LINE Desktop MCP

English | 繁體中文

🍴 About this fork

This is a fork of dtwang/line-desktop-mcp by Geoffrey Wang, kept under the original MIT License (see LICENSE.md; copyright remains with the original author). All upstream functionality is preserved; this fork adds a macOS reading path and an idle-aware scheduler, and hardens a handful of automation bugs. Fork maintained by Sung Yeh (@Sung-tsan).

Installing this fork (not the upstream npm package): the macOS OCR/idle features below live in this repository only — they are not published to npm. npx line-desktop-mcp@latest pulls Geoffrey Wang's upstream package, which does not include them. To use this fork, clone this repo, run bash native/build.sh, and point your MCP client at node <path>/src/server.js (see the macOS install section).

Why this fork

1) Bug fixes contributed against the upstream GUI-automation path

• osascript calls had no timeout. If LINE was blocked by a modal/pop-up, an unattended osascript could hang forever. Every AppleScript call now runs through a single osa() wrapper with a hard timeout (src/automation/macos-line-automation.js).

• Chat selection wasn't verified. Selecting a chat by name could silently open the wrong room. Selection now reads back the opened chat's header and aborts on mismatch, and the OCR path matches the sidebar row by exact name.

• send_message_auto had no guard (a prompt-injection / accidental-send risk for an agent-driven tool). It is now disabled by default behind the LINE_MCP_ALLOW_AUTO_SEND=true environment gate (src/server.js).

• AppleScript pitfalls hardened: identifiers must not start with an underscore (set _x … raises compile error -2741), and a handler that uses a top-level variable must declare it global on both sides — both are easy to trip over and are avoided throughout the AppleScript in this fork.

2) New reading method — screenshot + OCR + idle-aware scheduling (macOS)

On newer macOS LINE builds, two hard facts make the classic approaches fail:

• Message text is self-drawn and never enters the Accessibility tree (AXStaticText values are empty), so AX cannot read message content. Upstream reads by driving the clipboard (Cmd+A / Cmd+C), which steals focus, keyboard, and the clipboard.

• LINE drops its onscreen window ~1 second after losing focus, so background screenshots are not possible either.

This fork's macOS read path instead uses:

• Foreground screenshot + Apple Vision OCR (excellent Traditional-Chinese accuracy, zero LLM tokens for text extraction) — no clipboard hijack, no keystrokes. Chat switching is done with a real click (cliclick) because AXPress / set-selected do nothing on these sidebar rows.

• Idle-aware scheduling: a launchd agent fires a few times a day, but the runner waits for you to be idle (HID idle ≥ N minutes) before it scans, so it never interrupts you mid-work.

• Cursor + focus restoration: it snapshots the frontmost app and mouse position before scanning and restores them afterward (even on error).

Honest caveat: this read path is not background or non-hijacking. To capture, it briefly brings LINE to the foreground and moves the mouse to click a chat; it just restores your app and cursor when done, and is idle-gated so it only runs while you are away from the keyboard.

Limitations of the OCR/idle path

• macOS only (uses Apple Vision, screencapture, cliclick, launchd). The Windows path is unchanged from upstream.

• Requires a one-time Screen Recording permission. For scheduled runs, the node binary launchd spawns needs this permission of its own — granting it to your terminal is not enough.

• LINE must be logged in, in expanded-chat-window mode, with a window that can be opened (not minimized to the menu bar), on the primary display.

• Pull-based, not real-time: it reads on a schedule / on demand, not on push.

• Sender/time separation from OCR is best-effort heuristic; the raw OCR lines are always kept so nothing is silently lost.

Usage (macOS OCR scan)

# 1. Build the native helpers (needs Xcode Command Line Tools)
bash native/build.sh

# 2. Grant Screen Recording permission to your terminal (and, for scheduled
#    runs, to the node binary) under System Settings > Privacy & Security.

# 3. See which chatrooms are visible, to build your blocklist (no reads):
node scripts/scan-once.js --dry

#    Put names you never want scanned into src/scan/state/blocklist.json
#    (copy from blocklist.example.json).

# 4. Run one sweep now (only scans if you've been idle >= 5 min; --force to override):
node scripts/scan-once.js --force
#    Outputs land in src/scan/state/out/  (structured JSON + a Markdown daily report).

# 5. Install the idle-aware schedule (launchd, 3x/day by default):
scripts/install-schedule.sh install     # status | uninstall

The three MCP read tools (get_line_chatroom_history_*) are rerouted to this screenshot+OCR engine on macOS; their names and parameters are unchanged.

Related MCP server: line-oa-mcp-ultimate

繁體中文

透過 MCP（Model Context Protocol），使 AI 工具能夠與 LINE Desktop 整合，並執行訊息的讀取與發送操作。

⚠️ 重要說明

這個專案不是 LINE 官方的 line-bot-mcp-server

如果你要找的是官方版本，請前往：https://github.com/line/line-bot-mcp-server

與官方版本的差異

• 官方 line-bot-mcp-server：透過 LINE Messaging API 操作 LINE Bot

• 本專案 line-desktop-mcp：透過 MCP 在 Windows 或 Mac 上直接操作 LINE Desktop 應用程式

重要聲明

1. 本專案與 LINE 官方無任何關聯
   This project is NOT officially affiliated with LINE.

2. 無需申請 LINE Developers 或使用 Channel Access Token
   本專案透過已經完成登入的 LINE Desktop 應用程式進行操作，不需要申請開發者帳號或 API Token。

關於專案

LINE Desktop MCP 是一個基於 Model Context Protocol 的整合工具，讓 AI 工具（如 Claude Desktop, n8n ），能夠直接與 LINE Desktop 應用程式互動。透過此專案，您可以：

• 📖 讀取 LINE 聊天訊息

• ✉️ 發送 LINE 訊息（手動或自動）

• 🤖 將 LINE 整合到您的 AI 工作流程中

功能特色

• 🤖 AI 整合：透過 MCP 協議與 Claude Desktop、 n8n 等 AI 工具無縫整合

• 💬 訊息操作：支援讀取和發送 LINE 訊息

• 🖥️ 桌面整合：直接與 LINE Desktop 應用程式互動

• 🔄 自動化支援：可選擇手動確認或自動發送訊息

系統需求

基本需求

• LINE Desktop：v9.10 或以上版本

• 作業系統：

  • Windows 10 或以上版本

  • macOS Ventura 13.0 或以上版本（需要 AppleScript 支援）

與 Claude Desktop 整合

• Claude Desktop App：最新版本

• Claude 訂閱方案：Pro 方案

與 n8n 整合

• n8n：支援 MCP 的版本

安裝方式

Windows

1. 安裝 Node.js

   • 參考微軟官方文件：https://learn.microsoft.com/zh-tw/windows/dev-environment/javascript/nodejs-on-windows

2. 安裝 AutoHotkey v2

   • 下載並安裝：https://www.autohotkey.com/

3. 設定 Claude Desktop

   • 開啟 Claude Desktop 設定檔

   • 在 mcpServers 中加入以下設定：

{
  "mcpServers": {
    "line-desktop-mcp": {
      "command": "npx",
      "args": ["line-desktop-mcp@latest"]
    }
  }
}

macOS（本 fork 的 OCR／閒置排程功能）

⚠️ npx line-desktop-mcp@latest 會裝到上游 npm 套件，不含本 fork 的 OCR／ 閒置排程。請改用下面的 clone + 本地路徑方式。

1. 安裝 Node.js 與 cliclick

   • brew install node cliclick

2. Clone 並建置原生工具

   git clone https://github.com/Sung-tsan/line-desktop-mcp.git
   cd line-desktop-mcp && npm install && bash native/build.sh

3. 設定 Claude Desktop（指向本地 fork，非 npx）

   • 開啟 Claude Desktop 設定檔，在 mcpServers 中加入：

{
  "mcpServers": {
    "line-desktop-mcp": {
      "command": "node",
      "args": ["/絕對路徑/line-desktop-mcp/src/server.js"]
    }
  }
}

4. 授予「螢幕錄製」權限（系統設定 › 隱私權與安全性），OCR 讀取與排程掃描皆需要。

macOS 的截圖＋OCR 掃描與閒置排程用法，見上方英文 Usage (macOS OCR scan) 段。

進階設定

Streamable HTTP 模式

除了預設的 stdio 模式外，本專案也支援透過 Streamable HTTP 方式運行。此模式特別適合在 n8n 等支援 MCP 的平台中使用。

啟動 Streamable HTTP 模式：

# 本機使用（僅 localhost）
npx line-desktop-mcp@latest --http-mode --port 3000

# 開放外部連線（需搭配 token）
npx line-desktop-mcp@latest --http-mode --host 0.0.0.0 --port 3000 --token YOUR_SECRET

參數說明：

• --http-mode：啟用 Streamable HTTP 模式，使用 HTTP streaming 而非 stdio

• --port <port>：指定 HTTP 伺服器的 port（預設：3000）

• --host <host>：指定綁定的網路介面（預設：127.0.0.1）

• --token <secret>：設定 Bearer Token 驗證密鑰。當 --host 設為非 loopback 位址時為必填，以確保安全性

MCP 端點配置：

本機連接：

http://127.0.0.1:3000/mcp

Docker 中的 n8n 連接（同一台機器）：

http://host.docker.internal:3000/mcp

傳輸方式：

• POST 請求：發送 JSON-RPC 訊息並透過 SSE stream 接收回應

• 支援 session 管理，每個連接會獲得唯一的 session ID

n8n 工作流程範例：

如果您想在 n8n 中使用 LINE Desktop MCP，可以下載我們提供的範例工作流程檔案：

• 📥 下載 n8n 工作流程範例

此範例展示如何在 n8n 中整合 LINE Desktop MCP 建立聊天機器人工作流程。

使用方式

在 Claude Desktop 的對話中，您可以使用以下方式操作 LINE：

1. 讀取聊天內容

請幫我讀取 LINE 群組『專案討論』的訊息，並作出總結

2. 發送訊息（手動確認）

請幫我撰寫一個問候，發送到 LINE 群組『專案討論』中

Claude 會先撰寫訊息內容，等待您確認後再發送。

3. 發送訊息（自動送出）

請幫我撰寫一個問候，發送到 LINE 群組『專案討論』中，並自動發送

Claude 會撰寫訊息並自動完成發送動作。

使用注意事項

重要提醒

1. 避免干擾自動化操作
   本工具透過圖形使用者介面（GUI）進行自動化操作。在自動化程式執行期間，請勿同時使用滑鼠進行其他操作，以免干擾程式運作。

2. LINE Desktop 視窗配置要求
   請確保 LINE Desktop 使用「展開聊天視窗」模式。在此模式下，聊天視窗會固定顯示在聊天室清單的右側，而非以獨立視窗方式開啟。

3. 多顯示器環境配置
   若您使用多個顯示器,請將 LINE Desktop 應用程式放置於主要顯示器（第一個顯示器）上,以確保自動化功能正常運作。

授權條款

本專案採用 MIT 授權條款 - 詳見 LICENSE.md 檔案

作者

Geoffrey Wang — original author

• GitHub: @dtwang

• Threads: @geoff_spacetime

Sung Yeh — fork maintainer (macOS OCR / idle-scheduler)

• GitHub: @Sung-tsan

English

Integrate AI tools with LINE Desktop through MCP (Model Context Protocol) to enable message reading and sending operations.

⚠️ Important Notice

This project is NOT the official LINE line-bot-mcp-server

If you're looking for the official version, please visit: https://github.com/line/line-bot-mcp-server

Differences from Official Version

• Official line-bot-mcp-server: Operates LINE Bot through LINE Messaging API

• This project line-desktop-mcp: Directly operates LINE Desktop application on Windows or Mac through MCP

Important Disclaimer

1. This project is NOT officially affiliated with LINE
   本專案與 LINE 官方無任何關聯。

2. No need to apply for LINE Developers or use Channel Access Token
   This project operates through the already logged-in LINE Desktop application, without requiring developer account registration or API tokens.

About

LINE Desktop MCP is an integration tool based on the Model Context Protocol that allows AI tools (such as Claude Desktop, n8n ) to interact directly with the LINE Desktop application. With this project, you can:

• 📖 Read LINE chat messages

• ✉️ Send LINE messages (manual or automatic)

• 🤖 Integrate LINE into your AI workflows

Features

• 🤖 AI Integration: Seamlessly integrate with AI tools like Claude Desktop, n8n through the MCP protocol

• 💬 Message Operations: Support for reading and sending LINE messages

• 🖥️ Desktop Integration: Direct interaction with the LINE Desktop application

• 🔄 Automation Support: Choose between manual confirmation or automatic message sending

System Requirements

Basic Requirements

• LINE Desktop: v9.10 or above

• Operating System:

  • Windows 10 or above

  • macOS Ventura 13.0 or above (requires AppleScript support)

Integration with Claude Desktop

• Claude Desktop App: Latest version

• Claude Subscription: Pro plan

Integration with n8n

• n8n: Version with MCP support

Installation

Windows

1. Install Node.js

   • Follow Microsoft's official guide: https://learn.microsoft.com/en-us/windows/dev-environment/javascript/nodejs-on-windows

2. Install AutoHotkey v2

   • Download and install: https://www.autohotkey.com/

3. Configure Claude Desktop

   • Open Claude Desktop configuration file

   • Add the following to mcpServers:

{
  "mcpServers": {
    "line-desktop-mcp": {
      "command": "npx",
      "args": ["line-desktop-mcp@latest"]
    }
  }
}

macOS (this fork's OCR / idle-scheduling features)

⚠️ npx line-desktop-mcp@latest installs the upstream npm package, which does not include this fork's OCR / idle scheduler. Use the clone + local path below instead.

1. Install Node.js and cliclick

   • brew install node cliclick

2. Clone and build the native helpers

   git clone https://github.com/Sung-tsan/line-desktop-mcp.git
   cd line-desktop-mcp && npm install && bash native/build.sh

3. Configure Claude Desktop (point at the local fork, not npx)

   • Open the Claude Desktop config file and add to mcpServers:

{
  "mcpServers": {
    "line-desktop-mcp": {
      "command": "node",
      "args": ["/absolute/path/line-desktop-mcp/src/server.js"]
    }
  }
}

4. Grant Screen Recording permission (System Settings › Privacy & Security) — required for OCR reading and scheduled scans. See Usage (macOS OCR scan) above.

Advanced Configuration

Streamable HTTP Mode

In addition to the default stdio mode, this project also supports running via Streamable HTTP. This mode is particularly suitable for use with platforms like n8n that support MCP.

Start Streamable HTTP Mode:

# Local use (localhost only)
npx line-desktop-mcp@latest --http-mode --port 3000

# Allow external connections (token required)
npx line-desktop-mcp@latest --http-mode --host 0.0.0.0 --port 3000 --token YOUR_SECRET

Parameters:

• --http-mode: Enable Streamable HTTP mode, using HTTP streaming instead of stdio

• --port <port>: Specify the HTTP server port (default: 3000)

• --host <host>: Specify the network interface to bind (default: 127.0.0.1)

• --token <secret>: Set Bearer Token authentication secret. Required when --host is set to a non-loopback address for security

MCP Endpoint Configuration:

Local connection:

http://127.0.0.1:3000/mcp

n8n in Docker (same machine):

http://host.docker.internal:3000/mcp

Transport Method:

• POST requests: Send JSON-RPC messages and receive responses via SSE stream

• Supports session management with unique session IDs for each connection

n8n Workflow Example:

If you want to use LINE Desktop MCP in n8n, you can download our sample workflow file:

• 📥 Download n8n Workflow Example

This example demonstrates how to integrate LINE Desktop MCP in n8n to create a chatbot workflow.

Usage

In Claude Desktop conversations, you can interact with LINE in the following ways:

1. Read Chat Messages

Please read the messages from LINE group 'Project Discussion' and summarize them

2. Send Messages (Manual Confirmation)

Please write a greeting and send it to LINE group 'Project Discussion'

Claude will compose the message and wait for your confirmation before sending.

3. Send Messages (Automatic)

Please write a greeting and send it to LINE group 'Project Discussion', and send it automatically

Claude will compose the message and automatically complete the sending action.

Usage Precautions

Important Reminders

1. Avoid Interfering with Automation
   This tool performs automation through the graphical user interface (GUI). During automated operations, please refrain from using the mouse for other tasks to prevent interference with the program's execution.

2. LINE Desktop Window Configuration
   Please ensure that LINE Desktop is configured in "Expanded Chat Window" mode. In this mode, the chat window remains docked to the right side of the chat list, rather than opening as a separate independent window.

3. Multi-Monitor Setup
   If you are using multiple monitors, please ensure that the LINE Desktop application is positioned on the primary display (first monitor) for the automation to function correctly.

License

This project is licensed under the MIT License - see the LICENSE.md file for details

Author

Geoffrey Wang — original author

• GitHub: @dtwang

• Threads: @geoff_spacetime

Sung Yeh — fork maintainer (macOS OCR / idle-scheduler)

• GitHub: @Sung-tsan