What does a safe home for AI agents look like?

Prototype A learning project, not production software. Not beta, not supported, not security reviewed.

KubeClaw builds an IPv6-only Kubernetes cluster with strict network boundaries and outbound-only access, to find out what containing an agentic workload actually takes.

Agentic AI runs arbitrary code with tool access. The interesting question is not whether that needs a boundary, but which boundary actually holds. KubeClaw is an attempt to answer that by building one end to end – OpenTofu provisions the infrastructure, Ansible maintains it, and Cilium restricts egress to the services each workload genuinely needs.

Contain the blast radius

Run agent workloads in isolated Kubernetes namespaces with host hardening, private networking, and explicit Cilium egress policies.

Read the security model

IPv6-first infrastructure

Provision Hetzner nodes without public IPv4 addresses and use NAT64/DNS64 when IPv4-only services are unavoidable.

Read the architecture

Built to be taken apart

Every decision is written down, including the ones that turned out to be wrong. Read it, disagree with it, and build something better.

Read the project status

Take the ideas, not the cluster

The documentation is organized around the decisions behind the design, the commands that build it, and the checks that tell you it works. Use it to learn the shape of the problem -- then build a version you have reviewed yourself.