openclaw-guardian_skill

Overview

This skill provides a safety layer plugin for OpenClaw that intercepts high-risk AI operations and enforces a voting-based approval flow before execution. It uses a fast keyword rule engine to auto-pass low-risk calls and invokes lightweight or full multi-Guardian voting for the small fraction of operations that need human-like review. The design minimizes latency and token cost by only running LLM-based Guardians for about 5% of operations.

How this skill works

Every tool call is scored by a zero-cost keyword rule engine that assigns a numeric risk score. Low scores go through a fast lane and execute immediately; medium scores trigger a single Guardian for a light review; high scores trigger a full vote where multiple Guardians with distinct perspectives evaluate the action and majority rules. Audit logging and configurable policies let you track decisions and customize thresholds.

When to use it

• Protect destructive operations such as file deletion or mass data modification
• Gate access to credentials, API keys, or other sensitive secrets
• Prevent unauthorized configuration changes or privilege escalations
• Control external network requests or data exfiltration attempts
• Add automated, low-latency safety checks to production AI agents

Best practices

• Tune keyword rules and risk thresholds to your environment to keep false positives low
• Start with conservative Light Review thresholds, then tighten Full Vote triggers as you observe behavior
• Keep default-policies.json in source control and document any team overrides
• Monitor the audit log (SHA-256 hash-chain) for incident review and compliance
• Limit which tools are subject to the Guardian to reduce unnecessary review overhead

Example use cases

• An agent attempts to delete project files — system scores high and triggers Full Vote before execution
• A tool requests an external webhook — keyword rule flags medium risk and a single Guardian reviews parameters
• Configuration change requests are routed through the Guardian to ensure the user has required permissions
• Credential or API key exposure is blocked by Privacy-focused Guardian perspective and logged for later audit
• Large-scale database write operations are evaluated for reversibility and cost before approval