denoland/clawpatrol

What it solves

It provides a security firewall specifically designed for AI agents. It prevents agents from performing dangerous or unauthorized actions in production environments by intercepting and gating their network traffic before it reaches the target system.

How it works

Claw Patrol acts as a proxy (gateway) that parses network traffic at the wire level. It uses rules written in HCL (HashiCorp Configuration Language) and CEL expressions to evaluate whether an action should be allowed, denied, or paused for human approval. It can be deployed as a standalone gateway, a process-specific tunnel (using netns on Linux or NetworkExtension on macOS), or a host-wide tunnel via WireGuard.

Who it’s for

Developers and organizations deploying AI agents that need to interact with production infrastructure like Kubernetes, SQL databases (Postgres, ClickHouse), and HTTP APIs, where strict security controls and security policies are required.

Highlights

• Wire-level traffic parsing for protocols like SQL and Kubernetes.
• Rule-based gating using HCL and CEL expressions.
• Flexible deployment options including per-process tunnels and host-wide routing.
• Support for human-in-the-loop approval for sensitive actions.