iphone-use: A Self-Hosted Way for AI Agents to Drive a Real iPhone

What it is for

iphone-use is aimed at developers and automation users who want an AI agent to operate a real iPhone. Its README describes support for Claude Code, Codex, and other MCP clients. An agent can read screen text and elements, tap, type, and swipe, including in apps that do not expose an API. The project is not limited to a simulator or a web browser: a daemon on the Mac works with an XCTest runner on the iPhone.

That makes the project relevant to several audiences. Mobile developers can use it for real-device login and regression flows. Individual users can automate repeatable phone tasks. Teams can combine agent-driven checks with manual control of the same test device. The project also exposes an HTTP API, while a browser page and a native iOS app provide human-facing control paths.

Documented capabilities

The README lists app launching, screen reading, tapping, typing, and swiping. Long text can be entered in one step, and a completed task can be recorded as a replayable flow. Later, that flow can run without calling a model for every action. The project also describes action status reporting: an operation may be reported as applied, not sent, or unknown, with information about whether retrying is considered safe. It says covered elements are refused rather than reported as successful.

Initial setup uses USB, after which the phone can be driven over Wi‑Fi through an encrypted tunnel, according to the README. Human control is available from a browser or another iPhone. The documented interface includes a live H.264 view, mouse and keyboard input, QR-code pairing, and an element wireframe when capture-restricted screens cannot show normal video. Agents and people use the same daemon, with a single-owner mechanism intended to prevent simultaneous control.

The official v0.19.2 release was published on October 17, 2026. Its notes mention pinned CI dependencies, additional centralized scanning requirements, and an authentication compatibility fix. Repository metadata shows the project is not archived, but it does not confirm a recognizable SPDX license.

Limits and deployment considerations

Environment requirements should be checked before adoption. The README calls for macOS 15 or later, Xcode signed in to an Apple ID, an iPhone running iOS 15 or later, Developer Mode enabled, and a USB cable for initial setup. The user must also unlock and trust the phone and handle signing-related steps. Those tasks cannot simply be delegated to an agent. The supplied materials do not confirm a broader hardware matrix, a performance guarantee, or universal app compatibility, so the examples should not be treated as proof that every app will automate reliably.

A real phone may expose messages, health information, banking pages, payment screens, and account sessions. The README mentions authenticated HTTPS reverse proxies or VPNs for remote access, but that is not a substitute for an organization’s access controls, network isolation, account separation, and audit plan. Avoid placing production credentials, payment actions, or private account data into an unverified agent flow.

The license also requires human review before integration. The repository metadata reports NOASSERTION for SPDX, so unrestricted commercial use or redistribution should not be assumed. A sensible evaluation path is to use a dedicated test phone, validate the target apps and signing process, check the network path, and measure replay stability before adding the project to CI or daily automation.

Sources

• GitHub repository metadata
• Project README
• Official release
• Official project page

Sources checked: 2026-10-19