ClawPatrol

Why it's ranked here

The project addresses agent access at the network boundary with unusually concrete controls. Rules can reason about SQL operations, Kubernetes resources, and HTTP request details instead of merely matching destinations. Its three deployment shapes make the same policy model useful for local workers and shared infrastructure, though setup complexity is substantial.

What's good

Protocol-aware facts make policies precise. A rule can distinguish SQL verbs and tables, Kubernetes verbs and namespaces, or HTTP methods and paths. WireGuard works without a separate gateway kernel module, while Tailscale mode avoids public inbound ports and adds user identity. Human approval can stop a request before it reaches the target.

Tradeoffs

This is infrastructure, not a lightweight agent library. The release-shaped Go binary is roughly 59 MB before the real dashboard bundle. Building requires the dashboard toolchain, and exercising per-process interception on macOS requires a signed system extension. WireGuard deployments also face NAT hairpin limits, while Tailscale mode depends on its hosted control plane.

How to use it well

Use it when platform engineers need enforceable controls around agents that access databases, Kubernetes, or HTTP services. Start with the annotated HCL example, test deny and approval paths, then choose per-process routing for isolated workers or host-wide routing for managed machines. It does not govern actions that never traverse the gateway, so retain host and application controls.

Observed

License

MIT

Primary language

Go

Packaging

Single statically linked Go binary with an embedded dashboard SPA

Install surface

Shell installer or source build with Make

Interface

CLI with gateway, join, and per-process run modes

Platform support

Per-process tunnels on Linux and macOS

Configuration

HCL policies with CEL conditions

Network modes

WireGuard and Tailscale