v0.3.0

Release 0.3.0

Ubuntu 24.04 is now a first-class guest: a three-node k3s cluster boots
on bare VMs, takes Cilium, Tetragon, and Longhorn, and still warm-forks.
A gym leases those nodes, a tree jailer contains them, and an MCP demo
plus live viewer show a cluster-fork incident end to end.

v0.2.2

Release 0.2.2

Inner-cluster fidelity and guest-kernel work. Add-ons survive a warm fork,
Cilium can replace kube-proxy inside the guest cluster, and the guest grows
iSCSI/NFS plus eBPF plumbing.

v0.2.1

Release 0.2.1

Patch release. k7d --version on a machine with a running daemon no longer
takes down that daemon's control socket.

v0.2.0

[0.2.0] — 2026-08-16

Second release. Bigger guests boot, container writes land on a real disk
instead of guest RAM, a restored guest knows what time it is, and an idle
vCPU costs ~1% of a core rather than 100%.

Added

• Writable storage on a host disk (k7d.katakate.org/scratch-size). A
  container's overlay upper now lives on a sparse ext4 image attached as a
  read-write virtio-blk device (default 8 GiB) instead of a RAM-backed
  layer. Writable capacity is no longer tied to the pod's memory limit, and
  filling it gives the writer ENOSPC instead of guest-wide memory
  pressure. A warm fork gets a reflink-CoW copy, isolated from parent and
  siblings; the image is unlinked with the VM.
• Per-container ephemeral-storage limits
  (k7d.katakate.org/container-scratch-size). A name=size list over
  Kubernetes container names, e.g. builder=2Gi,logger=64Mi, gives each
  named container its own image so an overrun stays that container's
  problem. Opt-in: each image spends one of the VM's virtio-MMIO device
  slots, and a pod using it is served by fork rather than the warm pool.
• Guest ephemeral-storage usage accounting over the agent protocol.
• packaging/install.sh for installing the daemon, shim, and guest
  artifacts from a release tarball.

Fixed

• Guests of 4 GiB and larger now boot. Guest RAM is split across the
  32-bit MMIO hole instead of overlapping the device window.
• A restored guest knows the wall clock. The guest is given kvmclock and
  the host a way to re-base it, so a VM resumed from a snapshot hours later
  no longer reports a date from 1999 — which broke TLS certificate
  validation for anything the guest fetched.
• An idle vCPU no longer costs a full host core. The guest uses the
  haltpoll cpuidle governor with a 200 µs window, bringing an idle vCPU to
  roughly 1% of a core.
• The daemon alone chooses which guest kernel and initramfs a VM boots, and
  refuses to create a VM without them rather than guessing.
• OCI cgroup mounts are translated to guest cgroup2 with Kata-like
  namespace nesting, and guest IPv6 is enabled.
• Pause/resume, fork-source adoption, teardown, and stats fixes found while
  integrating the k7 backend.

Install

tar xzf k7d-v0.2.0-x86_64-linux.tar.gz
cd k7d-v0.2.0-x86_64-linux
sudo ./install.sh

k7d v0.1.0

First release of k7d — a purpose-built microVM runtime for Kubernetes sandboxes: sub-second VM boot, CoW memory forks, snapshot trees, and a containerd shim (runtimeClassName: k7).

Install

curl -LO https://github.com/Katakate/k7d/releases/download/v0.1.0/k7d-v0.1.0-x86_64-linux.tar.gz
tar xzf k7d-v0.1.0-x86_64-linux.tar.gz && cd k7d-v0.1.0-x86_64-linux
sudo ./install.sh            # add --with-k3s to register the RuntimeClass

Requires Linux x86_64 with KVM (/dev/kvm). The tarball ships k7d, containerd-shim-k7-v1, the guest kernel, initramfs, and install.sh.

For the full sandbox platform (Kubernetes orchestration, CLI/API/SDK), see Katakate/k7 and the docs at docs.katakate.org.