Kubernetes on Proxmox VE in 2026: Every Way to Run It, Compared

Proxmox VE has no Kubernetes of its own. Its feature pages describe KVM for virtual machines and LXC for containers, and its roadmap has no Kubernetes item (checked on proxmox.com, 23 October 2026). So every Kubernetes cluster on Proxmox is VMs that you or a tool build. If you want someone else to run it, Ankra creates the VMs on your hosts and operates the cluster and your apps afterwards, and Cloudfleet runs the control plane for you while you add Proxmox VMs as nodes. If you want to own everything, the open source routes are Talos Linux (with Sidero Omni for fleets), Cluster API Provider Proxmox, Terraform with k3s or kubeadm, and Kubespray. Rancher can manage the result, but it has no official way to build it on Proxmox.

Whichever you pick, the hard part is rarely the first cluster. It is load balancers, storage, upgrades, spreading nodes over hosts, and what happens when one of those hosts dies.

What Proxmox gives you, and what it does not

Proxmox gives you very good virtualisation. VMs, templates, software-defined networking, built-in Ceph storage and a solid backup tool for VMs are all there. What it does not give you is the glue a cloud provider ships for Kubernetes.

• No official cloud controller manager. On a cloud, this component turns a LoadBalancer Service into a real load balancer. The community Proxmox cloud controller manager by Serge Logvinov labels and cleans up nodes, but it has no load balancer controller. Most teams add MetalLB or kube-vip to hand out IP addresses, or put everything behind an ingress controller.
• No official storage driver. The community Proxmox CSI plugin creates persistent volumes as Proxmox disks attached to the VM, with snapshots and resizing. It needs your Proxmox nodes joined in a Proxmox cluster.
• Spreading is up to you. Three control plane VMs on the same physical host is one power cut away from no cluster at all.

Keep these in mind as you read the options. The useful question for each one is which of these jobs it does for you.

The options at a glance

Option	What it is	Creates the VMs	Distribution	Who runs the control plane	Licence	Best for
Ankra	Platform that builds, deploys to and operates clusters	Yes	kubeadm (Cilium) or k3s	VMs on your hosts, operated by Ankra	Commercial (CLI open source)	Teams that want the cluster and the apps on it handled
Cloudfleet	Hosted control plane, your Proxmox VMs as nodes	You do, with its Terraform provider	Certified upstream Kubernetes	Cloudfleet hosts it	Commercial	A hosted control plane over on-premises nodes
Talos Linux + Omni	Immutable Kubernetes OS plus a fleet manager	With the Omni Proxmox provider, otherwise you	Talos	You, with Omni handling upgrades	Talos MPL-2.0, Omni BSL with paid plans	A locked-down OS and fleet upgrades
Cluster API Provider Proxmox	Cluster API provider from IONOS	Yes	kubeadm	You, from a management cluster	Open source (Apache-2.0)	Teams already running Cluster API
Terraform + k3s or kubeadm	Provider for VMs, your own install on top	Yes, from your code	Your choice	You	Open source	Teams that want to own every line
Kubespray	Ansible playbooks for any Linux machines	No	Upstream Kubernetes	You	Open source (Apache-2.0)	Ansible shops with VMs already in place
Rancher	Self-hosted multi-cluster manager	Only with community drivers	RKE2, k3s or what you register	You (and you run Rancher too)	Open source (Apache-2.0), paid Rancher Prime	Organisations already standardised on Rancher

Ankra

Ankra is an AI-powered platform that builds or imports Kubernetes clusters on the cloud or servers you choose, deploys your applications from Git, and operates them afterwards. On Proxmox VE it talks to the Proxmox API with a token you create, then clones the cluster’s VMs from a cloud-init template. Nodes without a template get an Ubuntu 24.04 one automatically. It installs kubeadm with Cilium (the default) or k3s with Flannel, Calico or Cilium, and can put etcd on dedicated VMs. There is no cloud bill, because the hardware is yours.

It is good at the Proxmox-specific chores. When you save the credential, Ankra sets up what each node is missing, a management bridge and a private ankra network with NAT and DHCP. With host spread, control planes, etcd members and workers are balanced over two or more hosts, and every node is labelled with its host so your workloads can spread too. If the Proxmox API is not on the internet, Ankra reaches it and the VMs through an SSH jump host you attach to the credential. Traefik, cert-manager and a DNS subdomain come up by default, so an ingress hostname gets its record and certificate automatically. Upgrades go one node at a time, control plane first, with an etcd snapshot before the control plane moves. Cloud Cost prices the cluster from a rate card you can edit, since Proxmox has no list prices. Every detail is in the Proxmox VE guide.

Above the cluster, Ankra connects a GitHub repository, builds and scans your code in pipelines that run in your cluster, deploys on every push and gives pull requests their own preview environments. That suits teams that build apps with AI tools and want them live on hardware they already own. The Security Center ranks vulnerabilities by how likely they are to be exploited, and an AI assistant explains failing pods and drafts fixes that wait for your approval. Proxmox clusters can also join UpCloud clusters in one Cilium ClusterMesh.

Where it fits less well is a team that wants Proxmox itself to keep the VMs alive. Ankra does not enrol VMs in Proxmox HA groups or move them after a host failure. Resilience comes from Kubernetes replicas spread over hosts instead. There is also no cloud controller, so LoadBalancer Services stay pending unless you add MetalLB or kube-vip, and most apps go through the ingress stack. Pricing is per worker vCPU, control planes are free, and the first 30 worker vCPU are free on every plan (see pricing). Enterprise customers can self-host Ankra, including in air-gapped sites, which matters for a lot of on-premises Proxmox estates.

Best for teams that own Proxmox hosts and want the cluster, the deploys and the day-2 work handled by one platform, without hiring a platform team. Details are on Kubernetes on Proxmox VE.

Cloudfleet

Cloudfleet is a Berlin company that runs a managed Kubernetes control plane and lets you attach nodes from clouds and your own hardware. On Proxmox, Cloudfleet hosts the control plane and handles etcd, certificates and Kubernetes version upgrades. Your Proxmox VMs join as self-managed nodes.

It is good at taking the control plane off your hands entirely. Nothing of it runs on your hosts, and the VMs need no inbound internet access. Its Proxmox guide creates the VMs with the Cloudfleet Terraform provider, the Telmate Proxmox provider and cloud-init, on Ubuntu 24.04. The same cluster can mix those VMs with nodes on Hetzner, AWS or Google Cloud.

It is a weaker fit if you want the control plane on hardware you own, or want someone else to create and size the VMs. Those Terraform files are yours to run. Application delivery, GitOps, backups and image scanning are also left to your own tools. We wrote up the wider differences in Ankra vs Cloudfleet.

Best for teams that want a hosted control plane with upgrades handled, and are happy to manage their Proxmox VMs with Terraform.

Talos Linux with Sidero Omni

Talos Linux is an immutable operating system built only to run Kubernetes, made by Sidero Labs and open source under MPL-2.0. It has no SSH, no shell and no package manager, and every node is managed through an API. Omni is Sidero’s fleet manager for Talos, as SaaS or self-hosted.

It is good at security and repeatable upgrades. Since May 2026, Sidero’s Omni Proxmox provider can create the Talos VMs for you. You describe a machine class and Omni adds or removes VMs as you scale. It is still a young 0.x project, and it needs Proxmox VE 9 for its optional HA rules.

It is a weaker fit without Omni, or for a business on a budget. The Talos guide for Proxmox has you create each VM by hand in the Proxmox UI and run talosctl yourself. Omni’s licence allows free non-production and home lab use, and anything your business depends on needs a commercial plan. Storage, load balancers, ingress, monitoring, backups and app delivery are yours to add either way.

Best for teams that value a locked-down OS and fleet-wide upgrades, and have the skills to assemble the rest.

Cluster API Provider Proxmox

Cluster API Provider Proxmox, often called CAPMOX, is an Apache-2.0 provider maintained by IONOS. It lets the Kubernetes Cluster API create and replace Proxmox VMs, so you describe clusters as Kubernetes objects.

It is good at declarative lifecycle. Clusters, machine counts and versions live in YAML that you can keep in Git. Its usage guide ships kubeadm templates with Cilium or Calico, a MetalLB variant, and Flatcar as a node OS.

It is a weaker fit for a first cluster. You need a management cluster to run Cluster API itself, a VM template built with image-builder, a pool of IP addresses and an IP address manager, and a fixed control plane address served by kube-vip. After that, add-ons, backups and delivery are yours.

Best for platform teams that already run Cluster API elsewhere and want Proxmox to be one more provider.

Terraform with k3s or kubeadm

Many Proxmox clusters are built with Terraform or OpenTofu. The bpg provider (MPL-2.0) is the most used by recent downloads and targets Proxmox VE 9. The older Telmate provider still has more stars, but its last stable release was in 2023 and its 3.x line is still in release candidates. Terraform clones the VMs from a template, then cloud-init or Ansible installs k3s or kubeadm.

It is good at control and at being yours. Every VM, disk and address is in code you can review, and you pick the distribution, the CNI and the OS.

It is a weaker fit for keeping it running. Terraform creates machines, it does not upgrade Kubernetes. Upgrades, certificate rotation, etcd backups, node replacement, load balancers and storage are runbooks you write and test yourself.

Best for infrastructure teams who already manage everything in Terraform and want no vendor in the path.

Kubespray

Kubespray is a Kubernetes SIGs project of Ansible playbooks, licensed Apache-2.0, that installs a production-shaped cluster on machines you already have. Proxmox is not one of its named platforms, but it does not need to be. It treats your VMs like any Linux servers.

It is good at flexibility and at being plain upstream Kubernetes. It has a long list of options for networking, container runtime and add-ons, and it upgrades a cluster the same way it installs one.

It is a weaker fit if you want VMs created for you, since that part stays with you. You also own the Ansible inventory and every variable in it, and backups, load balancers and delivery are yours to add.

Best for Ansible shops that create VMs another way and want a well-known installer.

Rancher

Rancher, from SUSE, is a self-hosted multi-cluster manager, open source under Apache-2.0 with a paid Rancher Prime edition. On Proxmox, most teams build clusters another way and register them in Rancher.

It is good at managing many clusters from one place, with RBAC and sign-in. Registered k3s and RKE2 clusters can be upgraded from Rancher.

It is a weaker fit for building clusters on Proxmox. Rancher’s built-in node drivers cover clouds and vSphere, not Proxmox, and the community Proxmox drivers are mostly old. Rancher Turtles, SUSE’s Cluster API extension, does not list Proxmox among its certified providers. You also run Rancher itself. SUSE’s own answer for on-premises virtualisation is SUSE Virtualization (formerly Harvester), which replaces Proxmox rather than running on it.

Best for organisations that already run Rancher and want their Proxmox clusters in the same console.

Moving from VMware or a cloud to Proxmox

A lot of teams arriving on Proxmox in 2026 are leaving VMware or a hyperscaler. Your manifests and Helm charts mostly move as they are. The parts that change are the ones a platform gave you.

• Load balancers. Services of type LoadBalancer need MetalLB, kube-vip or an external load balancer. Cloud annotations on those Services do nothing.
• Storage classes. vSphere, EBS or Azure disk classes become the Proxmox CSI plugin, Ceph through Rook, or local volumes. Check which workloads need shared read-write storage.
• Identity. Cloud workload identities have no Proxmox equivalent, so apps that called cloud APIs with them need their own credentials in Kubernetes Secrets.
• Spreading over hosts. Plan where control planes and replicas land so one host failing does not take down the cluster or an app.

Velero can back up workloads and volumes from the old cluster and restore them on the new one. With Ankra, you can import the existing cluster and clone it onto a new Proxmox VE cluster in one step. Provider pieces such as the cloud controller are skipped, stacks that held encrypted values ask you to re-enter them, and you review any value that still pins the old storage class or load balancer.

How to choose

• If you want the cluster, deployments, security scanning and day-2 operations from one platform on hosts you own, pick Ankra.
• If you want a hosted control plane and are happy to create the VMs with Terraform, pick Cloudfleet.
• If OS hardening and fleet upgrades matter most and you can build the rest, pick Talos with Omni.
• If you already run Cluster API, add Cluster API Provider Proxmox.
• If your team lives in Terraform and wants to own every layer, use the bpg provider with k3s or kubeadm.
• If you are an Ansible shop with VMs already in place, use Kubespray.
• If Rancher is already your standard, register your Proxmox clusters in it.

To try Ankra before connecting your Proxmox hosts, start a playground cluster.

Frequently asked questions

What is the best way to run Kubernetes on Proxmox?

Proxmox VE has no built-in Kubernetes, so you run it in VMs. For a managed option, Ankra creates the VMs on your hosts and operates the cluster, while Cloudfleet hosts the control plane and you add Proxmox VMs as nodes. For an open source route, Talos with Omni, Cluster API Provider Proxmox and Terraform with k3s are the common choices.

Should I use k3s, kubeadm or Talos on Proxmox?

k3s is the lightest and easiest to install, kubeadm is plain upstream Kubernetes, and Talos replaces the whole OS with one built only for Kubernetes. Talos gives the strongest lock-down but needs its own tools. Ankra installs kubeadm with Cilium by default or k3s, and Cluster API Provider Proxmox uses kubeadm.

How do I get LoadBalancer Services working on Proxmox?

Proxmox has no official cloud controller manager, so a LoadBalancer Service stays pending until something hands out an IP address. MetalLB and kube-vip are the usual choices on a local network. Many teams skip it for web apps and route traffic through an ingress controller instead.

Can Kubernetes on Proxmox survive a host failure?

Yes, if the VMs are spread over at least three hosts. Run three control planes on different hosts and at least two replicas of each app, spread by host. Ankra’s host spread places VMs this way and labels every node with its host, and Proxmox’s own HA can restart VMs on another host when you manage the VMs yourself.