Icemoon — control a real iPhone with AI

app.icemoonv1.0.1更新於 Oct 2, 2026

Control real, physical iPhones with AI: screen reading, human-like touch, no jailbreak.

概覽

AI 產生的概覽

讓助理讀取以 USB 連接的 iPhone 螢幕,並用點擊、滑動與輸入執行多步驟任務。

功能
Icemoon 透過內建於 Icemoon macOS 應用程式的本機伺服器,把實體 iPhone 暴露給 MCP 用戶端。助理以結構化 UI 元素而非像素的方式讀取螢幕,再以合成的觸控手勢操作,透過真實 iOS 鍵盤輸入,並可啟動應用程式與導覽。約 53 個工具涵蓋裝置與工作階段管理、螢幕理解、觸控與輸入、應用程式與導覽、批次與計畫、裝置狀態、排程,以及 HTTP 請求與 TOTP 驗證碼產生等公用功能。
適用情境
當助理需要在真實硬體上操作真實 iOS 應用程式時使用:走查新手導引流程、重現或驗證介面步驟、為某個流程截圖或錄影,或依排程在多台已連接的 iPhone 上執行已儲存的自動化。
執行需求
Apple 晶片上的 macOS 13 或更新版本,已安裝並啟用的商業 Icemoon 應用程式(Farm 或 Scale 方案),以及一或多台以 USB 連接、執行原版 iOS 的 iPhone。npm 套件 icemoon-mcp 只是應用程式內伺服器的啟動器。選用環境變數:ICEMOON_APP 用於非預設的應用程式路徑,ICEMOON_DASHBOARD_URL 用於本機儀表板位址。
安裝前請注意
此伺服器為閉源,隨付費應用程式提供,沒有授權就無法使用。它控制實體裝置:工具可以點擊、輸入、解鎖、變更裝置狀態、設定模擬位置、讀寫剪貼簿、錄影、安排排程任務並發出外部 HTTP 請求。螢幕內容會傳送給所連接的 MCP 用戶端,工具呼叫會寫入本機稽核記錄。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 Icemoon — control a real iPhone with AI,將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Desktop only,透過 STDIO。 STDIO 服務會啟動本機處理程序,因此需要 SourceWeft 桌面主機。

其他 MCP 客戶端

參照 儲存庫 中的啟動說明。

README

Icemoon MCP Server — control real iPhones from Claude

Give your AI hands. Icemoon exposes physical iPhones to Claude (and any MCP-compatible client) through a local Model Context Protocol server. The AI reads the device screen, plans multi-step tasks, and executes them with human-like touch — on stock iOS, no jailbreak.

This repository is the official documentation and integration reference for the MCP server that ships inside the Icemoon app. The server itself is bundled with the app (commercial, closed source); everything you need to connect it to your MCP client is documented here.

Website: icemoon.app · MCP overview: icemoon.app/mcp · Install guide: icemoon.app/install


What it does

┌────────────┐     MCP (stdio)     ┌──────────────────┐     USB      ┌────────────┐│   Claude   │ ◄─────────────────► │  Icemoon MCP     │ ◄──────────► │  iPhone(s) ││  (or any   │                     │  server + app    │              │  stock iOS ││ MCP client)│                     │  on your Mac     │              │            │└────────────┘                     └──────────────────┘              └────────────┘
  • Sees the screen as structure, not pixels — real UI elements (buttons, labels, text fields) with identifiers and coordinates.
  • Touches like a hand, not a script — every swipe follows a unique Bezier curve, taps carry natural timing jitter, typing goes through the real iOS keyboard.
  • Plans in the open — for multi-step tasks the AI creates a plan you watch progress live on the Icemoon dashboard.
  • Drives a fleet — connections are pooled; the AI switches between connected iPhones instantly or targets a specific device per call.
  • Runs locally — the MCP server and the devices live on your Mac. Nothing goes to Icemoon's servers.

Requirements

  • macOS 13+ on Apple Silicon (M1 or newer)
  • Icemoon installed and activated — the MCP server is included in the Farm and Scale plans
  • One or more iPhones connected over USB (stock iOS, no jailbreak — setup guide)

Quick start (Claude Desktop)

  1. Install Icemoon and connect an iPhone — step-by-step guide.
  2. In the Icemoon app, open AI Copilot → Install to Claude Desktop. The app registers the server (named automation-studio) in ~/Library/Application Support/Claude/claude_desktop_config.json automatically, with its bundled Python runtime — nothing to install by hand.
  3. Restart Claude Desktop. The device tools are now available in every chat.

Ask something like:

Open Settings on my iPhone, turn on Do Not Disturb, and verify it stuck.

Via npm (any MCP client)

icemoon-mcp is a tiny launcher: it finds the installed Icemoon app and hands stdio to the server bundled inside it. No Python to install, no paths to look up.

json
{  "mcpServers": {    "icemoon": {      "command": "npx",      "args": ["-y", "icemoon-mcp"]    }  }}

The launcher itself is MIT-licensed and has zero dependencies — read it, it is about a hundred lines. It does nothing on its own: the Icemoon app must be installed and licensed, otherwise it prints where to get it and exits.

If an AI agent is doing the setup for you, point it at llms-install.md — the prerequisites to check, the config to write, and what each failure message means.

VariableDefaultPurpose
ICEMOON_APP/Applications/AutomationStudio.appPath to the app bundle, if installed elsewhere
ICEMOON_DASHBOARD_URLhttp://127.0.0.1:8088Where the running Icemoon app listens

Manual configuration (Claude Code, Cursor, other MCP clients)

If you would rather not go through npm, any MCP client can run the bundled server directly — point it at the bundle's runtime and script:

json
{  "mcpServers": {    "automation-studio": {      "command": "/Applications/AutomationStudio.app/Contents/Resources/mcp_python/bin/python3",      "args": ["/Applications/AutomationStudio.app/Contents/Resources/mcp_server.py"],      "env": {        "DASHBOARD_URL": "http://127.0.0.1:8088",        "PYTHONPATH": "/Applications/AutomationStudio.app/Contents/Resources/mcp_libs"      }    }  }}

Notes:

  • DASHBOARD_URL is the local Icemoon dashboard (default port 8088); the app must be running.
  • Exact paths can differ per build — the easiest way to get the correct entry is to run the one-click install once and copy the generated block from claude_desktop_config.json.

Tools

53 tools across eight groups — full reference with parameters in docs/tools.md. The same list is published as a machine-readable server card.

GroupToolsWhat the AI gets
Devices & sessionlist_devices connect_device disconnect_device device_status get_metricsDiscover running devices, pooled connections, per-call port= targeting
Screen understandingget_visible_elements identify_screen get_screen_info get_screen_map get_element_at_position find_by_id find_by_label get_nav_map screenshot_saveStructured UI reading — elements, screen fingerprints, app navigation maps
Touch & inputtap tap_by_id tap_by_label tap_relative double_tap touch_hold swipe swipe_relative scroll pinch type_text smart_type keyboard_dismiss press_buttonHuman-like gesture synthesis; typing through the real iOS keyboard
Apps & navigationopen_app kill_app get_app_list navigate_to wait_for_screennavigate_to walks to a target screen in one call: detect → path → execute → verify
Batching & plansrun_batch sleep_delay foreach_elements plan_create plan_add_step plan_update_step plan_getWhole action sequences in one round-trip; live plan progress on the dashboard
Device statelock_device unlock_device set_location pasteboard_get pasteboard_set video_record set_action_recordingLock/unlock, clipboard, screen recording, simulated GPS, action recording
Scheduling & scriptsschedule_task list_scheduled_tasks delete_scheduled_task list_saved_scriptsFire saved automations later, on a calendar, across the fleet
Utilitieshttp_request get_2fa_codeFetch external data mid-task (SSRF-safe URL validation); generate TOTP codes for sign-in flows

Example prompts

  • Go through the onboarding of the app I just installed and screenshot every screen.
  • Run the saved “morning report” automation on all connected iPhones and tell me if any device fails.
  • Navigate to the checkout screen in our app on every device and record a video of the flow.
  • Every day at 9:00, open the app, refresh the feed, and save a screenshot — schedule it.

Security & privacy

  • The server binds locally and talks only to the Icemoon dashboard on 127.0.0.1; devices are attached over USB.
  • Every tool call is written to a local audit log (sensitive parameters masked).
  • http_request validates URLs and blocks private-network targets.
  • Screen content reaches only the MCP client you connect (e.g. Claude), only while it's working for you.

FAQ

Is this open source? The docs, examples, and manifest in this repo are MIT. The server ships inside the commercial Icemoon app.

Does it work without the Icemoon app? No — the MCP server is the AI-facing surface of the app's automation engine.

Android? No, iOS only. That focus is where the human-like input and no-jailbreak setup come from.

Which plans include it? Farm and Scale — see pricing.

Related


Icemoon — automate real iPhones with human-like touch, from your Mac, no jailbreak.

來源:README.md,提交 375f92b

工具

0
工具後設資料尚未被收錄。

版本歷史

1
  1. v1.0.1最新Sep 16, 2026