Lobster Trap: OpenClaw in Containers from Local to K8s and Back — Sally Ann O'Malley, Red Hat
Summary
Sally Ann O'Malley, a software engineer at Red Hat with 10 years of experience in containers, Linux security, and Kubernetes, presents a practical demo of running OpenClaw (an open-source AI agent framework) inside containers using Podman locally and Kubernetes for team-scale deployment. The talk argues that containerization is the natural and correct way to deploy AI agents, applying the same principles that make containers valuable for any application: reproducibility, portability, secret isolation, and clean sandbox environments.
O'Malley introduces the concept of a "forever claw" — a persistent, personalized OpenClaw instance running as a long-lived container. Her own setup includes two sub-agents: Joy (a Jyotish astrology expert providing weekly readings) and Bruno (a Boston Bruins sports briefing agent). These personal agents serve as a demonstration vehicle for the broader architectural patterns she advocates.
The core technical stack she presents is Podman locally with Kubernetes or OpenShift for distribution, using the same container image throughout. Secret management gets two layers of isolation: Podman secrets store API keys on the host machine, while OpenClaw's built-in secret ref feature creates internal pointer references inside the container — ensuring API keys never appear in logs or environment variable dumps. For Kubernetes deployments, the same pattern applies with Kubernetes secrets mapped as secret refs to environment variables. Persistent state is handled through Podman volumes locally and PVCs (Persistent Volume Claims) in Kubernetes, enabling a clean backup and recovery story. O'Malley runs nightly automated backups of her forever claw using a systemd-equivalent service on Mac.
O'Malley built a custom local installer tool (available on GitHub) that wraps the Podman command with opinionated defaults: it accepts API keys via Podman secrets or Docker environment variables, creates OpenClaw secret refs automatically, supports provider selection (OpenRouter with Gemma as primary, Anthropic as fallback, with local endpoint options), and optionally sets up an OpenTelemetry collector with Jaeger for observability. The installer also supports SSH sandbox configuration, allowing OpenClaw to run all its commands in a remote SSH workspace.
The most compelling production case study comes from an Nvidia team, disclosed by O'Malley with permission from a colleague she met at PyTorchCon. The team runs model evaluations using OpenClaw, with approximately 10 engineers each running their own OpenClaw instance in Kubernetes. The result: work that previously required six people is now handled effectively by one engineer, freeing the rest to focus on creative and experimental work rather than tedious evaluation code.
O'Malley articulates a vision for standardized workplace agent deployment: a curated baseline OpenClaw container image per organization that includes company-approved MCP servers, authenticated integrations (e.g., Google Drive access), and team-specific skills — deployable to new hires on day one and then personalized by each individual. This replaces the current informal approach of sharing markdown files, config directories, and YAML and hoping colleagues can reproduce the setup. The container image becomes the team baseline, and the demo shows transitioning from a local Podman instance to a kind cluster and OpenShift with minimal configuration change — the same image, flipped flag.
Key Takeaways
- •Use Podman secrets (or Docker env vars) combined with OpenClaw's secret ref feature to create two-layer API key isolation — keys never appear in logs or plain environment variables.
- •The same container image can be used locally with Podman and promoted to Kubernetes or OpenShift by flipping a flag, giving teams a single artifact for all environments.
- •Store all OpenClaw runtime state in Podman volumes (or Kubernetes PVCs) to enable clean backup, recovery, and migration without coupling state to the container lifecycle.
- •An Nvidia team of 10 engineers runs model evals with each engineer operating their own OpenClaw in Kubernetes, achieving the throughput that previously required six engineers.
- •OpenClaw's SSH sandbox feature allows agents to execute all commands in a remote SSH workspace — useful for isolating agent actions from the local host.
- •A curated, company-approved OpenClaw container image — pre-loaded with MCP servers, auth, and team skills — can standardize new-hire onboarding and replace ad-hoc markdown/config sharing.
- •OpenRouter with Gemma is a viable primary provider for OpenClaw, with Anthropic as a fallback; local endpoints can also be added directly in the installer config.
- •A custom installer wrapping the Podman run command with opinionated defaults reduces OpenClaw spin-up to under 2 seconds and abstracts secret mapping and provider configuration.
- •Running AI agents in containers provides a natural sandbox — host file system and network access must be explicitly granted — improving security posture compared to native installs.
- •Mount an agent tools/skills/MCP server directory as a volume so that all capabilities are available at container startup without rebuilding the image.
Transcript
[music] >> Hey, um I'm Sally. I work at Red Hat. I've been there for about 10 years and uh the first 7 years, awesome, totally cool. I was working on containers and uh Linux security stuff and Kubernetes. Um in OpenShift. That's what I did for the first 7 years. And then uh about 3 years ago, well about 5 years ago, I moved to the emerging tech org and that was awesome, too, because now I'm not totally tied to a product. I get to just work on what I want. I get to just try out new things. Awesom...
Related Sessions
