ttguy0707/CyberClaw

What it solves

CyberClaw addresses the "black box" nature of AI agents by providing a transparent, controllable, and secure architecture. It prevents unexpected or dangerous AI actions through a zero-trust execution model and provides full traceability of the agent's decision-making process.

How it works

The system is built on a LangGraph-based agent loop that integrates several key mechanisms:

• Two-Phase Invocation: Instead of executing a tool immediately, the agent first calls a help mode to read the tool's SKILL.md manual before proceeding to the run mode.
• Dual-Watermark Memory: It combines a long-term user profile (user_profile.md) with short-term conversation summaries stored in SQLite to maintain context without token explosion.
• Behavior Auditing: Every action—including LLM inputs, tool calls, and system actions—is logged in JSONL format and can be monitored in real-time via a dedicated Rich terminal UI.
• Heartbeat Engine: A background process manages scheduled and recurring tasks independently of the main chat interface.
• Security Sandbox: All file and shell operations are restricted to a specific office/ directory with regex-based command blocking and path interception.

Who it’s for

• Enterprise Users: Those requiring strict audit logs and permission controls for AI deployments.
• AI Researchers: Developers analyzing agent decision chains and safety boundaries.
• Power Users: Individuals wanting a personalized assistant with scheduled automation and cross-platform shell capabilities.

Highlights

• Zero-Trust Execution: Reduces P0 incident rates by 80% by forcing the agent to read documentation before acting.
• Ecosystem Compatibility: Directly supports skills from OpenClaw and Claude Code.
• Transparent Monitoring: Real-time visibility into the agent's internal logic via a separate monitor terminal.
• Cross-Platform: Native support for both Windows (PowerShell/CMD) and Unix (Bash/Zsh) with automatic path and command adaptation.