Give AI a Cloud Phone: Paste a URL and It Can Tap the Screen, but 2FA Is Still on You

Phone Harness has released MCP version 0.3.3. Any MCP-supporting agent can now control a real iPhone or Android in the cloud. MCP is Model Context Protocol, used to connect models or agents to external tools. The phones connected this time are not web-based simulators, but full phones running in the cloud.

Usage is lightweight. In tools like Claude, ChatGPT, Cursor, and Codex, add it as a custom connector, paste a URL, and log in. The agent can see the screen and can tap, type, and swipe. The project is at phone-harness/mcp, version 0.3.3.

There are two connection methods. iPhone connects via iPhone Mirroring, and Android uses adb. In replies, shawn says that once connected, the agent can access the phone immediately, and it can also find a more efficient way to control it. Someone asked whether every tap sends raw frames; he replied that phone-harness improves itself, raw frames can be returned if wanted, and an accessibility tree can also be exposed when needed.

These capabilities are very direct for mobile testing. shawn says one agent can control multiple real phones in parallel, sessions can persist, and it is suitable for mobile QA. Multiple agents collaborating to control the same device—he himself thinks it could get weird, and he asks what scenario would be most suitable.

The boundaries are also clear. Someone asked what happens when an app pops up 2FA; shawn replied that it goes back to the user. One netizen worried about retaining login state: people are cautious about handing a phone to someone else, but handing it to AI makes it easier to let go. Another netizen said phone control is a great MCP test because the agent has to act, observe, and recover, and the failure modes are more interesting than the demo. Another netizen added that the 20% of failures usually come from apps requiring SMS verification codes.

In the demo, a few taps go smoothly. In real runs, what often stalls the flow is verification codes, login state, and multiple agents fighting over a device. Phone Harness turns the phone into a tool that agents can call; its value lies in testing and workflow automation, and so do the headaches.