Coderefercoderefer
CoursesWebinarsBlogAbout
Coderefercoderefer

Discover. Learn. Automate. Grow.

Learn

  • Courses
  • Webinars
  • Blog
  • Search

Company

  • About
  • Contact
  • FAQ

Legal

  • Privacy Policy
  • Terms of Service
  • Refund Policy
© 2026 Coderefer. All rights reserved.
HomeBlogPinky: Let Claude Code Control Your iPhone or Android (Human-in-the-Loop MCP)
ai-toolsNovember 16, 20268 min read

Pinky: Let Claude Code Control Your iPhone or Android (Human-in-the-Loop MCP)

Pinky is a free open-source MCP gateway from DIYgod that lets Claude Code or Codex control your real iPhone or Android phone with a human approval gate for risky actions.

V

Vamsi Tallapudi

Manager, Architect Technology at Cognizant

Pinky: Let Claude Code Control Your iPhone or Android (Human-in-the-Loop MCP)

Pinky is a free, open-source desktop app from DIYgod (the developer behind RSSHub) that lets Claude Code, Claude Desktop, Codex, Cursor, or any other MCP-compatible agent control your actual iPhone or Android phone, while a separate approval layer blocks the agent from touching payments, messages, or deletions until you click "allow." It's not a simulator and it's not a jailbreak tool. It mirrors your real phone's screen on your Mac, Windows, or Linux desktop and lets an AI agent tap, type, and scroll through real apps, with a gateway process standing between the agent and anything risky.

DIYgod open-sourced Pinky on November 6, 2026. As of this week it's sitting at roughly 2,200 GitHub stars, 240-ish forks, and over 1,350 commits, with new releases rolling out every few hours (version 2026.1115.4 landed November 15). For a pre-alpha project, that's a fast climb, and it tells you there's real appetite for "let my agent use my phone" beyond just dev-tool nerds.

What does Pinky actually do?

Pinky runs as a native desktop app (no Electron, no webview) that connects to your phone over USB or Wi-Fi and exposes an MCP server any agent can talk to. The agent sees your phone's screen, issues taps and swipes, and gets a fresh screenshot back after every action. You watch the whole thing happen in Pinky's window on your computer.

The install footprint is tiny by 2026 standards: 16-22 MB download, about 70 MB of idle RAM, basically 0% CPU when nothing's happening. DIYgod built the UI with his own rendering layer (called MyGo) instead of bundling a browser engine, which is why it's this light. Compare that to most "agent desktop" tools shipping a full Chromium instance just to draw a window, and it's a nice change of pace.

Platform support as of this week:

  • Desktop: macOS 13+ (Apple Silicon and Intel), Windows 10/11 64-bit, Linux 64-bit (glibc 2.29+, GTK 3)
  • Mobile: Android 10 through 17, over USB or Wi-Fi; iOS 17.4+ over USB or Wi-Fi, plus simulators and emulators on both sides

On newer iPhones (iOS 27+) there's nothing to install on the phone at all: Pinky talks to it directly. On older iOS you need a companion app installed once via a Mac with Xcode. Android needs a small companion app too, installed over USB the first time. It's a different trade-off than wiring an ESP32 into an iPhone agent setup, which skips the desktop tether entirely but needs its own hardware.

How does the human-in-the-loop approval actually work?

This is the part that matters more than the phone-control trick itself. A lot of "AI controls your phone" demos are cool for about ten seconds until you remember prompt injection exists: a malicious text message or a sketchy web page could try to talk the agent into doing something you never asked for.

Pinky's answer is to keep the rulebook out of the agent's reach entirely. The approval logic lives in Pinky's own gateway process, completely separate from the agent and the model generating its actions. When the agent tries something Pinky flags as risky, the action comes back with an approval_pending status and just... stops. Nothing executes until you click approve in Pinky's window. Critically, the agent itself cannot grant that approval, even if the screen content it's reading tries to instruct it to. The policy sits one layer below the model, where prompt injection can't reach it.

By default, Pinky requires your approval for:

  • Payments and money transfers
  • Deleting things
  • Sending messages or making calls
  • Installing apps
  • Changing device settings
  • Opening links

You can tighten this in Settings so every single action needs a click, or loosen specific categories if you trust a given task. Either way, the defaults assume you don't fully trust the agent, which, if you've used any coding agent for more than a week, is the correct assumption.

On top of the approval gate, Pinky hides banking, wallet, and password manager apps from the agent by default — it literally can't see them unless you grant access for a specific task. It also blocks the agent from typing into password or one-time-code fields. On Android 13 and newer, agents can work on a hidden background display while your actual phone screen stays dark and locked, so you're not watching your phone light up and get poked at every few seconds.

Every task gets recorded with a screen recording and step markers, which is genuinely useful when something goes wrong and you want to see exactly where the agent went off the rails.

What agents can you connect to Pinky?

Pinky auto-configures for:

  • Claude Code
  • Claude Desktop
  • Codex
  • Cursor
  • Any other MCP-compatible coding agent or custom script

There's also a built-in "Ask Pinky" mode that runs tasks using your own API key for a model of your choice: DeepSeek, Qwen, Kimi, GLM, Claude, GPT, Gemini, or even a local model through Ollama or LM Studio. Pinky itself ships with zero model baked in, which I actually like. It's infrastructure, not a product trying to lock you into one AI provider.

The agent-facing side is an MCP server speaking JSON-RPC over stdin, with around 26 tools, all prefixed phone_ — things like phone_connect, phone_tap, and phone_task_steer. Every tool call returns the updated screen state so the agent always knows what it's looking at next.

If you want to run something unattended for a bit, Pinky caps background tasks at 100 steps, 20 minutes, or $2 in model spend, whichever hits first. That's a sane guardrail. It stops a loop from quietly burning your API budget or wandering around your phone for an hour while you're not watching.

Is Pinky actually ready to use?

DIYgod labels the project pre-alpha, and the project's own acceptance docs are refreshingly honest about what's shaky. Background displays on Android 10 run apps you can't actually read until Android 11. iPhone Wi-Fi mirroring drops to 3-5 second screenshot intervals and loses video. File access only works on Android, not iPhone. Touch ID and Windows Hello prompts still need a real human finger (no surprise there, and frankly it'd be alarming if they didn't).

The docs are explicit that this is built "for one phone you own, with you in the loop" — not for unattended automation, not for device farms, not for anything hands-off. If you came here looking for a way to run a hundred phones scraping something overnight, this is the wrong tool on purpose.

How does Pinky compare to Poki?

Poki, built by JosssphZhou, solves a similar problem from a different angle: instead of running on your desktop and tethering to a phone, Poki's agent runs directly on the iPhone itself after a one-time setup, so tasks can execute without a connected computer. It's iPhone/iPad-only (iOS 18+), AGPL-3.0 licensed, and sitting around 1.4k stars with a 0.4.0 beta tag. It also ships an MCP server so a desktop agent can drive the phone remotely when you want that mode instead.

The practical difference: Pinky is cross-platform (iOS and Android) and desktop-first, with the gateway approval model as its headline feature. Poki is iPhone-native and works untethered once installed, with its own built-in approval gates for payments and Wallet actions, plus support for more model providers out of the box (ChatGPT, Claude, Gemini, DeepSeek, OpenRouter). If you're deep in the Apple ecosystem and want something that keeps working when your Mac is closed, Poki's approach makes sense. If you need Android support or you like having your desktop as the control room, Pinky is the one to grab.

Neither project is mature. Both are explicit about needing a human watching. Treat both as "cool, early, don't trust it with your banking app" rather than production tooling.

Should you try Pinky?

If you're already running Claude Code or Codex for development work and you've ever wished you could hand it "go check if the package arrived" or "open the app and change this setting," Pinky is worth ten minutes of your time. It's free, it's small, and the approval gateway is a legitimately smart piece of security design — not just a checkbox someone added to look responsible.

I wouldn't connect it to a phone with your primary banking app logged in and walk away, and the project doesn't want you to either. But for low-stakes phone chores (testing your own app's install flow, navigating settings, checking something in an app you don't want to pick up your phone for), it's a genuinely useful building block. Just know you're an early adopter on pre-alpha software, and read docs/ACCEPTANCE.md in the repo before you decide what you trust it with.

If you're building your own agent tooling and want to see how a prompt-injection-resistant approval gate should actually be architected, Pinky's gateway-separate-from-model design is worth studying regardless of whether you ever install the app. It's the same instinct behind why Claude Code's permission prompts exist in the first place — keep the thing that can say "yes" outside the reach of anything the model reads.

Watch

Pinky GitHub Overview: How Its Offline MCP Gateway Controls Mobile Actions
ai-tools mcp claude-code 2026

Comments

    Get AI tricks that save you hours every week

    New AI tools, automation workflows, and course drops — straight to your inbox. Join 2,400+ builders.

    Read Next

    Google Antigravity /teamwork vs Claude Code Subagents: Which Wins for Long Projects?
    ai-tools

    Google Antigravity /teamwork vs Claude Code Subagents: Which Wins for Long Projects?

    Google Antigravity's /teamwork needs a paid plan and can flood your folder with files. Claude Code subagents are cheaper but need manual orchestration. Here's which wins for long projects.

    November 16, 20268 min read
    Qwen4-Max Preview: 1M Context, No Benchmarks, Gone by Nov 25
    ai-tools

    Qwen4-Max Preview: 1M Context, No Benchmarks, Gone by Nov 25

    Qwen4-Max Preview packs a 1M-token context and agent-first coding claims, but Alibaba shipped it with no benchmarks and a hard November 25 shutoff. Here's what's real.

    November 16, 20266 min read
    Expo Exact Explained: One App for Web, macOS, iOS and Linux
    ai-tools

    Expo Exact Explained: One App for Web, macOS, iOS and Linux

    Expo's new Exact framework runs one Contract app natively on web, macOS, iOS and Linux in a ~20KB runtime, built for AI coding agents to write, run and test it.

    November 15, 20267 min read
    View all posts →