Developer DIYgod has made the Pinky project available, a desktop application that allows artificial intelligence agents engaged in coding and general tasks to control real iPhone or Android devices via the Model Context Protocol (MCP). The user monitors the process and has the ability to approve or block potentially risky actions.
The repository on GitHub, released under the Apache-2.0 license, was created on November 6th. According to a Chinese AI publication reported in Tencent News, the post about this project garnered around 220,000 views on X, and the project received over 1000 stars on GitHub within the first 25 hours. By November 11th, GitHub had noted approximately 1700 stars, and the publication reported the release of 30 updates in a single day.
The Pinky application functions on macOS, Windows, and Linux operating systems. The README file states that it supports Android versions from 10 to 17 and iPhones on iOS 17.4 or later versions, requiring the installation of a helper component via Xcode for iOS 17.4–26. MCP-compatible agents include Claude Code, Claude Desktop, Codex, and Cursor, as well as any other tools or scripts supporting MCP. Pinky itself does not contain its own model and acts as a bridge between the agent and the phone.
Android devices can connect via Wi-Fi after initial cable connection. Starting with Android 13, Pinky uses a hidden second screen, allowing the agent to work while the phone's display remains locked. In the case of iOS, the agent gains access to the main phone screen. An additional Tailscale option provides access to Android phones outside the local network. Developers can also allow the agent to install custom APK files, read logs and crash reports, and then retest.
The main advantage is the thoughtful security system. The approval process is controlled by the Pinky gateway, not by instructions given to the agent. By default, payments, transfers, and purchases require manual confirmation by the user; users can add messages, make calls, delete data, and install applications. Banks, wallets, and password managers are hidden from the agent by default, and fields for entering passwords and one-time codes cannot be filled. Every action is recorded with a description, screenshot, and recording, and the session can be exported as an HTML report. The user can take control at any time, which will pause the agent's operation.
During testing conducted by the publication, the agent was able to publish a short video on a video application in approximately four minutes. The README states that screenshots and history are stored on the computer for 30 days, while anonymous usage statistics and crash reports are transmitted. A warning was also issued that some applications detect automation, and users should adhere to platform rules.



