Deno has released Claw Patrol, an open-source firewall that intercepts AI agent traffic at the protocol level before it reaches production databases, Kubernetes clusters, or GitHub repositories.
The company has been deploying AI agents against live infrastructure to triage PagerDuty alerts and push fixes to Deno Deploy. That means agents with real credentials touching real Postgres, GCP, ClickHouse, and Kubernetes — a setup Deno describes as "dangerous to say the least." Claw Patrol sits in the path of those connections via WireGuard or Tailscale, terminating TCP sessions and parsing application protocols — HTTP, Postgres, SSH — to apply allow/deny rules. Destructive actions are routed to a secondary LLM for review, then held for human approval before they execute.
The interesting design choice is where it intercepts traffic: at the network and protocol layer, not the AI layer. Most agent safety tooling works one level up — system prompts, LLM gateways, MCP proxies — and none of those handle a Postgres connection tunneled through a Kubernetes port-forward or a raw SSH session. By parsing application protocols directly, Claw Patrol can enforce rules that model-layer guardrails simply cannot reach, while keeping the safety layer decoupled from the model itself.
Written in Go, configured in HCL, and MIT licensed, Claw Patrol is early — built to scratch Deno's own itch — but that itch is about to get a lot more common.