An agent runtime
Agents need somewhere to live that is not your main session, and not someone else's container.
2
personas, the workstation and the headless node, built from one codebase.
100%
of your prompts, files and keys stay on your hardware by default, not opt-in.
201MB
idle memory on the headless node. Measured, not a target we hope to hit.
Most agent work is grunt work: summarizing, refactoring, extracting, drafting, testing. A local model on hardware you already own handles it at zero marginal cost, offline, in private. The hard fraction (deep reasoning, big context, the final call) goes to a frontier model over the API.
vinOS wires that split in at the runtime layer, so it holds whether
you are one person at a desk or thirty workers on a rack.
Ollama serves local models on the metal,
Claude Code stands ready for the calls that earn their
price, and vinos-mcp governs which tools an agent can
reach. The API bill covers judgment, not volume.
The distro underneath differs because the job differs. The vinOS layer
on top (bash, systemd units, config drop-ins, the vinos
CLI) is byte-identical on both.
vinos-dev
The workstation you sit in front of.
linux and linux-t2pacman -Syushipping, v1.4.0
vinos-vm
The headless node you never log into.
linux-image-genericin development
Why two distros and not one? Because Arch on a cloud VM has an adoption ceiling somewhere under one percent, and Ubuntu on a developer's ThinkPad is a compromise nobody asked for. Rather than force one base to be bad at half its job, vinOS keeps the base honest and moves the portable part, the runtime, above it.
That is the whole architectural bet. A desktop distro cannot become a fleet runtime later; a runtime can grow a desktop. We built the runtime first.
Not a fork of anything. A runtime layer over an unmodified base: plain bash, plain systemd units, plain config files. Read all of it in an afternoon, walk away and nothing follows you.
┌───────────────────────────────────────────────────────────────┐ │ your agents # local · frontier · MCP tools │ │ ollama · claude code · vinos-mcp registry │ ├───────────────────────────────────────────────────────────────┤ │ vinos runtime # identical on both personas │ │ bin/vinos-* · systemd units · /etc/vinos/* │ ├──────────────────────────────┬────────────────────────────────┤ │ vinos-dev # workstation │ vinos-vm # headless node │ │ hyprland · waybar · walker │ cloud-init · agent worker │ │ foot · mako · greetd │ no session, no compositor │ ├──────────────────────────────┼────────────────────────────────┤ │ arch linux # rolling │ ubuntu 24.04 # LTS │ │ linux · linux-t2 (Apple T2) │ linux-image-generic │ └──────────────────────────────┴────────────────────────────────┘
One ISO, any x86_64 machine. The laptop on your desk, the workstation your OEM shipped preloaded, the box in the basement that has been idle since 2021. Every one of them is a node that costs nothing per hour.
16 GB of RAM and a decade-old CPU run a 7B model comfortably. A GPU or a fat server runs 30B. Drivers are detected and services tuned, because a node you have to hand-configure is a node you will never add a second of.
Including the Intel MacBooks Apple walked away from.
vinOS ships the linux-t2 kernel in the ISO and installs
it to disk on Apple hardware automatically. Keyboard, trackpad, Wi-Fi
and Touch Bar work from the live USB through to the installed system.
Known T2 limits: suspend is unreliable upstream (the workarounds are documented), and Touch ID does not work under Linux. Everything else does.
Nothing tricky. If you have flashed a USB before, you can install this.
Grab vinos-1.4.0-x86_64.iso from the
releases page.
About 4.0 GB, with a published checksum. Verify before you flash.
Etcher, Rufus, or dd. Any 8 GB stick works. Five minutes.
Boot the stick. When the desktop loads, open a terminal and run:
sudo vinos-install-diskThe base stays small. Each bundle is a plain list of packages at
/usr/share/vinos/bundles/, installed by one command.
ai
ollama · claude-code · open-webui · llm
dev
neovim · lazygit · docker · gh · nodejs · rustup · uv · direnv
media
mpv · imv · yt-dlp · ffmpeg · spotify · pinta
office
libreoffice-fresh · zathura · obsidian · localsend
gaming
steam · lutris · gamemode · mangohud · protontricks
productivity
obsidian · dbeaver · postman · flameshot · syncthing
comms
signal-desktop · slack · zoom · discord
browser
chromium · firefox · brave-bin
Install one with vinos-install ai. Nothing here is
preinstalled, and nothing here is required.
Every binding is wired in the default Hyprland config and works on first boot. Press Super+K to see it inside the OS.
Same kernel, same repos, same packages you could install yourself. The difference is what is already done, and whether that work carries over to the machine you never log into.
| vinOS | Arch (stock) | DIY Hyprland | |
|---|---|---|---|
| Live installer ISO | yes | yes | no |
| Hyprland pre-configured | yes | no | bring your own |
| T2, NVIDIA, ThinkPad drivers | baked in | manual | manual |
| Agent stack in one command | yes | no | no |
| First-boot user creation | TUI installer | archinstall | manual |
| Live theme switcher | 5 themes | no | no |
| Opt-in app bundles | 8 bundles | no | no |
| Sudo and security defaults | hardened | defaults | defaults |
| Time to a working desktop | about 15 min | about 45 min | a weekend |
| Same layer runs headless | one codebase | rebuild it | rebuild it |
| MCP tool registry | vinos-mcp | no | no |
Flash it to a USB, boot any x86_64 machine (including the T2 MacBook in your drawer) and you have a node running local models in about fifteen minutes. Add the second one when the first has earned it.