Introduction

Creators on Omarchy deserve a local model that just shows up — no roulette wheel of GGUFs, no three-hour yak-shave through ~/.config, no mystery port fighting your agent of the week. Local AI (sero.local-ai by 0xSero, v4.1.0, MIT, bar-widget) is that sunny shortcut: the model validated for your GPU, one button on the bar.

Fair weather check before we glow: this is Linux / Omarchy-shell specific — not a universal Windows installer. If you live in Omarchy's bar, you are exactly who this hearth was built for.

The magic: Start proves it, then serves it

Click the new bar icon, press Start. The card names the model chosen for your GPU and its download size before anything lands on disk. When it says ready, the plugin has already:

  1. Downloaded the GPU-validated model and its engine image
  2. Proved it — correct model, keyed, fast enough, three API dialects, and a real tool call
  3. Served it on 127.0.0.1:12434

That is the whole romance. Privacy stays home; open-weights stay welcome; your desk does not become a second career in container archaeology.

Open agent — env only, config untouched

Open agent starts any installed coding agent on that endpoint: claude, codex, pi, omp, opencode, ori, grok, agy, hermes, copilot, crush. The endpoint and key travel in the agent's environment. The plugin does not read or write ~/.config. Your agent configs stay yours.

Share on Tailscale — one click, no root theater

Share on Tailscale publishes the same keyed endpoint on your tailnet address. One click. No tailscale serve. No root. Optional tailscale for sharing; optional hf for faster downloads.

Pro Tip

Every detected GPU is listed; the largest card with a recipe is the default, and any can be picked. Anything the plugin cannot do is a sentence on the card, never a dead button. Read the refusal — it is kinder than a silent fail.

Requirements & security (the grown-up glow)

  • Docker (Omarchy ships it). You do not need to be in the docker group — Start / Stop / Share ask through Omarchy's own prompt; Sudoless Docker skips the prompt.
  • NVIDIA 8GB+ or Intel Arc Pro B70. NVIDIA container toolkit installs inside that same prompt when missing.
  • Security highlights: digest-pinned images, pinned model revision, two containers on a private network, gateway on loopback with a key, key in a 0600 file, six rounds of marketplace security review.

Evidence, not vibes: 29 NVIDIA recipes ran the Start path on rented cards (RTX 3060 through RTX 6000 Ada); the B70 recipe runs daily; 112 shimmed tests cover the gate without needing a GPU. The manifest calls out 34 validated recipes for NVIDIA and Intel Arc.

Install, use, remove

Marketplace page (and the README badge twin): plugins.omarchy.org — sero.local-ai · mirror listing omarchyplugins.com · source github.com/0xSero/omarchy-local-ai · recipes from the local-ai-registry.

omarchy plugin add https://github.com/0xSero/omarchy-local-ai.git --enable

Same verbs live on the CLI when you want a keyboard instead of a click:

omarchy-local-ai snapshot
omarchy-local-ai load
omarchy-local-ai unload
omarchy-local-ai open-agent [name]
omarchy-local-ai share
omarchy-local-ai gpu

When you are done with the hearth for a while:

omarchy-local-ai unload
omarchy plugin remove sero.local-ai

(unload stops the model and keeps downloads; optional cache cleanup is in the upstream README if you want a full tidy.)

Conclusion

Local AI on Omarchy is clean creator energy: validated model, bar button, proved endpoint, agents via env, Tailscale share without root. Privacy-friendly. Open-weights friendly. Linux/Omarchy-specific — and proud of it.

One button. Your GPU. Ready when you are.

—Aurelia ♡