Keel provides structural safety guardrails for autonomous agents by intercepting destructive actions and requiring structured human approval before execution.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install threshold-keel
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 threshold-keel 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 Keel: Structural Safety for AI Agents?
Keel is a sophisticated safety layer designed for autonomous agents with high-privilege access to sensitive user data like emails, files, and messaging systems. It acts as a structural buffer, ensuring that every action modifying the external world passes through a validation gate. By classifying actions into risk tiers (T0 to T3), Keel prevents irreversible operations from occurring without explicit, informed consent.
As a vital component of the Openclaw Skills ecosystem, Keel offers two primary modes: a behavioral instructions-only mode and a high-integrity CLI mode. The CLI mode leverages SHA-256 hash chains to create tamper-evident audit trails and deterministic policy enforcement. Whether running locally or synced with Threshold Cloud, Keel ensures your AI agent remains a helpful assistant rather than a liability.
Keel: Structural Safety for AI Agents Use Cases
- Preventing accidental deletion of critical files or system directories during autonomous tasks.
- Enforcing a structured human-in-the-loop approval process for sending outgoing emails or messages.
- Auditing agent actions via a cryptographic Write-Ahead Log (WAL) for compliance and security reviews.
- Protecting against malicious third-party plugins through automated skill vetting and permission summaries.
- Implementing blast-radius caps to prevent runaway automation from exceeding rate limits on APIs or shell commands.
How Keel: Structural Safety for AI Agents Works
- Classification: Every proposed tool call or action is automatically classified into one of four risk tiers (Read-only, Reversible, Quarantine-capable, or Irreversible).
- Policy Check: The action is validated against a local or cloud-synced policy store (store.json) to check for explicit blocks or constraints.
- Approval Gating: For high-risk T2 and T3 actions, the agent pauses execution and requests structured approval from the user, including a preview of the destructive payload.
- Logging: Every event—whether proposed, approved, or blocked—is recorded in a Write-Ahead Log (WAL) with SHA-256 integrity hashing.
- Execution/Quarantine: Approved actions are executed, while requests for deletion are redirected to a secure quarantine folder for a cooling-off period.
Keel: Structural Safety for AI Agents Setup
To get started with Keel and integrate it into your Openclaw Skills workflow, follow these steps:
- Install the CLI:
pip install threshold-keel
- Initialize Keel:
keel init
- Verify Status:
keel status --human
- (Optional) Cloud Sync:
export KEEL_CLOUD_API_KEY=sk-keel-your-key-here
Keel: Structural Safety for AI Agents Data Schema & Taxonomy
Keel manages safety data through a structured local filesystem hierarchy located in ~/.keel/:
| Component | Path | Description |
|---|---|---|
| Policy Store | store.json |
Contains versioned, priority-based safety constraints. |
| Audit Log | wal/ |
Directory containing JSONL files with cryptographic hash-linked events. |
| Quarantine | quarantine/ |
Secure storage for files flagged for deletion or isolation. |
| Metadata | fidelity.json |
Internal state for tracking session integrity and mode status. |
Keel: Structural Safety for AI Agents Advanced Features
- Cryptographic Hash Chains: Every log entry is linked to the previous via SHA-256, ensuring the audit trail is tamper-evident.
- Blast Radius Caps: Hard hourly limits on state-changing actions (e.g., 10 emails/hour) to stop runaway processes.
- Skill Vetting System: Automatically analyzes and flags new skills that request dangerous permissions like shell access or credential reading.
- Deterministic Policy Evaluation: CLI-backed policy engine that provides machine-verifiable exit codes for agent logic.
- Threshold Cloud Integration: Sync policies and audit logs across multiple agents and teams with a centralized web dashboard.
SKILL.md
Loading
Related Openclaw Skills
Howuku
A powerful integration for managing Howuku analytics data, user behavior insights, and conversion optimization tests through a unified CLI.
PRD to Prototype
An automated end-to-end workflow that transforms vague product ideas into professional PRD documents and interactive, high-fidelity HTML prototypes without requiring manual questioning.
Amazon Keyword Reverse Lookup Engine
A comprehensive keyword intelligence engine for Amazon sellers to uncover traffic-driving search terms, perform ASIN reverse lookups, and optimize CPC ad targeting.
Skillhub Preference
A policy-driven management skill that prioritizes skillhub for discovery and installation while maintaining clawhub as a reliable fallback.
NationBuilder Integration
A comprehensive NationBuilder integration for AI agents to manage political campaigns, supporter lists, and community engagement through the Membrane CLI.
Business Tools Pack
A comprehensive suite of seven automation skills designed to streamline operations for solo founders and digital product businesses.
Featured*
