A wire-level bouncer that stops agents from breaking prod

What it does

Claw Patrol is an inline security gateway that sits between autonomous agents and production infrastructure. It parses traffic at the wire level using protocol-specific inspectors—SQL verbs and table names for Postgres and ClickHouse, resource types and namespaces for Kubernetes, methods and paths for HTTP—and gates each request against HCL rules written as CEL expressions. You can block destructive queries outright or pause operations like kubectl delete pod until a human approves them.

The interesting bit

Instead of forcing a full network rearchitecture, it can wrap a single process tree in a per-process sandbox on Linux via network namespaces or macOS via NetworkExtension. Only the wrapped agent’s traffic transits the gateway, which makes it practical for ad-hoc tooling rather than just fleet-wide rollouts.

Key highlights

• Protocol-aware rules: CEL expressions evaluated against wire facts, not just Layer 3 metadata
• Supports deep inspection for Postgres, ClickHouse, Kubernetes, and HTTP traffic
• Can operate as a central proxy, a host-wide tunnel, or a per-process network sandbox
• Single Go binary that accepts clients over WireGuard or Tailscale
• MIT licensed, maintained by the Deno team

Caveats

• Linux and macOS only; Windows support is not mentioned in the README
• The human-in-the-loop approval flow is referenced but not explained in the repository
• Detailed configuration docs live entirely on an external site

Verdict

Worth a look if you are wrapping Claude or other agents and pointing them at production databases or Kubernetes. Skip it if your tooling is already covered by existing authz layers and you do not need wire-level visibility.