Openclaw Guardian

OpenClaw Guardian

The missing safety layer for AI agents.

What It Does

Guardian sits between the AI's decision and actual execution, automatically assessing risk and routing dangerous operations through independent Guardian Agents for voting-based approval.

95% of operations pass instantly (zero latency, zero cost). Only the ~5% that are potentially dangerous trigger Guardian review.

Architecture

Tool Call → Risk Assessor (keyword rules, 0ms)
              ↓
   Score 0-30  → Fast Lane (just execute)
   Score 31-70 → Light Review (1 Guardian, ~1-2s)
   Score 71-100 → Full Vote (3 Guardians, ~2-4s, majority rules)

Guardian Perspectives (Full Vote)

Installation

1. Clone into your OpenClaw workspace:

cd ~/.openclaw/workspace
git clone https://github.com/fatcatMaoFei/openclaw-guardian.git

2. Register the plugin in openclaw.json:

{
  "plugins": {
    "load": {
      "openclaw-guardian": {
        "path": "workspace://openclaw-guardian",
        "enabled": true
      }
    }
  }
}

3. Restart OpenClaw gateway.

Configuration

Edit default-policies.json to enable/disable:

{ "enabled": true }

File Structure

• index.ts — Entry point; registers before_tool_call hook
• src/blacklist.ts — Keyword rule engine for risk scoring
• src/llm-voter.ts — Tiered LLM voting with parallel execution
• src/audit-log.ts — SHA-256 hash-chain audit logger
• openclaw.plugin.json — Plugin manifest
• default-policies.json — Default risk policies (user-customizable)

Token Cost

Average overhead: ~15-35% of total token usage.

License

MIT