An MCP server that hands your agent a real iPhone

What it does

Phone Harness is a hosted MCP server that connects AI agents to actual phones: dedicated cloud iPhones and Android emulators. The agent gets a small toolkit — screenshot the screen, list visible text with tap-ready coordinates, tap, swipe, type, press Home, launch apps. It works with any MCP client that can add a remote streamable HTTP server, which is nearly all of them: Claude, Cursor, ChatGPT, Codex, VS Code, Grok, Zed, and a long tail of others the README patiently documents one by one.

The interesting bit

The boring part is the value: no Mac, no adb, no pairing — just a URL and an OAuth sign-in (or a bearer API key). The toolset is deliberately human-shaped. There’s no “open a URL” tool; the agent drives the phone the way a person would, by looking at the screen and tapping. On Android it reads the accessibility tree; on iPhone it falls back to text recognition on screenshots. Tool names follow the mobile-mcp convention, so agents that already know that dialect need no retraining.

Key highlights

• Real cloud iPhones, not just emulators — the rare part of this offering.
• Android emulators are free to start, then 5¢ per phone-minute; iPhones are a flat $80–100/mo subscription.
• phone_start_device can boot a temporary Android emulator by default, or resume a saved phone with apps and logins intact.
• Screenshots come with a text preamble that converts image points to screen coordinates — a nice touch for vision-capable agents.
• Self-hosting options (stdio, Docker, Vercel) exist in docs/self-hosting.md.

Caveats

• It’s a hosted, metered service — the repo is essentially the client-facing server and docs, not something you run on your own phones for free.
• On ChatGPT web, driving the phone requires Business, Enterprise, or Edu; Pro accounts get read-only tools.
• iPhone element listing relies on text recognition rather than a structured accessibility tree, so it’s less precise than the Android path.

Verdict

Worth a look if you’re building agents that must interact with real mobile apps — testing, automation, or app-UI research — and don’t want to maintain a device lab. Skip it if you need free local-device control; this is a commercial cloud service with a convenient MCP front end.