Summary

Clawd on Desk provides a real-time, animated desktop pet visualizing AI coding agent activity (Claude Code, Codex CLI, Copilot CLI). It enhances developer workflow with immediate, non-intrusive feedback on agent states like thinking, typing, and errors, improving situational awareness.

How It Works

Utilizing Electron, Clawd is an always-on-top, transparent desktop app integrating with AI agents via command hooks or log polling. Agent events map to 12 distinct animations via a state machine. Key features include real-time state awareness, eye-tracking cursor following, interactive clicks, and a "Permission Bubble" for Claude Code's tool requests, overlaying dialogues directly on the desktop.

Quick Start & Requirements

• Installation: Clone, npm install, npm start.
• Prerequisites: Node.js.
• Supported OS: Windows 11, macOS.
• Agent Integration: Claude Code/Codex CLI auto-integrate; Copilot CLI requires manual ~/.copilot/hooks/hooks.json setup.
• Remote Access: Supports remote agent sessions via SSH reverse port forwarding.
• macOS DMG: Unsigned, requires manual bypass.

Highlighted Details

• Multi-Agent Coexistence: Tracks Claude Code, Codex CLI, and Copilot CLI simultaneously.
• Interactive Permission Bubble: One-click interface for Claude Code tool permissions.
• Remote Session Support: Visualizes remote AI agent activity via SSH tunneling.
• Mini Mode: Compact view hiding at screen edge, peeking on hover.
• Eye Tracking: Follows user's cursor.

Maintenance & Community

A community-driven project welcoming contributions. Roadmap includes exploring Codex terminal focus and sound effects.

Licensing & Compatibility

• License: MIT License, permitting broad commercial use.
• Compatibility: Generally compatible with closed-source applications. macOS DMG requires manual installation bypass.

Limitations & Caveats

Codex CLI lacks terminal focus and has ~1.5s latency on Windows (log polling). Copilot CLI requires manual hook setup and lacks permission bubbles. macOS DMG is unsigned. Core Electron process lacks automated testing.