SIAgentNews.com
← All news

Pinky lets an AI drive your phone but makes you approve payments

A new open-source tool lets AI agents operate a real phone, but payments, deletions and installs wait for a human click that the agent cannot fake.

· First reported by dev.to

16 to 22 MBsize of the Pinky desktop download

The short version

  • Pinky is an agent that operates a real phone, from the developer behind RSSHub.
  • Payments, trades, transfers, deletions, installs and calls wait for your click.
  • Banking, wallet and password-manager apps are hidden from agents by default.
  • Runs on Android 10 to 17 and iOS 17.4 and later; free under Apache-2.0.
  • It is pre-alpha, and we have not tested it.

Why it matters

Agents that act on your behalf raise a simple question: what happens when the AI gets it wrong? Pinky shows one answer, which is to put a hard checkpoint in front of actions you cannot undo. Whatever agent tool you use, that is a useful standard to hold it to.

An agent that works your phone

Pinky is a new project from DIYgod, the developer known for RSSHub. As dev.to reported, it is an agent that operates a real phone. You tell it what to do, and it taps and types through the apps for you.

The pitch starts from a worry many people share. AI agents can already book, buy and message on your behalf, but a phone holds some of the most sensitive data you own. One bad call by the model could mean a payment sent or something deleted, and neither is easy to take back.

Two ways to use it

The first way is the simple one. You put your own model API key (the password that lets software use an AI service) into the Pinky app on the phone, then give it tasks in a chat window.

The second way uses your computer. You install Pinky on a desktop, connect the phone, and let an outside agent drive it. Pinky names Claude Code, Claude Desktop, Codex and Cursor as examples. These connect through MCP, the Model Context Protocol, a common way to plug tools into AI apps.

The phone app can also work alone. It supports a range of models, from DeepSeek, Qwen, Kimi and GLM to Claude, GPT and Gemini. It also works with local models run through Ollama or LM Studio, which keep the AI on your own machine.

Safety enforced in code, not in a prompt

The most interesting part is how Pinky handles risk. According to the post, many phone agents borrow the usual approach from desktop agents. They guard certain files and web addresses, then rely on the instructions given to the model to make it ask a person before anything risky. That fails when the model is weak or its context drifts.

Pinky instead builds the human check into its gateway, a checkpoint that sits between the agent and the phone. Payments, trades, transfers, deletions, installs and calls all stop there until you click to approve. Because the gateway enforces this, the agent cannot approve its own request, even if text on the screen tries to talk it into doing so.

Other limits apply by default:

  • Banking, wallet and password-manager apps are invisible to agents.
  • Agents cannot type into password or one-time-code fields.
  • You can unlock the phone from your computer without the agent ever seeing your PIN.

A second screen and a paper trail

On Android 13 and later, the agent works on a hidden virtual screen, so your own screen stays free. You can keep chatting or scrolling while it works. A mirror on your computer lets you watch, and if the agent drifts off course, you click the mirror to pause it, fix things by hand, and give the task back.

Each run is also recorded. Pinky keeps the steps in plain words, along with screenshots and a screen recording, and you can export them as an HTML report. The post says all of this stays on your computer.

What you need, and what to be careful about

Desktop apps exist for macOS, Windows and Linux. Phones supported are Android 10 to 17 and iOS 17.4 and later. The post says the app is light, with a download of 16 to 22 MB, and that on a Mac it idles at about 70 MB of memory and 0% CPU. Scheduled tasks and opt-in memory are built in.

It is free, open source under the Apache-2.0 license, and needs no account. The post also says plainly that Pinky is pre-alpha, meaning very early software that is likely to be rough. We have not tested it, and the claims above come from the project's own description, so treat them as the builder's account.

Even so, the design idea carries over. The post argues that the next stage of the agent race may be less about what an agent can do and more about whether you trust it to do it. A hard stop at money, deletion and installation is a concrete way to earn that trust.

What to watch

  • Whether Pinky's approval gateway holds up once people outside the project test it.
  • Whether other phone agents move their safety rules out of prompts and into code.

Try it

Before you let any agent act for you, write down the actions you could not undo, such as payments, deletions and sent messages. Check whether the tool can pause on each one, and if it cannot, wait.

Sources and links

First reported by dev.to. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news