Claw Patrol

Claw Patrol is an open-source agent proxy from the team at Deno. It sits between AI agents and production systems, parses their traffic at the wire, and evaluates each action against rules written in HCL. It also injects credentials on the agent's behalf and keeps an audit log of every action.

What It Is

Claw Patrol is a security firewall for agents. Rather than only filtering model output, it inspects what agents actually do downstream: HTTP requests, Postgres and ClickHouse queries, and Kubernetes API calls. Rules are CEL conditions over protocol-level facts such as HTTP method, path, headers and body, SQL verbs, tables and functions, and Kubernetes resource, verb and namespace. A rule can allow, deny, or route a request to an approver.

How the workflow runs

Agents are wrapped by prefixing their command, for example clawpatrol run codex or clawpatrol run claude. The project documents three deployment shapes: clawpatrol gateway runs the proxy from a single binary loading an HCL config, clawpatrol join tunnels a whole host to a gateway over WireGuard, and clawpatrol run opens a per-process tunnel on Linux (netns) or macOS (NetworkExtension). Gateways accept clients over WireGuard or Tailscale and are self-hosted.

Approval flows and testing

Ambiguous requests can be deferred to an LLM judge with a custom prompt, with verdicts cached, or to a human who votes in Slack, the dashboard, or a custom webhook; human approvals time out closed. Real actions can be recorded from the dashboard as JSON fixtures, and clawpatrol test replays them against the rule engine in CI, printing a diff and failing the build when a policy change flips a verdict.

Open-source model

The project is MIT licensed, and the site states the proxy holds secrets and sees all agent traffic, so it needs to be auditable. A demo admin dashboard is available for browsing captured requests. The project positions itself as complementary to sandboxes, which it says can be stacked with it.