GuardClaw
GuardClaw is a deterministic runtime security layer for AI agents that blocks risky actions before they execute, without using an LLM to make those decisions.
About
GuardClaw is a deterministic runtime security layer for AI agents that blocks risky actions before they execute, without using an LLM to make those decisions. It sits between your agent and its tools, applying seven independent defense layers — including threat intelligence, input validation, policy enforcement, capability tokens, sandboxed execution, human-in-the-loop approvals, and cryptographic audit receipts. The key design choice is that security decisions are made by policy rules rather than probabilistic AI inference, which eliminates the irony of using another AI to police the first one. It runs locally so your prompts and data never leave your infrastructure.
Developers deploying OpenClaw or other MCP-compatible agents in environments where the agent has write access to real systems — production databases, email, code repositories, or financial tools. Essential for any team that needs a compliance audit trail or wants to prevent prompt injection attacks from causing irreversible damage. <!-- Screenshot pending -->
Pros & Cons
Pros
- check Deterministic rule-based enforcement means security behavior is predictable and auditable, unlike LLM-based safety layers
- check 1,743 built-in detection patterns cover prompt injection, data exfiltration, tool abuse, and known CVEs out of the box
- check Local-first execution keeps sensitive prompt content and tool calls entirely within your own infrastructure
- check Human-in-the-loop approval layer lets you gate high-risk actions — useful for agents with write access to production systems
- check Tamper-evident audit logs provide a compliance-ready record of every agent action and security decision
Cons
- close Free tier caps at 1,000 requests per day and 5 custom policies, which may be insufficient for active production agents
- close Proprietary software — you cannot inspect or modify the detection logic, which creates a trust dependency for a security-critical component
- close Adds latency to every tool call; in high-frequency agent workflows this overhead compounds
- close Currently requires CLI installation via Homebrew or PowerShell, which is not suitable for fully automated deployment pipelines without extra scripting
More Security
Other tools in the same category.
