DIYgod open-sources Pinky, enabling AI Agent to operate iPhones and Android phones

AT A GLANCE

DIYgod open-sourced Pinky, saying it can let AI operate iPhone and Android phones via MCP, and has added human approval for sensitive operations.

Article

The Ultimate Mobile Agent Has Arrived? DIYgod Open-Sources Pinky, Handing Both iPhone and Android Over to AI

Beating AI Flash News: RSSHub author DIYgod has open-sourced Pinky, enabling AI Agents such as Codex and Claude Code to directly operate real iPhones and Android phones. The desktop side supports Mac, Windows, and Linux. Once connected via MCP, the AI can view the screen, open apps, tap, swipe, and input text.

Previously, we have introduced quite a few iPhone Use projects. Some rely on the XCTest/WDA testing interface, some simulate keyboard and mouse input via Bluetooth, and others require external hardware. Pinky integrates multiple technical approaches into a single desktop tool.

On the Android side, it uses ADB + scrcpy, combined with an optional accessibility assistant app to read UI elements. On Android 13 and above, it can also create a hidden virtual screen, allowing the AI to operate apps in the background while the user continues using the phone normally. This part is quite similar to the previously introduced Mika.

For iPhone, it selects different approaches based on the system version. iOS 27 and above directly calls Apple's CoreDevice developer protocol to read the interface and send touch commands, without needing to install a test program. The technical approach is similar to Poki. iOS 17.4 to 26 still relies on WDA and requires signing and installation via Xcode.

Pinky also specifically addresses the problem of AI misoperations. It identifies sensitive buttons such as "Pay" and "Transfer," pauses before actually tapping them, and waits for the user to manually approve. If there is no confirmation within 60 seconds, execution is rejected. After approval, it rechecks the interface to prevent changes in the amount or button. Sending messages and deleting content can also be set to ask first.

This approval mechanism is written into the control program and does not rely on the model following prompts. Banking apps, wallets, and password managers are hidden from the Agent by default, and password and verification code input fields are also prohibited from being auto-filled. Users can take over operations at any time, and every step is screen-recorded, with recordings saved on the user's own computer.

Project: https://github.com/DIYgod/pinky

Original link: https://m.theblockbeats.info/flash/376241

Key points

The report states that DIYgod, the author of RSSHub, has open-sourced Pinky.
The report claims that Pinky enables AI Agents such as Codex and Claude Code to operate real iPhones and Android phones.
The desktop version supports Mac, Windows, and Linux, and through MCP, it allows AI to view the screen, open apps, click, swipe, and type text.
The Android end uses ADB and scrcpy, and may optionally use an accessibility assistant application to read interface controls.
The report claims that Android 13 and above can create a hidden virtual screen, allowing AI to operate apps in the background.
The report states that iOS 27 and above use the CoreDevice developer protocol; iOS 17.4 to 26 rely on WDA and require signed installation via Xcode.
Pinky will pause and wait for user approval before clicking sensitive buttons such as "Pay" and "Transfer"; if not confirmed within 60 seconds, it will refuse to execute.
The report claims that banks, wallets, and password managers are hidden from the Agent by default, and that password and verification code input fields prohibit autofill.
The user can take over the operation at any time; every step is screen-recorded, and the recordings are saved on the user's own computer.
The report provides Pinky's GitHub project address: https://github.com/DIYgod/pinky.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

Pinky is a desktop tool suite: after connecting an AI Agent that supports MCP to a mobile phone, the AI can view and operate the phone interface just like a user. It adopts different connection schemes for Android and various iOS versions, and places high-risk actions such as payments and transfers into a manual confirmation step.

Why it matters to readers

If the capabilities described in the reports can be stably reproduced, Pinky could extend the operational scope of desktop AI Agent to real mobile applications; the approval and takeover mechanisms written into the control program also attempt to reduce the risk of the model mistakenly tapping sensitive buttons.

Beginners can think of it as an open-source tool that "lets AI remotely operate a phone," but the installation method varies depending on the phone's system and version, and sensitive operations still require user supervision.

Risks and unknowns

The evidence is a media news flash that does not include official project documentation or actual test results, and its functionality and compatibility have not been independently verified.
The availability of the hidden virtual screen, UI reading, and different iOS version solutions on specific devices is unknown.
It is unknown whether sensitive button identification, interface re-checking, and default hiding of applications can cover all risk scenarios.
The specific security protection measures for screen recording, local records, and interface data are not specified.

DIYgod open-sources Pinky, enabling AI Agent to operate iPhones and Android phones

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center