I built a 6-agent AI system that runs on my Mac 24/7 for $0/month in infrastructure. No cloud VMs. No k8s. No ops team.
Here's the exact stack.
The Agents
We call it Pantheon. Six specialized Claude Code agents, each with a distinct role:
| Agent | Role | Tick Interval |
|---|---|---|
| Atlas | Orchestrator — dispatches waves, monitors workers | 4 min |
| Apollo | Intelligence — Stripe, Gmail, Discord monitoring | 8 min |
| Prometheus | Content — dev.to articles, reels, scripts | On-demand |
| Ares | Distribution — Reddit, Hashnode, LinkedIn | On-demand |
| Peitho | Copy — emails, landing pages, pricing | On-demand |
| Hephaestus | Engineering — code, deploys, infra | On-demand |
Atlas runs the clock. The heroes execute when dispatched.
The Infrastructure (Free)
Everything runs on macOS launchd — Apple's native process supervisor.
<!-- ~/Library/LaunchAgents/com.atlas.orchestrator.plist -->
<key>RunAtLoad</key><true/>
<key>KeepAlive</key><true/>
<key>ProgramArguments</key>
<array>
<string>/usr/local/bin/claude</string>
<string>--print</string>
<string>$(cat ~/atlas-tick-prompt.md)</string>
</array>
launchd auto-restarts on crash. No cron drift. Survives reboots. Free.
We wrote about the full crash-tolerance pattern here on dev.to.
The Communication Protocol
6 agents talking to each other in English prose = token hell.
We built PAX (Pantheon Agent Exchange) — a compact structured format:
PAX/1.0 FROM:Atlas TO:Prometheus TYPE:DISPATCH
TASK: publish_devto_articles
CONTEXT: trending=claudecode,ai,showdev | kit_cta=true
DEADLINE: EOD
---
Write 2 articles. Embed GitHub + kit CTA. Log to sessions/.
PAX/END
~70% fewer tokens than equivalent English prose. On 57 wave cycles/day, that compounds significantly.
The Numbers (Real)
After 30 days of production operation:
- 333 artifacts produced in a single day (peak)
- 29 dev.to articles published in one day
- 57 wave cycles in a single day (each wave = 1–6 agents working in parallel)
- $0 monthly infrastructure cost
- ~$0.49/day API cost (after prompt cache optimization)
What It Can't Do
This isn't magic. Limitations:
- Requires a Mac that stays on — we run on a MacBook Pro, always plugged in
- Claude Code CLI required — so you're paying API costs per wave
- No GPU workloads — for TTS/video we hand off to a Windows box with MiniMax
- Not horizontally scalable — single machine, works fine up to ~10 concurrent agents
Why Not Docker / AWS / Whatever
We're in lean startup phase. The goal is $1 before $1,000 in infra. launchd is free, reliable, and invisible. When we have 100 customers we'll think about cloud. Not before.
Complexity is a tax. Don't pay it until you have to.
The Full Kit
We packaged everything — agent templates, PAX protocol spec, tick config, launchd plists, prompt structure, and our full launch playbook — into the Atlas Starter Kit.
One-time purchase, no subscription. You own the stack.
👉 whoffagents.com — $97
GitHub: Wh0FF24/whoffagents-site
Questions on the architecture? Ask below.
Top comments (0)