for machines you own · build once · install anywhere, even offline
Build your whole stack once. Install it on every machine you own.
kldload is a build tool for your own servers and workstations. On one machine with internet, it gathers everything the others will need — the Linux you choose (Fedora, Debian, RHEL or Arch), the kernel, ZFS, graphics drivers, virtual machines, Kubernetes — checks that those versions actually work together, and packs it all into a single installer image.
Boot that image from a USB stick or over the network and you get a finished machine, the same every time — with no internet needed, because the full image carries its own package mirrors. When a machine breaks, you do not repair it by hand: you reinstall it from the same image and bring its data back from a snapshot.
Build.
One command, ./deploy.sh build, picks a kernel that ZFS is
proven to build against and freezes it, with every package the
machines will need, into one ISO. Or download one already built.
Install. Boot it from USB, or point your machines at it over the network and walk away. Each one installs, sets itself up and reboots ready to use.
Keep it. Every machine runs on ZFS: a snapshot before every update, rollback in one command, clones and replication built in. Recovering a machine is the same as installing one.
Good for a homelab that has outgrown one box, a small datacentre or edge site, an air-gapped or regulated network, or a lab full of identical machines.
Checksums, system requirements and what is on each ISO · sha256
Four editions
One installer, one result — ZFS on root with ZFSBootMenu, the kernel and its modules pinned as a set — and four payloads around it. Pick by how much you want done for you and whether there is a network.
The whole thing
A pushbutton golden image and a developer workstation.
Two complete offline mirrors (Debian and Fedora), the Kubernetes lab images, the local model runtime and every console. Installs with the network unplugged and comes up with virtualisation, an encrypted mesh, kernel tracing and a local AI stack already serving. Fully leverages the whole set of kernel modules and extensions — ZFS, NVIDIA, WireGuard, eBPF, Firecracker — built, signed and loaded against the shipped kernel.
For: a golden image you clone from, a workstation you develop on, an estate with no WAN.
Fedora workstation
The Fedora desktop, offline, with all the tools.
The Fedora mirror and nothing else on the payload side: the same installer, consoles, appliances and the NVIDIA and Secure Boot chain, without the Debian mirror, the Kubernetes images or the model runtime. Fedora installs with no network; the Kubernetes lab and the AI stack pull what they need when there is one.
For: a Fedora machine with ZFS on root and the kldload desktop, on a smaller stick.
Net installer
General build and install, over the network.
Every tool and every profile, no mirrors. Pick Fedora, Debian, the RHEL family or Arch and it installs from that distribution's own repositories, then brings up the same consoles and appliances the full image does. It needs a network for the whole install and takes longer; it fits on a 4 GB stick.
For: any distribution, any machine with a connection, the fastest download.
Core
Fedora, ZFS on root and ZFSBootMenu. Nothing else.
Like a minimal CentOS install, with ZFS. The installer, the pool layout, the signed boot chain and boot environments — and no consoles, no web UI, no appliances, no mirrors. Installs Fedora or the RHEL family over the network and hands you a plain system to build on your own way.
For: a server that wants ZFS and rollback and none of the rest.
| On the ISO | The whole thing | Fedora workstation | Net installer | Core |
|---|---|---|---|---|
| Download | 15 GB | 5.9 GB | 2.2 GB | 1.4 GB |
| USB stick | 32 GB | 8 GB | 4 GB | 2 GB |
| ZFS on root, ZFSBootMenu, boot environments | ✓ | ✓ | ✓ | ✓ |
| Kernel + modules pinned as one set, Secure Boot chain | ✓ | ✓ | ✓ | ✓ |
| Installs with no network | Debian, Fedora | Fedora | — | — |
| Distributions | Fedora, Debian, RHEL family, Arch | Fedora offline; the rest over the network | Fedora, Debian, RHEL family, Arch | Fedora, RHEL family |
| Consoles: zxplore, wgxplore, vmxplore, ztxplore, web console | ✓ | ✓ | ✓ | — |
| Thirteen appliances, Firecracker clones | ✓ | ✓ | ✓ | — |
| NVIDIA, WireGuard, eBPF, KVM on first boot | ✓ | ✓ | ✓ | — |
| Kubernetes lab images on the ISO | ✓ | from the network | from the network | — |
| Local model runtime (Ollama, Open WebUI) | ✓ | from the network | from the network | — |
| Recommended machine | The whole thing | Fedora workstation | Net installer | Core |
|---|---|---|---|---|
| CPU | 16 cores | 16 cores | 8 cores, 16 with the labs | 4 cores |
| RAM | 32 GB+ | 32 GB+ | 16 GB, 32 with the labs | 8 GB |
| Disk | 512 GB NVMe | 512 GB | 256 GB | 64 GB |
| GPU | NVIDIA for the local model | NVIDIA for the local model | optional | — |
| Boot | UEFI, Secure Boot supported. Legacy BIOS is not. | |||
Those are what the desktop and its labs want — KVM, Kubernetes, klab and the appliances run real VMs on ZFS clones. A server profile, or a VM with 8 GB, installs and runs the core of it. Minimums and the full table.
Just get it
Build your own, or use one of the four finished images.
Build your own
1. git clone https://github.com/kldload/kldload.git
2. PROFILE=desktop ./deploy.sh build
3. sudo ./deploy.sh burn /dev/sdX
4. Boot it. Done.
EDITION=net, EDITION=core or
DARKSITES=fedora build the other three; ./deploy.sh menu
is a checklist that shows the size before you commit to it.
Or use a finished image
1. Pick an edition and download it —
resumable, because a transfer that dies at 14 GB should not mean starting over.
2. Write it to a stick.
3. Boot it, pick a distribution and a profile.
4. Reboot into a working machine. Done.
vmxplore 0.5.0 — thirteen appliances, one click each
Web StackLAMPWriteFreely JellyfinPlexSeedbox IcecastSDR StationTvheadend AdGuard HomeSyncthing VDI DesktopRDP Desktop
Each is a real upstream project that boots as its own VM on its own
ZFS pool — verified by its own checks, enrolled on its own
WireGuard mesh with a certificate from the estate CA, in the Ansible
inventory and on the dashboards — and hands you a URL. Type a
pool name and the datasets are laid out for you; a Plex library gets
one dataset per title, so one film is one zfs send.
Shut a tile down and clone it as Firecracker microVMs: about
200 ms each, ten of them serving in half a minute. The
VDI desktop clones into a wall of
live desktops, each naming itself on screen.
Things you can go and do
Each of these is a page that walks the whole task.
Golden images, clones and replication cannot solve every problem. They solve a great many of them, and they turn most of the rest into something you can roll back.
Download and system requirements · Screenshots · Why I use kldload · The Bridge · Tutorials · Reference