kcore: virtualisation built for automation, not for a web console

The hypervisor Evaluation ISO available

A controller holds desired state; node agents run Cloud Hypervisor on each NixOS host; operators work through kctl or the same gRPC APIs. Add a node by booting the ISO and pointing it at the controller.

Host OS	NixOS, read-only root. Each update is one transaction; a failed update boots the previous generation. Every generation is kept.
API	gRPC with mTLS, certificate-CN authorisation per method. kctl apply -f for declarative YAML: VMs, networks, SSH keys, containers, security groups, disk layouts, Ceph clusters, cluster updates.
Guests	Linux and Windows (Server and desktop). Cloud-init. One physical GPU per VM via VFIO passthrough; NVIDIA vGPU is not supported.
Storage	Local directories, LVM or ZFS on the host. Ceph for shared disks; live migration between hosts requires Ceph. Snapshots, clones and rollback on every backend.
Networking	NAT, VXLAN and bridge networks. Security groups with ingress rules and DNAT, attached to VMs or networks.
Clustering	Multi-node, with control-plane replication across datacenters on a CRDT-based state layer. Behaviour under partitions and node failure is specified in TLA+ and exercised by failure-injection tests.
Security	LUKS disk encryption on hosts, TPM2-sealed where hardware allows. Cluster PKI with a root CA that never leaves the operator's machine, sub-CA for automatic node certificate renewal.
Tooling	Terraform provider (kcore_vm and friends), MCP server for AI agents, Crossplane provider, kcore-migrate for VMware and Proxmox imports.
Pricing	Per CPU socket per year. Community ISO free on GitHub Releases. Standard, Premium and Enterprise tiers add signed ISOs, the enterprise channel and support.
Not yet	Integrated backup (see kcore Vault below), vSphere-style HA, direct disk pull from vCenter without Broadcom's VDDK.

Off-cluster backup Private channel · in development

Snapshots inside a cluster protect against operator mistakes and bad upgrades. They do not protect against losing the Ceph pool, the node holding an LVM or ZFS volume, or an attacker with cluster-admin. kcore Vault moves guest data out of the cluster.

Model	A Vault coordinator exposes a gRPC service (kcore.vault.Vault). It drives the open-core snapshot and export RPCs on the cluster: expose a read-only snapshot device, list changed extents between two snapshots, stream blocks out, stream blocks back in on restore.
Incremental	Changed-block tracking between snapshots means only extents that changed since the last run are read and transferred, on Ceph RBD, ZFS and LVM-thin volumes alike.
Consistency	Crash-consistent by default; quiesced snapshots freeze guest filesystems over SSH first.
Trust boundary	Vault clients authenticate with certificates from the Vault CA, not the cluster CA. The coordinator reaches the cluster with a scoped backup-operator identity. A compromised cluster has no credential that reads its backups.
Open hooks	The cluster-side RPCs are part of open-core kcore, so any backup tool can drive them. Vault is the supported, paid implementation on top.
Today	In Community kcore, kctl backup copies controller state and SnapshotPolicy schedules in-cluster snapshots. The Vault coordinator exists; the storage engine, mover, dashboard and CLI follow. Immutable backup targets are planned.

Bare-metal provisioning Available

Installing one host from a USB stick is fine. Installing forty across three sites is not. kcore Provisioning turns racked servers into kcore nodes without anyone touching them.

Hardware control	Discovers servers and their inventory over Redfish and IPMI, sets boot order, and controls power. Tested against Dell PowerEdge hardware, including the quirks of older iDRAC and iPXE firmware.
Network boot	UEFI HTTP boot server with dynamic GRUB configuration per machine, serving the EFI bootloader and image artefacts. No TFTP, no PXE legacy stack.
Images	Image registry with SHA-256 verification on download and on serve; corrupted artefacts are detected and removed. Size limits and rate limiting protect the server.
Machines	Machine registration by MAC and BMC address, boot-target assignment, and status tracking in SQLite.
PKI	Issues host certificates so a freshly installed node joins the cluster with mTLS from first boot.
Security	mTLS client authentication with bearer-token fallback, role-based authorisation, input validation on every identifier, path-traversal and command-injection protection, and an audit log of every operation.
Interfaces	gRPC API, kctl integration, Terraform, and a read-only web dashboard (Leptos, server-rendered) for monitoring.
Implementation	Rust. Controller and HTTP boot server as separate services; packaged as Nix flakes for reproducible deployment.

• kcore-migrate

  Inventories a vCenter or Proxmox estate and writes one kcore_vm Terraform resource per guest, plus a gaps report for anything the record cannot express. Converts exported OVA, OVF, VMX and VMDK with virt-v2v, or changes container format with qemu-img. Written in Go.

• Terraform provider

  A real provider built on the Terraform Plugin SDK, talking to the controller over gRPC. VMs are resources; terraform apply does the rest.

• MCP server

  Lets an AI agent operate a cluster through the same declarative surface as kctl apply, with a discipline built in: advise on the questions an operator still has to answer, plan, then apply only with explicit confirmation. Ships as a Cursor plugin and a VS Code extension.

• Crossplane provider

  Reconciles Kubernetes custom resources (VMs, networks, SSH keys, security groups, workloads) against the controller API, so a kcore estate can be managed from Argo CD like everything else.