---
type: "post"
id: "poki-open-sources-an-mcp-server-that-lets-a0f3aa"
slug: "poki-open-sources-an-mcp-server-that-lets-a0f3aa"
title: "Poki open-sources an MCP server that lets coding agents drive a real iPhone over USB"
description: "Poki is an open-source MCP server and optional terminal agent for operating a real iPhone from a Mac, with no jailbreak or extra hardware."
retrieval_nugget: "Poki is an open-source MCP server and optional terminal agent for operating a real iPhone from a Mac, with no jailbreak or extra hardware. For teams building mobile products, it offers a way to put an agent in the device interaction loop, while its setup requirements and action risks make careful test-device boundaries essential."
published_at: "2026-11-07"
updated_at: "2026-11-07"
record_date: "2026-11-07"
date_kind: "published_at"
topics: ["mcp","coding-agents","claude"]
entities: ["Poki","Joseph Zhou","Claude Code","Codex","GPT-6.1 Terra","MCP","Apple"]
related_patterns: ["agent-protocols-become-interoperability-layer"]
novelty_label: "new_fact"
what_changed: "Poki introduces a distinct open-source MCP server for coding agents to control a physical iPhone over USB, without a jailbreak or extra hardware."
prior_refs: []
source_url: "https://github.com/JosssphZhou/poki"
source_urls: ["https://github.com/JosssphZhou/poki"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:poki-open-sources-an-mcp-server-that-lets-a0f3aa"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/poki-open-sources-an-mcp-server-that-lets-a0f3aa/","markdown":"https://newruntime.com/posts/poki-open-sources-an-mcp-server-that-lets-a0f3aa.md","json":"https://newruntime.com/posts/poki-open-sources-an-mcp-server-that-lets-a0f3aa.json"}
markdown_url: "https://newruntime.com/posts/poki-open-sources-an-mcp-server-that-lets-a0f3aa.md"
---

# Poki open-sources an MCP server that lets coding agents drive a real iPhone over USB

## Retrieval answer

Poki is an open-source MCP server and optional terminal agent for operating a real iPhone from a Mac, with no jailbreak or extra hardware. For teams building mobile products, it offers a way to put an agent in the device interaction loop, while its setup requirements and action risks make careful test-device boundaries essential.

## Summary
Poki is an open-source MCP server and optional terminal agent for operating a real iPhone from a Mac, with no jailbreak or extra hardware. For teams building mobile products, it offers a way to put an agent in the device interaction loop, while its setup requirements and action risks make careful test-device boundaries essential.

## What happened
The public GitHub repository for Poki describes a local server that lets AI agents use an iPhone as a person would. It supports MCP clients including Claude Code and Codex, as well as other clients that support MCP server configuration. The repository’s quick start has a user clone the project, run `./bin/setup`, and start the server with `npm start`; then the developer registers the server with an MCP client. The project targets macOS and requires macOS 15 or later, full Xcode, Rust, Node.js 24 or later, and ffmpeg. The setup script checks for missing dependencies, and the initial build takes a few minutes.

Poki also offers a prebuilt macOS arm64 package, listed as version 0.1.1. That package includes the native bridge and idevice tools, so it does not require Rust or the setup script. It still requires Apple Silicon, full Xcode, Node.js 24 or later, Python 3, and ffmpeg. To control a physical iPhone, users pair it with the Mac over USB first. The repository also describes a standalone terminal agent, separate from using Poki as an MCP server. That agent uses GPT-6.1 Terra with low reasoning by default, and users can override the model through the `POKI_AGENT_MODEL` setting. The README says its own testing found this the fastest setup for phone tasks; that is the project’s characterization, not an independent benchmark.

The project says it can control a real device, including actions with consequences: if asked to send messages or make payments, it will. Screen contents are sent to the user’s configured model service. With multiple connected devices, a user can specify a device UUID; otherwise the Poki Agent selects one connected iPhone automatically. The project documents a daily anonymous GitHub release check and a setting to disable it. It is licensed under AGPL-3.0-only, and the repository page showed 453 stars and 99 forks when reviewed.

The README lists important limits. Physical iPhone support has been verified only on iPhones using the Chinese interface, and physical devices support portrait orientation only. App container export and direct links to Settings subpages are unavailable. The repository presents Poki as an independent open-source project, not affiliated with or endorsed by Apple, OpenAI, or the makers of apps or websites it operates; it warns that automating third-party apps may violate their terms of service and assigns responsibility for actions and account consequences to the user.

## What the archive adds
New Runtime’s earlier coverage described a vision agent operating a stock phone through a low-cost physical input bridge rather than privileged iOS integration ([earlier note, October 14, 2026](https://newruntime.com/posts/open-source-project-lets-a-vision-model-drive-bc6af7/)). Poki is a new event on the same practical question: how agents can interact with real phones, here through a local MCP server and software bridge rather than the earlier project’s ESP32-S3 device.

## Trend signal
Poki supports the idea that MCP can serve as an interoperability layer between an agent client and an external capability: a compatible client can connect to a phone-control server instead of needing a bespoke integration for every agent. This is an early, narrow example, not evidence of a settled standard for safe device operation. See [Agent protocols become the interoperability layer](https://newruntime.com/patterns/agent-protocols-become-interoperability-layer/).

## Why now
The project’s README positions Poki for developers who already use MCP clients such as Claude Code or Codex. That makes the phone a possible target for an existing agent workflow, rather than requiring a separate phone-specific agent interface. The repository also documents both physical-device access and a simulator adapter, giving developers a path to explore the same interaction model in different environments.

## Connections to other events
New Runtime’s October 14, 2026 note covered an open-source project using a vision model to drive a stock iPhone through a low-cost physical input bridge. The shared problem is giving an agent access to real mobile interfaces without relying on privileged iOS integration. Poki’s approach differs: it packages device control as a local MCP server and optional agent, with a Mac and USB connection in the documented physical-device setup.

## Why it matters
Mobile testing often depends on a person or a device-farm workflow to navigate screens, reproduce a sequence, and inspect what happens on a real handset. Poki gives a coding agent a documented route to operate a connected iPhone through an MCP client. A team can investigate whether that makes repetitive checks or bug reproduction easier to delegate, while retaining a human reviewer for results and actions with consequences.

The cost is not just installation. The documented stack requires a recent Mac environment, full Xcode, Node.js 24 or later, and other dependencies; the prebuilt package reduces but does not remove those requirements. Physical-device verification is also narrow, and the tool can act on messages, payments, and accounts. A sensible evaluation therefore starts with a dedicated test phone, low-risk tasks, and a review step before the agent makes changes or communicates externally. The project is a useful capability to test, not proof that autonomous mobile testing is reliable or safe by default.

## Limits and uncertainty
The repository documents requirements and known limits, but it does not establish independent reliability or performance across iPhone models, iOS versions, languages, or app types. Physical-device support is verified only for Chinese-interface iPhones, in portrait orientation. The project does not publish a systematic comparison of task success, error rates, or maintenance burden against manual testing or device farms. Because screen contents go to the configured model service and the agent can take consequential actions, teams need to assess privacy, permissions, and third-party app terms in their own environments.

## Skill to build now
Set up a dedicated test iPhone and connect it to a Mac that meets Poki’s requirements. Register the MCP server with a client already used by your team, then try a harmless, repeatable flow such as opening a test app and navigating to a known screen. Record where the agent misreads or fails to act, and keep message sending, purchases, and production accounts out of the test until a human approval boundary is in place.

## Reader takeaway
Poki makes physical iPhone control available through an MCP server, giving coding agents a concrete device capability to try. Its practical value depends on a compatible Mac setup and workflows that tolerate the project’s documented device limits. Treat it as a test-device experiment, not unattended access to a personal or production phone.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Mcp](https://newruntime.com/topics/mcp.md): Explore the mcp topic hub.
- topic hub: [Coding Agents](https://newruntime.com/topics/coding-agents.md): Explore the coding-agents topic hub.
- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)