DIYgod releases Pinky: an open-source bridge that lets AI agents control iPhone and Android

Agents have learned to code, browse websites and work with files. But as soon as a task hinges on the phone, everything breaks: messages in a messenger, an order in a delivery app, a payment in a banking client. Much of this does not exist on the web, and an ordinary browser agent cannot reach it. Chinese developer DIYgod has published a project called Pinky that tries to close this gap.

What it is

According to the author, Pinky lets any AI agent use your iPhone or Android. The tool itself runs on Mac, Windows and Linux, meaning the phone is connected to the computer and the agent works through it. The project is free and open source. It is a native app, not a heavy wrapper: the installation package is around ten megabytes.

The key phrase here is "any agent." Pinky is conceived not as yet another standalone assistant, but as a layer between the agent and the device. You choose the model or framework you like, and Pinky gives it hands to work with the phone. The author does not disclose the specific connection protocols or the list of supported agents in the post; you will need to check the repository.

Why the emphasis on control

Handing an agent your phone is scarier than handing it a browser: it holds money, correspondence and personal data. That is why half of the announcement is devoted to the brakes. Before a payment, Pinky stops and waits for your confirmation. The same confirmations can be enabled for sending messages and for deletion. This is the familiar human-in-the-loop scheme, applied only to genuinely dangerous actions rather than to everything indiscriminately.

The second mechanism is the ability to take over control at any moment. If the agent goes the wrong way, you simply take the phone back. The third is screen recording at every step. This is useful for debugging: you see not an abstract log but what the agent actually tapped. And if something goes wrong, you have proof of exactly where.

Limitations and open questions

So far this is an announcement post, not a review. We do not know how reliably the agent handles complex apps, how the project behaves with the protections of banking clients, or what limitations iOS imposes. It is also unclear how permissions are structured and where the video is stored, which is an important question for screen recording that captures personal data. The post's reach (around 72 thousand views) signals interest, but not product maturity.

A separate risk is prompt injection. If the agent reads other people's messages from the screen, they may contain text with instructions. Confirmation before payment softens the consequences, but does not eliminate the problem for actions you have not put behind a confirmation.

What to do about it

Agent developers should look at the repository and see how confirmations and recording are built: these techniques will be useful in other projects. Product managers may want to note the pattern of "pause on irreversible actions plus video audit" as a baseline standard of trust. It is best to try it on a test device or a spare account, and to enable confirmations for everything involving money, correspondence and deletion. If Pinky proves stable, open-source agents will gain the missing mobile layer without being tied to a single vendor.