SynapseOS

What is SynapseOS?

Most “AI operating systems” are a chat window on a normal desktop. SynapseOS starts one layer down. A local inference daemon called synapd owns the model and serves every other component over the SYN socket protocol, so the AI is a system service in the same sense that systemd-resolved is — something the rest of the OS can depend on rather than an app you open.

The desktop is synui, a wlroots Wayland compositor written for this system rather than adapted to it — one that knows the AI daemon exists and holds a live subscription to the security monitor’s verdict feed.

• Local inference, no accounts

  The installer downloads one model — Mistral 7B Instruct (Q4_K_M, ~4.1 GB) is the recommendation, and you can decline — and runs it through llama.cpp. Nothing to configure afterwards, and nothing leaves the machine.

• A shell that takes English

  synsh is a normal POSIX-ish shell until you stop typing commands. Describe the outcome instead and it resolves the intent against the local model — then shows you what it is about to run.

• A compositor of its own

  synui runs on wlroots 0.20 and renders through scenefx 0.5: tiling and monocle layouts, per-output workspaces, XWayland, layer-shell, glass, blur, shadows, and an optional CRT post-process pass.

• Security that enforces in-kernel

  synguard classifies syscall events, scores threats and publishes verdicts on a feed the desktop subscribes to. Since 0.2.4 it can also deny in-kernel through a BPF-LSM gate, behind warmup, heartbeat and deny-budget guards.

• Telemetry from a real kernel module

  synapse_kmod is a DKMS module exporting syscall telemetry and AI scheduling hints through /sys/kernel/synapse/ — syscall_log, ai_hints, stats, status, config, version.

• A terminal that links no GL

  syntty is the default terminal: wl_shm, xdg-shell, xkbcommon and libc, cells to pixels on the CPU. 359 KB, a window in 5.8 ms, tabs, images, and OSC 133 prompt marks that synsh emits at the other end. A live image cannot count on a GL context.

• Applications of its own

  A file manager (synfiles), a package manager (synpkg), settings (syn-settings), an editor (syn-edit) and a disk utility (syn-disks) — each written for this system, and each a window and a command line over one binary rather than two programs that disagree.

• A ten-foot interface

  Big screen mode (Super+F10, or the pad’s Guide button) puts the Steam library, Big Picture, a browser, music, and any Plex or Jellyfin server on the network on a television — drivable from a controller, including as a mouse.

• An installer that respects your disk

  syn-install offers a whole-disk install with optional LUKS2 full-disk encryption, or — on UEFI, where the disk already holds another OS — a non-destructive dual-boot into existing free space, reusing the machine’s ESP. Four filesystems, three bootloaders.

How it fits together

One model, resident once, shared by everything. Components do not each load their own copy — the AI coding assistant reuses the model already sitting in synapd rather than spending a second allocation of VRAM.

  User
   │
   ▼
 synsh ─── natural language / commands ──┐
   │                                     │
   ▼                                     │
 synapd  (local LLM — Mistral 7B)        │
   │  inference over SYN socket protocol │
   ├──► synguard      security verdicts ─┤
   ├──► synnet        network policy     │
   └──► synapse_kmod  kernel sysfs       │
            │                            ▼
            ▼                    synui (Wayland)
     /sys/kernel/synapse/
     syscall_log, ai_hints, stats,
     status, config, version

The desktop

synui draws its own display settings, wallpaper picker, dock, cursor picker, sound panel and control panel. The status bar and the desktop widgets are a native quickshell shell. Everything below is compositor-drawn — there is no third-party panel or settings app in the picture.

Some of what the keyboard reaches

The full table — sixty-odd bindings — is in the README, and Super+C generates it live from the running bind table.

Components

Each lives in its own directory in the tree with its own PKGBUILD.

Download SynapseOS

Latest release: 0.2.9 — the terminal and the television. Free, x86_64, about 3.7 GiB. The model is downloaded during installation, not carried in the image.

Get the ISO from GitHub Releases

Reassemble and verify

The ISO is split into parts to fit GitHub’s 2 GiB asset limit. Download all .part* files, then:

cat SynapseOS-0.2.9-x86_64.iso.part* > SynapseOS-0.2.9-x86_64.iso
sha256sum -c SynapseOS-0.2.9-x86_64.iso.sha256

Write it to a USB stick

sudo dd if=SynapseOS-0.2.9-x86_64.iso of=/dev/sdX bs=4M status=progress oflag=sync

Or try it in a VM first

git clone https://github.com/velle999/SYNAPSE.git && cd SYNAPSE
QEMU_RAM=8G ./archiso/build_scripts/qemu-test.sh

The script auto-detects the newest ISO, uses KVM when available, boots UEFI via OVMF, and attaches a persistent 20 GB test disk.

System requirements

• CPU — x86_64. CPU inference is the default and works everywhere.
• RAM — 8 GB minimum; a resident 7B model is the reason.
• Disk — ~25 GB for an install, plus room for the model.
• Network — needed once, during installation, to fetch the model.
• GPU — optional. NVIDIA via a CUDA build, AMD and Intel via Vulkan.
• Firmware — UEFI or BIOS. Dual-boot installs need UEFI.

Frequently asked questions

Does SynapseOS send anything to the cloud?

No. The installer downloads a model once; after that the file lives on the machine and inference runs locally through llama.cpp. There are no API keys to configure and no account to create, and the AI works with the network cable out.

Is it just Arch with an app on top?

No. The inference daemon is a system service and the components query it over a socket: the shell resolves intent through it, the compositor talks to it, the security monitor uses it to classify events. The desktop, the shell, the security monitor, the network daemon and the kernel module are all written for this project.

Can I run it on a machine with no GPU?

Yes — that is the default. The ISO ships a CPU build of the inference backend so it starts on any x86_64 machine. GPU acceleration is a separate package you install afterwards if you have the hardware.

Is it free and open source?

Yes. SynapseOS-authored code is GPL-2.0-or-later, except the kernel module, which is GPL-2.0-only because the Linux kernel is. The whole tree is on GitHub.

Can I use it as my daily driver?

The author does. That is not a recommendation that you do — it is alpha at 0.2.x, it moves fast, and it will break. Run it in a VM first.

How do updates work?

An installed system needs two commands, because they cover different halves of it: sudo pacman -Syu for Arch, and syn-update check / syn-update apply for the SynapseOS components. There is a GUI for the second one in the start menu.