Noureddine RAMDI / Clawd Cursor: Unified cross-platform desktop control for AI agents

Created Mon, 04 May 2026 10:23:01 +0000 Modified Sat, 23 May 2026 20:41:27 +0000
980 Words
github-stars typescript desktop-automation cross-platform ai-agents nodejs security

AmrDab/clawdcursor

Clawd Cursor tackles a problem that AI agent developers often run into: how to give AI models native, low-latency control over the desktop environment in a way that works across operating systems and without involving cloud dependencies. It provides a model-agnostic, OS-agnostic skill that lets AI agents manipulate the mouse, keyboard, screen, windows, and browser locally, via MCP or REST APIs, with strong security measures and a rich set of tools.

What Clawd Cursor does and its cross-platform architecture

At its core, Clawd Cursor is a desktop automation skill designed to be consumed by AI agents capable of tool-calling, such as Claude Code, Cursor, Windsurf, OpenClaw, or custom SDKs. It exposes two tool catalogs: a compact one with 6 compound tools totaling about 1,500 tokens in prompt footprint, and a granular catalog with 74 individual tools for fine-grained control.

The architecture is skill-first and local-first. The skill runs locally on 127.0.0.1, ensuring no cloud round-trips. This approach reduces latency and security risks associated with network transmission of sensitive desktop control commands.

A key technical component is the PlatformAdapter abstraction. This interface unifies the differences between Windows, macOS, and Linux, handling native GUI control under the hood. Instead of duplicating logic or maintaining separate code paths, the PlatformAdapter lets the skill provide a consistent API surface to the AI agent, regardless of the underlying OS.

Another distinctive feature is the unified blind/hybrid/vision pipeline. This setup consolidates three different GUI automation strategies into a single event loop, simplifying state management and improving reliability. Blind mode can perform actions without visual feedback, hybrid mode mixes blind with vision-based checks, and vision mode uses screen analysis. Combining these modes in one pipeline is engineering work worth noting.

The skill supports reading accessibility trees, which allows more semantic understanding of UI elements and better automation capabilities. It also integrates with browser windows, further broadening its control scope.

Technical strengths and design tradeoffs

Clawd Cursor’s model-agnostic design means it doesn’t depend on any particular AI backend. Instead, the skill focuses on providing a stable, secure interface to desktop control that any AI agent can invoke. This separation is practical and future-proof.

Security hardening is a focus in the latest release (v0.8.7). The skill adds shared safety gates for direct tool calls, aiming to prevent misuse or accidental harmful commands. This is crucial because desktop automation inherently risks disrupting user workflows or privacy.

The code quality appears solid, particularly around the PlatformAdapter abstraction. Handling Windows, macOS, and Linux under one interface reduces duplication and surface for platform-specific bugs. However, the tradeoff is complexity in the adapter implementation itself — it must contend with widely differing OS APIs and quirks.

The dual tool catalogs offer a nice balance. The compact catalog is efficient for prompt-based agents where token limits matter, while the granular catalog gives more control when needed. This design shows an awareness of real-world agent constraints.

Running entirely locally is both a strength and a limitation. It improves privacy and latency but requires the skill to be installed and maintained on each machine. It also means the skill must handle platform-specific dependencies and permissions, which the README documents clearly (e.g., macOS accessibility and screen recording permissions, Linux OCR and Wayland input tools).

Quick start

Installation is straightforward and well-documented for all three major OSes. Here are the commands exactly as provided:

Windows

powershell -c "irm https://clawdcursor.com/install.ps1 | iex"

macOS

curl -fsSL https://clawdcursor.com/install.sh | bash
clawdcursor grant     # Accessibility + Screen Recording

Linux

curl -fsSL https://clawdcursor.com/install.sh | bash

The installer clones the skill into ~/clawdcursor, runs npm install, builds, and registers a global clawdcursor shim via npm link. Runtime state such as auth tokens, pidfiles, and logs are stored in ~/.clawdcursor/.

Prerequisites include Node.js 20 or newer, and OS-specific dependencies like Xcode CLI tools on macOS, and tesseract-ocr, python3-gi, and ydotool or wtype on Linux for OCR and accessibility. The skill does not configure the AI agent host automatically; users must wire it in following provided documentation.

verdict

Clawd Cursor is a solid, practical tool for AI researchers and developers who need reliable, cross-platform desktop automation that respects privacy by staying local and supports multiple AI agent frameworks. Its unified PlatformAdapter and consolidated automation pipeline are engineering highlights that make the codebase manageable despite the inherent complexity of multi-OS GUI control.

The tradeoffs around local installation and dependency management are clear, but for many use cases where cloud round-trips are unacceptable, this is a reasonable price. The security hardening efforts are welcome, given the risks of desktop automation.

If your projects involve AI agents requiring native GUI interaction on diverse desktop environments, Clawd Cursor is worth exploring. It’s less suited for quick experiments without setup or for purely cloud-based workflows. The documentation and installation scripts are well-crafted, making onboarding relatively smooth for developers comfortable with Node.js environments and OS-level permissions.

Overall, Clawd Cursor stands as a practical, well-architected example of desktop automation skill design for the AI agent era.

→ GitHub Repo: AmrDab/clawdcursor ⭐ 297 · TypeScript