In today's dynamic infrastructure environment, the need for fast, repeatable and customizable VM deployments is essential. Many teams already know this pattern from other virtualization platforms — building VM images with Packer and publishing them through content libraries to speed up provisioning while keeping their existing VM management practices intact.

The question is: what does that same discipline look like once you move onto KubeVirt based platform in my case Harvester?

In this talk I'll show how we use Packer with the new (pre-release) packer-plugin-kubevirt to build golden VM images directly inside a Harvester HCI cluster — no separate virtualization tooling needed. It's the same "build once, deploy many" philosophy, just running natively on Kubernetes.

This approach is largely hypervisor-agnostic because it separates concerns cleanly into layers: the OS layer (built once via Packer + Kickstart, patched centrally) and the app layer (deployed on top, independent of the underlying hypervisor). Once you adopt that separation, the same golden image discipline applies whether you're running on any KubeVirt-based platform, or eventually migrating between them.

I'll walk through a real working pipeline: starting from an ISO, automating the install with Kickstart, producing a reusable image on Harvester and adding that into a CI pipeline so that every patch cycle just becomes a pipeline run. I'll then take it a step further and show how we've started exposing that same Harvester/KubeVirt infrastructure to an MCP server, so an agent can interact with the cluster conversationally.

Building the golden image once and reusing it as the foundation for the migrated workloads, instead of doing a one-off lift-and-shift for each VM.

For organizations with large existing VM footprints, this gives a path to combine modern, pipeline-driven deployment with the operational practices teams already rely on.

This is based on hands-on work and customer environments running kube-virt based virtualisation platform and I'll share what worked, what didn't and the rough edges of using a pre-release plugin in a real pipeline.