Podclaw

Podclaw is a homelab-friendly playground for running OpenClaw on rootless Podman inside Incus containers.

The goal is to give AI agents (Claude + "Ralph") a safe, disposable environment to iterate on OpenClaw deployments without putting real infrastructure or data at risk.

Design goals

• Sacrificial host only All experiments run on a dedicated box (k8s-delta) with no important data and no shared host path mounts.

• Rootless by default OpenClaw runs under a non-privileged user in rootless Podman, using Incus containers as the outer boundary.

• Agent-friendly, human-auditable Cloud-init, Incus profiles, and scripts are small, idempotent, and stored in this repo so both humans and AI agents can reason about them.

• Easy nuke-and-rebuild If (when) things go sideways, you can delete experiment containers or rebuild the entire host from scratch with minimal ceremony.

Components

• profiles/ Incus profile definitions for the sacrificial host:

  • ralph-bridged.yml -- bridged NIC, no host mounts.
  • ralph-nesting.yml -- enables nested Podman/Docker where needed.

• cloud-init/ Cloud-init configs for:

  • ralph-sandbox.yml -- Claude Code + tools sandbox on the victim host.
  • openclaw-podman-skeleton.yml -- base for OpenClaw + rootless Podman experiment containers.

• scripts/ Helper scripts for:

  • Configuring a fine-grained GitHub token for this repo only.
  • Launching and cleaning up experiment containers.
  • (Later) starting OpenClaw via Podman or systemd user services.

• NOTES.md Threat model, safety assumptions, and "what I'm willing to lose" written down explicitly.

High-level workflow

1. Prepare the victim host

   • Choose a sacrificial machine (k8s-delta) running Ubuntu 24.04 LTS.
   • Remove old Incus containers and any host path mounts.
   • Create minimal Incus profiles from profiles/.

2. Create the Ralph sandbox

   • Launch ralph-sandbox on the victim with cloud-init/ralph-sandbox.yml.
   • Install Claude Code CLI and the "Ralph" skills in that container.
   • Configure a fine-grained GitHub token that can only access this repo.

3. Iterate on OpenClaw experiments

   • Ralph/Claude edits cloud-init/openclaw-podman-skeleton.yml and helper scripts in this repo.
   • From ralph-sandbox, launch experiment containers with Incus using that cloud-init.
   • Inside those containers, run OpenClaw under rootless Podman and refine until it's boring and reliable.

4. Reset when necessary

   • Delete experiment containers or rebuild the entire victim host when the lab gets messy.
   • Keep k8s-alpha and any real workloads out of this blast radius.

Security notes (non-exhaustive)

• Podclaw assumes:
  • The sacrificial host has no important secrets or data.
  • There are no host path mounts from Incus containers into sensitive directories.
  • The Ralph sandbox user has no SSH keys to other machines.
  • OpenClaw gateways created here are initially bound to loopback; any LAN exposure is deliberate and proxied.

For more detailed security assumptions and boundaries, see NOTES.md.

Status

This repo starts as a lab notebook and harness, not a polished product. Expect rough edges, breaking changes, and plenty of "Ralph Wiggum" experiments as we converge on a clean, repeatable pattern for OpenClaw on rootless Podman in Incus.

Hardware

• k8s-delta (sacrificial lab host): GMKtec M6, 64 GB RAM, 2 TB NVMe, Ubuntu 24.04 LTS
• k8s-alpha (real homelab, out of scope): Minisforum HX370, 96 GB RAM, 2 TB NVMe