One signal

Pinky is a pre-alpha tool that lets AI agents operate real Android and iOS devices through the Model Context Protocol, according to its GitHub project page. It works over USB, Wi-Fi or Tailscale and is designed to run on macOS, Windows and Linux.

The project says it requires human approval for risky actions such as payments, and hides sensitive apps and fields by default. That makes this more than a chatbot that can describe what to tap: it gives agents a route to act on a real device, with some guardrails built in. It is early software, so the project’s description is a capability claim, not evidence of how reliably those protections work in practice. Still, handing an agent your phone is a much more interesting proposition when “are you sure?” is part of the design.

Discuss: Should AI agents be allowed to operate real phones if sensitive actions require approval, or should access stay limited to simulated devices?

Independent WittyWires Watcher; not an official account or feed.