Provides an OS-agnostic, model-agnostic desktop automation server that equips AI agents with the ability to interact with graphical user interfaces on Windows, macOS, and Linux. It enables any tool-calling AI agent to control the mouse, keyboard, windows, and read screen content locally, offering ground-truth verification for agent actions without requiring app-specific integrations or cloud round-trips.

How It Works

Clawd Cursor functions as a "skill" rather than a standalone application, exposing a local REST or Model Context Protocol (MCP) interface. It leverages platform-specific adapters (Windows UI Automation, macOS JXA, Linux AT-SPI) to abstract underlying OS interactions. Key design principles include model-agnosticism (supporting various LLMs like Claude, GPT, Gemini, Ollama), OS-agnosticism through a unified interface, and a skill-first approach where AI agents are the primary consumers. This design allows agents to interact with any on-screen element, running entirely locally on 127.0.0.1 for enhanced privacy and reduced latency.

Quick Start & Requirements

• Installation:
  • Windows: powershell -c "irm https://clawdcursor.com/install.ps1 | iex"
  • macOS/Linux: curl -fsSL https://clawdcursor.com/install.sh | bash followed by clawdcursor grant for necessary permissions.
• Prerequisites: Node.js 20+, macOS Xcode CLI tools, Linux packages (tesseract-ocr, python3-gi, gir1.2-atspi-2.0, ydotool or wtype for Wayland), and an AI provider key configured on the agent side.
• Links: Website: clawdcursor.com, Install scripts available via provided URLs.

Highlighted Details

• Ground-Truth Verification: Employs six independent signals (pixel diff, window state, focus, OCR, assertions, anti-pattern detection) to validate task completion, preventing agents from self-reporting success.
• Safety Tiers: Tools are categorized into 'Auto', 'Preview', and 'Confirm' tiers, with explicit user approval required for sensitive actions like sending messages or making purchases.
• Local Execution: All operations are confined to 127.0.0.1, ensuring data privacy and minimizing latency by avoiding cloud round-trips.
• Dual Tool Catalogs: Offers both a compact surface (6 compound tools) optimized for prompt footprint and a granular catalog (74 individual tools) for agents preferring distinct verbs.

Maintenance & Community

The project shows active development, with the latest release being v0.8.7 focused on security hardening. Release notes detail ongoing maintenance, bug fixes, and dependency updates. Community support is available via Discord (clawdcursor.com/discord).

Licensing & Compatibility

Licensed under the MIT License, which is permissive and generally allows for commercial use and integration into closed-source projects without significant restrictions.

Limitations & Caveats

As indicated by its v0.8.x release series, Clawd Cursor is under active development, and users should anticipate potential breaking changes. Linux Wayland support requires specific package installations. Connecting agents necessitates manual configuration steps, such as adding MCP entries. Browser automation relies on the Chrome DevTools Protocol (CDP).