Pinky Lets AI Agents Tap a Real iPhone, Behind a Human Gate
Pinky is an Apache 2.0 MCP bridge letting agents drive a real iPhone or Android phone, with payments and transfers held behind a human click.

Pinky, an Apache 2.0 desktop bridge, lets an MCP agent operate a real iPhone or Android phone while a human approves the risky taps.
DIYgod open-sourced it on November 6, 2026. Version 2026.1115.4 landed on November 15 at 02:38 UTC, with 2,158 stars, 235 forks, 1,352 commits and 151 releases behind it — new builds roll every few hours.
What you need to run it
The download is about 35 MB and needs no account, no API key and no GPU. It runs on Linux, macOS or Windows. The phone on the other end has to be an Android 10 to 17 device or an iPhone on iOS 17.4 or later.
Pinky ships no model of its own. Its phone app runs tasks with whatever model you pick — DeepSeek, Qwen, Kimi and GLM, or Claude, GPT and Gemini, or local models in Ollama or LM Studio. That keeps the cost and latency decision on your side of the tool.
The agent surface: 26 tools, all prefixed phone_
The agent-facing side is an MCP server over standard input speaking JSON-RPC 2.0. It exposes 26 tools, every one prefixed phone_, from phone_connect and phone_tap through phone_task_steer. Every action returns the new screen, so the agent sees the result of its own tap.
The documentation names Claude Code, Claude Desktop, Codex, Cursor, VS Code, Copilot CLI, Windsurf and Gemini CLI as supported agents, connected over standard input or HTTP. If your team already runs one of those, the harness does not change.
Approvals live in the gateway, not the prompt
This is the design decision worth copying. Approvals sit in Pinky's gateway rather than in the prompt, so an agent cannot approve itself even if text on the screen talks it into trying.
By default, payments, trades and transfers wait for a click. Settings can extend that gate to messages, orders, deletions, calls, installs, links leaving the app, and setting changes. Pinky says it spots these buttons in 10 languages.
Banking, wallet and password-manager apps are invisible to agents, and an agent cannot type into a password or one-time-code field. If it needs one, it has to ask, and the user can allow it for that task only.
A Tokenstrom code review found that only the desktop interface can grant an approval, that requests are denied after 60 seconds, and that each approval is tied to the session, device, tool and arguments and used once. A single click therefore cannot be replayed against a different command.
What the human sees while it runs
On Android 13 and later, agents work on a hidden second screen, so the phone stays usable. Older Androids and iPhones use the main screen instead. Clicking the mirror or touching the phone on Android pauses the agent.
Every run is recorded in plain words, with screenshots and a screen recording marked per step, exportable as an HTML report. The data stays on the user's computer.
That record is what makes the difference between a demo and something you would point at a real account. Without a per-step trail, a bad run leaves nothing to reconstruct.
What it is not
The project calls its own status pre-alpha and points to docs/ACCEPTANCE.md for what works today. Read that as scoped to one phone you own, with you in the loop — not unattended infrastructure and not a device farm.
Two caveats come from the reporting rather than the project. The approval model is a claim from the authors, not an independently tested result. And button detection is the weak link: a gate only helps if it recognises the risky tap in the first place.
The release cadence cuts the same way. A nine-day-old project shipping several builds a day is not one to track by "latest" — pin the version you tested and read the notes before moving.
The pattern other MCP servers are converging on
The lesson the project itself draws is to put the safety control where the agent cannot reach it. A second MCP server published the same day applies that to spend instead of taps: each of its tools takes public file URLs plus a max_charge_usd limit per call, defaulting to $1, so an agent in a loop cannot spend more than you allow. Its author discloses building both the server and the services behind it.
Different risk axis, same structure. The limit is enforced outside the model's context, which is the only place a limit survives contact with text the model is reading.
What to watch
Three things decide whether this graduates from hobby tooling. Whether anyone independently tests the button detection, since the gate rests on it. Whether docs/ACCEPTANCE.md stabilises enough to tell you what a given release actually handles. And whether the hidden-second-screen path stays an Android 13-and-later privilege, because on an iPhone the agent is still competing for the screen you are watching.
Until then, treat the approval gate as damage limitation rather than a guarantee, and keep the phone one you can afford to have tapped wrong.
More from DangMua