A deterministic governance layer for controlling AI agent tool execution and enforcing security policies.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install policy-engine
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install policy-engine using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Policy Engine?
The Policy Engine is a production-hardened governance layer designed to provide fine-grained control over AI agent actions. It acts as a deterministic gatekeeper by hooking into the tool call lifecycle, ensuring that every command executed by an agent complies with predefined security standards. Whether you are building complex multi-agent systems or simple automation scripts, this skill provides the necessary guardrails to prevent unauthorized tool usage and malicious command execution.\n\nBy leveraging Openclaw Skills like the Policy Engine, developers can implement strict tool profiles, path restrictions, and risk-based tiers. This ensures that agents operate within safe boundaries, protecting system integrity while maintaining high operational transparency through detailed audit logging and escalation tracking.
Policy Engine Use Cases
- Restricting sub-agents to read-only access for research tasks using Openclaw Skills.\n- Preventing accidental or malicious system-level commands like rm -rf or fork bombs.\n- Enforcing workspace boundaries to prevent directory traversal attacks.\n- Testing new security configurations in dry-run mode without disrupting workflows.\n- Scaling multi-agent teams with different permission levels and model routing.
How Policy Engine Works
- The engine intercepts an agent's request via the before_tool_call hook integrated within Openclaw Skills.\n2. It validates the requested tool against specific allowlists and pre-defined risk tiers (T0, T1, T2).\n3. Tool arguments are scanned against deny patterns to block dangerous strings or forbidden system paths.\n4. File paths are canonicalized and checked against allowed directory prefixes to prevent unauthorized access.\n5. Every decision is logged for audit purposes, and blocked attempts trigger escalation counters for the session.\n6. The tool call is either permitted to proceed, blocked with a remediation message, or logged without action in dry-run mode.
Policy Engine Setup
Install the skill using the command line:\n\nbash\nclawhub install policy-engine\n\n\nThen, enable the plugin in your openclaw.json configuration file to start using these Openclaw Skills:\n\njson\n{\n \"plugins\": {\n \"policy-engine\": {\n \"enabled\": true\n \"}\n }\n}\n
Policy Engine Data Schema & Taxonomy
The Policy Engine organizes governance data and session metadata using the following structures:\n\n| Component | Description |\n| :--- | :--- |\n| Tool Profiles | Named sets of allowed tools (e.g., 'readonly', 'coder') mapped to specific agents. |\n| Risk Tiers | Classification of tools into T0 (Read), T1 (Write), and T2 (Exec/System). |\n| Path Allowlists | Directory prefixes permitted for file operations to prevent path traversal. |\n| Deny Patterns | Regex or string patterns used to block specific arguments within tool calls. |\n| Escalation Stats | Session-based tracking of blocked attempts and retry counts per agent. |
Policy Engine Advanced Features
- Hot-reloading of configuration changes via gateway commands without needing to restart the agent.\n- Break-glass mode via environment variables for emergency administrative bypass of Openclaw Skills.\n- Fail-open architecture ensuring tool calls proceed if the governance engine encounters an internal error.\n- Per-agent model routing to optimize costs and performance across different LLM providers.\n- Integrated slash commands (/policy status) for real-time status checks and session resets.
SKILL.md
Loading
Related Openclaw Skills
Immich Photo Library (claw2immich)
Seamlessly interact with your Immich photo library through an AI agent using natural language and advanced metadata filtering.
Pixel Lobster
A transparent desktop overlay featuring a pixel art lobster that provides real-time lip-syncing for AI agent speech output.
Local-First LLM Router
A smart routing agent that prioritizes local LLMs over cloud APIs to reduce costs and enhance data privacy.
Adaptive Routing
A local-first routing engine that prioritizes privacy and cost-efficiency by intelligently escalating LLM requests to the cloud only when necessary.
LanceDB Hybrid Search (Memory Plugin)
A high-performance long-term memory plugin that combines semantic vector search with keyword-based BM25 retrieval for AI agents.
Govee Lights Control
A powerful automation skill for controlling Govee smart lights via API using natural language or command-line scripts.
Featured*
