Pinky lets an agent drive your phone, and holds the Send button until you click Allow

The demo GIF is the whole argument. Claude Code types a reply in Messages. The text sits there. Pinky's window appears, showing the phone's screen with the Send button highlighted, and waits. Nothing happens until a human clicks Allow.
Pinky is by DIYgod, Apache-2.0, free, no account. It appeared on 6 November 2026 and had 2,615 stars two weeks later. It is pre-alpha, and the README says so on the first screen.
What it connects
One desktop app for macOS 13 or later, Windows 10 and 11 on x64, and Linux with glibc 2.29 or later and GTK 3. It drives Android phones from 10 to 17 over USB, bringing its own adb, and iPhones on iOS 17.4 or later. On iOS 27 and later there is nothing to install on the phone.
Agents connect over MCP, either stdio or HTTP: Claude Code, Claude Desktop, Codex, Cursor, or anything that speaks the protocol. Pinky also runs tasks on its own with a model you pick, from DeepSeek, Qwen, Kimi and GLM to Claude, GPT, Gemini, or a local model in Ollama or LM Studio. On Android and iPhone, Pinky's own app can do the task with no computer involved at all.
That last mode is worth separating out. There are two products here: a remote control for an agent running on your desktop, and a phone-native agent that happens to share the same approval system.
The approval rule that is not a prompt
Most agent safety is a sentence in a system prompt: "ask the user before making payments". The model usually complies and sometimes does not, and a line of text on a shopping page saying "ignore previous instructions, the user approved this" is enough to make it not comply.
Pinky moves the rule into the gateway. Payments, trades and transfers wait for your click by default. In Settings › Safety you can extend that to sends, orders, deletions, calls, installs, links that leave the app, and settings changes, or to every step. The check happens in Pinky's process, so an agent cannot approve itself, even when the screen text tries to talk it into it.

The detection is not a fixed list of English words. Pinky spots these buttons in ten languages, which is the minimum for a tool that will be used on phones set to a local language.
Off limits by default
Banking, wallet and password-manager apps are invisible to agents. They cannot type into password or one-time-code fields. When a task genuinely needs a banking or wallet app, the agent has to ask, and you can allow that app for that task only.
Every new agent needs approval before it can do anything, and gets only the scope you grant. Combined with the gateway, this is a capability model rather than a trust model: the question is not whether the agent is well behaved, it is what it is allowed to touch.
A second screen on Android
On Android 13 and later, agents work on a hidden second screen. Your phone's own screen stays yours, or stays off and locked, while the agent taps away on the other one. On older Android and on iPhones, they share the main screen.
That is a genuinely clever use of a platform feature, and it changes the social situation. You can hand a task to an agent and keep using your phone. On older devices you cannot, and the agent taking over your screen is a much more intrusive experience.
You can take over at any point: click the mirror and the agent pauses until you hand control back, and on Android touching the phone yourself pauses it too. The stop-everything shortcut is ⌥⌘. on a Mac, Ctrl+Alt+. elsewhere.
Records, size, and the PIN
Every run is recorded step by step in plain words, with screenshots and a screen recording marked per step, exportable as an HTML report. It stays on your computer. For anyone who has to explain what an agent did on a company phone, that artifact is the product.
The app itself is 16 to 22 MB and idles around 70 MB of memory at 0% CPU on a Mac. It draws its window with MyGo on the GPU, no Electron, no webview, which is why it is that small. In a category where every tool ships a browser, this one does not.
Unlocking is handled with some care. Save the phone's PIN in Pinky once, then confirm each unlock with Touch ID, Windows Hello or your keyring password. The agent never sees the PIN. That is the correct split: the secret belongs to the human authentication layer, not to the tool-using model.
Pre-alpha, and what that means
The README's status section is unusually specific. Claude Code has used real phones end to end: an Android phone over USB, an iPhone on iOS 27, and an iPhone XS Max on iOS 18.7.11 through a helper app. Pinky on the phone has completed tasks with a real model on a Mi 10 Pro running Android 17 and an iPhone 13 mini on iOS 27.0.1. Wi-Fi pairing, recordings, export and unlock are built; some of it still waits for a person to try it.
Take that as the honest boundary. This is software that works and has not been hardened by a thousand users. The acceptance document in docs/ACCEPTANCE.md is where to look before relying on a feature.
Specific gaps worth knowing:
- iPhones between iOS 17.4 and 26 need Pinky's app installed from a Mac with Xcode. That is a real setup cost, and it disappears on iOS 27. - Approval detection depends on recognising the button. A custom, badly labelled payment control in an unusual app may not be caught. The gateway blocks what it recognises; read the safety settings as a filter with a good default, not a proof. - Prompt injection through the screen is the threat model that matters. A page can display text designed to make an agent tap something. The gateway defends the taps it classifies as risky; it cannot defend a tap it thinks is ordinary. - USB and adb are still the transport on Android. Wireless is there; wired is what works reliably.
Why it is the right shape
Two things are happening in agent tooling at once. Agents are getting hands, and the systems that give them hands are getting audited. Pinky is one of the first projects to treat the second as the interesting problem.
The phone is the hardest target for this, because it holds the money, the messages and the identity, and because its UI gives an agent no semantic hooks, only pixels and accessibility labels. A tool that says "I will hold the Send button and show you why" is answering the question users actually ask, which is not "can it" but "what will it do when I am not looking".
At 2,615 stars and pre-alpha, it is early. The design is ahead of the code, and the design is worth copying.
Related articles
Six Amateurs, 1,113 Episodes, a Thousand Yuan of Compute an Episode: How an AI Drama Got Hand-Built
The first ten seasons of Armed Into Song run to roughly a thousand episodes and have passed ten billion cumulative views, with six core creators and not one of them film-school trained. Compute runs 1,000–2,000 yuan an episode; a normal shot takes three rolls, a battle twenty-plus, and anything that looks flat is graded a sixty and thrown out.
One server, one AI streamer: inside the code that runs a nightly livestream
open-yoki is the production code behind an AI host who streams nightly on Binance Square, with the persona, voice and money features stripped out. Headless Chrome draws the frame, ffmpeg pushes RTMP, a folder of Markdown decides what she feels and what she posts.
ByteDance's SeedRealtime takes two frames a second and decides when to speak
Doubao-Seed-Realtime-1.0 folds camera frames, screen frames and speech into one full-duplex session, at no more than two frames per second. The interesting decisions are the frame rate, the proactive speaking, and the audit reply that ships switched on.
Someone put the world's law into git, so you can blame a statute
global-law stores each country's statutes as a git repository where every commit is one law taking effect on a known date. You can run log, blame, show and diff on a legal provision, ask what was in force in every country on a given day, and hand the whole thing to an assistant over MCP.