Local AI Isn't Ready Yet - So I Built An Omarchy AI Shed And Run Claude From My Phone

DESCRIPTION:
People keep asking how I churn out so much software. This video is the honest answer. I seriously looked at going fully local with AI — did the research, ran the numbers, ran the models — and came away unconvinced. So instead I scavenged together an "AI shed", a box that sits in the corner doing all my agent work while I get on with my life, and I drive it from anywhere using tools that are already built into Omarchy. Plus a quick heads-up on the Omarchy 4.5 release candidate.

OMARCHY 4.5 RELEASE CANDIDATE

It's on the release channel now and it is a toy festival — enhanced applications, Glyph remote desktop, and a gazillion other things. I'm not reviewing it properly until it ships because that's not fair on the team. Grab it and have a play. Full review when it lands.

LOCAL AI VS CLOUD AI — THE NUMBERS

I put an intelligence index table up on screen. Opus 5.5 sits at 55 overall and 60 on coding. The best open models you can run yourself top out around 46 — and those are huge. What actually fits on a normal machine, a 64GB PC with an RTX 5060, scores about 34. Strix Halo with 128GB gets you to 40. To get anywhere near the top you're into RTX Pro territory or a 512GB unified-memory Mac, and the card alone is £15,000.

I did this dance twenty-five years ago in 3D. A single CGI frame at 720x576 could take an hour. A 750-frame commercial was 31 days on one box, so we rented render farms in the cloud and got it back in six hours. The difference now is we're not buying compute, we're buying intelligence — and the local stuff feels like a flashback to ChatGPT 3. Fine for a story about potatoes. Not fine for C++, graphics and a full application in one go.

My maths: £2,000 of local hardware plus £190 a month for Claude Max takes 81 payments — six or seven years — to reach the £17,000 a serious local rig costs. And every six months the cloud models get better while that hardware depreciates. I paid £7,000 for a Mac trash can once. There are good reasons to go local — air-gapped businesses, small models parsing your own code, simple always-on jobs, image work with Flux — but for an individual building serious software, it isn't there yet.

THE AI SHED BUILD

My old shed was a Dell 8950, but my son's gaming PC died (cat hair — don't), so he got the Dell and I built a replacement on the cheap. Ryzen 9 5950X — 16 cores, 32 threads, £300 now when it used to be £800. AM4 means DDR4, and I scavenged 64GB from other machines rather than paying today's RAM prices. Intel Arc B580 because I do a lot of Vulkan work and Vulkan is GPU-agnostic. Old SSDs, an old 750W PSU, a Lian Li case, a Thermalright cooler and some fans. I asked AI what would do the trick and that's what did the trick. No build video — the wiring is not pleasant. Who cares, it works.

HOW I DRIVE IT REMOTELY

Everything is on Tailscale — free for one seat with as many machines as you like, and your whole network just talks to itself. Glyph in Omarchy 4.5 is the game changer: proper remote desktop for Hyprland, so instead of tmux-ing everything I jump between machines graphically. I built a Glyph manager to switch between my two boxes, Colossus and Guardian (if you know the film reference, you know).

On the shed I leave sessions running: Codex building my World War II U-boat game, Claude working through OmaEdit. Codex has a remote server mode — turn it on and every session appears on your phone in the ChatGPT app. Claude is even simpler: type /rc before you walk away and the session continues on your phone. If I need to see something, I tell it I'm remote and it sends me the image or a video. I get notifications as it goes. I set goals with /goal for long unattended runs, and I always leave a couple of chats open so I can kick off a new project from the train to Hampstead Heath. By the time I'm home half of it's built. You can't escape human testing though.

Models: Codex high and Opus 5.5 day to day, Fable only for planning. Two accounts, and I still don't hit my limits.

COMING UP

OmaBroadcast — I've rebuilt OBS for Omarchy on FFmpeg 9.1. OBS is gone from this channel in a few weeks. Full video soon. The U-boat game is nearly done — I fed it the plans for every wartime vessel and the script to Das Boot. And for the SteamOS crowd: 3.9.5 has pushed over the air. As soon as Valve release the recovery ISO I'll fold the Nvidia drivers back in and redo the installer.

Takeaway: you need very good reasons to go local. Omarchy 4.5, Tailscale and a bit of imagination will get you further for a fraction of the money.

ABOUT THIS CHANNEL

I make videos about Linux, Omarchy, creative tools, gaming, and self-hosting. Subscribe if that's your thing.
00:00 - Start
00:09 - PRE AMBLE
01:27 - LOCAL V CLOUD
10:57 - LOCAL BUILD
15:42 - OMARCHY RS RC
17:29 - RUNNING IT ALL
22:37 - FINAL WORDS