.
Droidline's own README quotes screen tree reads at 53ms. The same read through adb and uiautomator takes 2,342ms. Finding an element by text is 16ms, against 2,368ms through adb. These are the project's figures and were not checked independently here.
The agent path
The MCP server is the part a reader building with AI agents will notice. The docs say Claude, Cursor and other agent clients can run Droidline as an MCP server. One line of config is enough. The .mcpb package also installs in Claude Desktop. The server is listed in the MCP Registry.
An agent given the MCP server can take a screenshot of a real phone. It can read the element tree, tap, type and wait. The project's docs say the agent's tool set can be reduced with --tools or --read-only. That limits what an unattended agent is allowed to do.
What it does not do
The README lists the limits in plain words. Droidline uses only permissions an ordinary app can get. So it cannot open another app's unexported screens. It cannot unlock a PIN lock. It cannot force stop an app or clear its data without pressing buttons in Settings. Games and some custom-drawn apps show no elements and need coordinate taps.
The repository runs CI on every commit. There are 14 incremental releases in 24 days. The docs are in English, Korean and Simplified Chinese.
A team that tests mobile apps on real phones has a new way to do it. So does a developer giving an AI agent hands on a phone. Neither needs a cable or root.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.