macos-sys-assist
A secure, constraint-based macOS OS-level automation MCP server for AI assistants.
README
macos-sys-assist
A focused macOS automation MCP server for reliable input simulation, window management, and window-specific screenshots — the three things AppleScript and bash do poorly.
What Is This?
macos-sys-assist is a Python-based MCP server that fills the gaps where bash + Chrome DevTools fall short. It uses pyobjc (native macOS APIs) and Core Graphics for low-level input simulation — more reliable than AppleScript's keystroke.
What It Does That bash/CDP Can't
| Capability | Why Not bash/CDP |
|---|---|
| Core Graphics click/type/key | AppleScript keystroke misses keys or fails silently. This uses CGEventPost — the same API macOS uses internally. |
| Window-specific screenshots | bash screencapture captures the full screen; cropping is tedious. This captures just the window you want. |
| Precise window geometry | osascript returns position inconsistently. This uses the Accessibility API for accurate pixel-level data. |
| Multi-app window layouts | Arrange 3+ apps at specific positions in one command. bash needs multiple chained osascript calls. |
What It Does NOT Do (Use bash Instead)
| Tool | Why Use bash |
|---|---|
| Finding files | find / mdfind are simpler |
| Reading files | cat / python3 -c |
| Opening files | open command |
| App queries | osascript -e 'tell app "System Events"...' |
| Clipboard | pbpaste / pbcopy |
| Screen resolution | system_profiler SPDisplaysDataType |
Quick Start
Installation
git clone https://github.com/YOUR_USERNAME/macos-sys-assist.git
cd macos-sys-assist
./setup.sh
Grant Permissions
- Accessibility — System Settings → Privacy & Security → Accessibility → Add Terminal/Python
- Screen Recording (for screenshots) — System Settings → Privacy & Security → Screen Recording → Add Terminal/Python
Configure Apps
Edit allowed_apps.json to control which apps can be automated.
Usage
Standalone Mode
./run.sh
OpenCode Integration
Add to opencode.jsonc:
"mcp": {
"macos-sys-assist": {
"type": "local",
"command": ["/path/to/macos-sys-assist/run.sh"],
"enabled": true
}
}
Direct Python Usage (via bash)
.venv/bin/python3 -c "
import sys
sys.path.insert(0, '.')
from macos.input import InputSimulator
InputSimulator().click_at(100, 200, 'left')
"
Tool Reference
Input Simulation (Core Graphics)
| Tool | Description | Security |
|---|---|---|
click_at(x, y, button, double) |
Click at screen coordinates | ⚠️ Confirmation |
type_string(text) |
Type text character by character | ⚠️ Confirmation, max 500 chars |
press_key(combination) |
Press key combo (e.g., cmd+tab) |
⚠️ Blocked combos enforced |
More reliable than AppleScript — uses CGEventPost instead of keystroke.
Window Management
| Tool | Description | Security |
|---|---|---|
move_window(x, y) |
Move active window to coords | ⚠️ Confirmation |
resize_window(width, height) |
Resize active window | ⚠️ Confirmation |
get_window_geometry(pid) |
Get window position/size (Accurate) | Read-only |
Uses Accessibility API for pixel-level accuracy. More reliable than osascript.
Screenshots (Requires Screen Recording Permission)
| Tool | Description |
|---|---|
screenshot(filepath, display_id) |
Capture full screen |
screenshot_window(pid, filepath) |
Capture specific window only — no cropping needed |
screenshot_region(x, y, w, h, filepath) |
Capture a screen region |
get_displays() |
Get all connected displays and resolutions |
When to Use This vs bash
✅ Use macos-sys-assist when:
- AppleScript
keystrokeorclickfails silently - You need a screenshot of just one window without browser chrome
- You're arranging 3+ app windows at specific positions for a workspace
- The task requires pixel-level coordinate accuracy
❌ Use bash when:
- Finding files (
find,mdfind,ls) - Reading files (
cat,python3 -c) - Opening files (
open) - Basic clipboard (
pbpaste,pbcopy) - Checking what app is frontmost (
osascript) - Launching apps (
open -a)
🔄 Use Chrome DevTools when:
- Interacting with web pages (clicking buttons, filling forms)
- Uploading files to websites (base64 injection into
<input type="file">) - Reading page content
- Navigating multi-page web flows
Configuration
allowed_apps.json
Controls which apps can be automated:
{
"allowed_apps": [
{
"bundle_id": "com.brave.Browser",
"name": "Brave Browser",
"allow_actions": true
}
],
"global_settings": {
"require_confirmation_for_click": true,
"require_confirmation_for_type": true,
"max_string_length": 500,
"blocked_key_combinations": [
"cmd+q",
"cmd+delete",
"ctrl+alt+delete"
]
}
}
Project Structure
macos-sys-assist/
├── server.py # Main MCP server entry point
├── config.py # Configuration management
├── security.py # Security validation layer
├── allowed_apps.json # Application allow-list
├── requirements.txt # Python dependencies
├── setup.sh # Installation script
├── run.sh # Wrapper script
├── macos/ # Native macOS API wrappers
│ ├── accessibility.py # App queries, PID lookup
│ ├── window.py # Window move/resize/geometry
│ ├── input.py # Core Graphics click/type/key
│ ├── screenshot.py # Screen capture (full/window/region)
│ └── task_engine.py # Multi-step task execution
└── tools/ # MCP tool definitions
├── information.py # get_window_geometry
├── actions.py # click_at, type_string, press_key, move/resize
└── screenshot.py # screenshot, screenshot_window, screenshot_region, get_displays
Roadmap
Completed ✅
- [x] Core Graphics input simulation (click, type, key)
- [x] Window management (move, resize, geometry)
- [x] Window-specific screenshots (no cropping)
- [x] Security layer (allow-list, blocked keys, confirmations)
Planned 📋
- [ ] Folder Watcher — Detect new files in Downloads, auto-organize by project
- [ ] System State — Battery, WiFi, disk space checks before long automations
- [ ] Window Layout Presets — Save/restore multi-app workspaces
- [ ] Calendar Integration — Meeting-aware automation scheduling
Security Model
Design Principles
- No Shell Access — All operations use native macOS APIs
- Explicit Allow-List — Only pre-approved apps can be controlled
- Human-in-the-Loop — Invasive actions require user confirmation
- Input Validation — Text length limits, key combo blocking
What's Blocked
| Threat | Mitigation |
|---|---|
| Unauthorized app control | Application allow-list |
| Destructive key combos | Blocked combinations list |
| Excessive text input | Maximum string length (500) |
| Unconfirmed actions | Confirmation prompts |
Troubleshooting
"Accessibility permission not granted"
- System Settings → Privacy & Security → Accessibility
- Add Terminal.app or
.venv/bin/python3 - Ensure toggle is ON
- Restart the server
"Screen Recording permission required"
- System Settings → Privacy & Security → Screen Recording
- Add Terminal.app or
.venv/bin/python3 - Ensure toggle is ON
- Restart the server
"App not in allow-list"
- Find the app's bundle ID:
osascript -e 'id of app "AppName"' - Add it to
allowed_apps.json - Restart the server
License
MIT License — see LICENSE
Acknowledgments
Built for the OpenCode AI assistant framework.
Uses the Model Context Protocol for tool integration.
Powered by pyobjc for native macOS API access.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。