July 17, 2026 · 4 min read
How I Actually Run My AI Infrastructure (No Expensive Hardware)
A non-technical founder's real AI setup: agents built around workflows, Claude + Codex orchestration, the six-step loop, and a VPS that never crashes.
How I Actually Run My AI Infrastructure (No Expensive Hardware)
Uh.
So I was debating on putting this out here because it’s literally my sauce. Very much an unorganized brain dump.
But I figured I would anyway — just to give value and let people know how a real non-technical gangster runs his AI setup (WITHOUT EXPENSIVE HARDWARE), since most people are still confused.
Here’s how to do it. Get prepared for the longest, most technical thing you’ll read today, and be even more confused. Let’s have some fun.
First: see this as building a moat
Before you start, you need to see this as building a moat.
Deeper moats > latest tech. My book “Entrepreneurs Cockpit” is about this concept. You need to raise this thing like a child and commit to its growth. Vibe Scaling isn’t a weekend project — it’s a system you feed until it feeds you.
The base software
If you’re gonna do this, your computer will need SOME of these installed:
- Homebrew
- Node.js
- Python
- Tailscale
- Docker
(DYOR on this, OK?)
Build agents around your actual workflows
Check that, then first things second: build all your agents around your actual workflows — the things that really matter and move the needle. Not around whatever tool is trending this week. Your agents should map to the jobs your business already does every day.
Build your arsenal
Then you build your arsenal. Your agents’ plumbing and toolkit. What does each agent need? Every agent gets the specific tools, skills, and access it needs to do its one job well.
Make sure all your environment variables are in Railway. Logins, APIs, all of it.
Then confirm your agent network and arsenal are backed up to local files, GitHub, Supabase, AWS, Hostinger, Vercel, Obsidian, or whatever else you’re using. Backups everywhere. If one surface dies, the network survives.
Wire Claude and Codex together
Go tell Claude Code in the terminal to find the Codex plugin for Claude Code and install it.
Then tell Claude to make sure it always uses:
Claude as the orchestrator.
Codex as the executor.
One brain directs. The other builds. Keeping those roles separate is what makes the whole thing stable.
Then set it to YOLO mode:
--dangerously-skip-permissions (or /settings > configure permissions setting)
Put it on Ultra Code “/effort” level.
The six-step loop
Then give it some insanely huge task and tell it to follow this process:
“With Claude as the orchestrator and Codex as the executor, use this six-step agent process:
1. Research — Do deep research around the real current state before anything.
2. Plan — Define the shortest safe path, success criteria, and rollback.
3. Reference — Check the agent and Arsenal infrastructure. Load the right skills, tools, agents, and existing systems.
4. Execute — Do the authorized work and capture evidence as you go.
5. Review — Verify against the intended outcome. Repair anything incomplete or wrong.
6. Loop — Feed review findings back into Research and repeat until there’s a proper stopping point.
(These steps are infinitely expandable – this is a concise version)The key principle:
Don’t stop because the first pass is complete.
Keep looping until the result is neither shallow nor wastefully overbuilt. Only stop for a genuine approval, credential, or business-decision gate.”
There are some additional tools for front end, back end, database, and caching to look into to equip your AI with.
Mirror it to a VPS
It doesn’t stop there.
Then what you do is you tell Claude to MIRROR this exact config / ability onto your VPS.
Then you do this for literally every one of your projects. You can run hundreds of agents parallel in one terminal session. And you can have dozens of terminal sessions running inside of Claude Code’s agent terminal. Codex can kinda do a similar thing in their desktop app, which I also really love.
My laptop crashed at five projects in running this. My VPS is solid though. Doesn’t crash when I do the same. It just took a long time to configure. And I think most people are too lazy to actually config it properly.
The honest risk disclosure
Note: I am crazy. And I am very risk tolerant. I gave full disk access and accessibility access locally to my AI infrastructure (Claude, Codex, and Hermes). Which honestly has been very interesting but also helpful. It still won’t use my card or post or send anything without my authorization though.
Don’t copy that part unless you understand what you’re signing up for.
One brain, every surface
So I run:
My phone.
My computer.
My VPS.
All of it is also hosted on a custom website, so I can log in and have the same experience no matter where I tap in from.
I have lockdown files, LLM collaboration Markdown files, a canonical One-World infrastructure principle, semantic hooks that invoke skills, and systems designed to sync the entire network back to one brain.
Then I can drive it however, whenever, and wherever I like. That’s the Cockpit. Out of the Cave, hands on the controls, the Fleet doing the work.
Feels like flying. After a year of confusion, I finally have such clarity on how to drive this thing. It’s been like learning to play golf.
Written by Josh Collier