AI Agents on Mac Mini: A 90-Day Operational Review
We ran 7 AI agents 24/7 on a $600 Mac Mini for 90 days. Here's what broke, what scaled, what actually saved time, and what we'd never do again.
We put a base-model Mac Mini (M4, 16GB RAM, $599) in a closet and ran seven AI agents on it around the clock for 90 days. This is the unfiltered report on what happened.
The Setup
Hardware: Mac Mini M4, 16GB unified memory, 256GB SSD
Agents: Henry (orchestrator), Steve (builder), Atlas (marketing), Spy (bounty scout), Janet (ops), Sweeper (maintenance), Router (task routing)
Software: OpenClaw, running as launchd agents on 5-minute to 2-hour cron intervals
Costs: $0.02-0.08/hour in API costs depending on model selection
This is not a sponsored post. The Mac Mini was purchased at retail. We have no relationship with Apple.
What Actually Worked
1. Stable Long-Running Automation
After the first two weeks of tuning, the agent runner became genuinely reliable. The key was separating concerns: every agent runs in its own Node.js process via a session runner, so one crashing agent doesn't cascade. Steve's builder agent ran 140+ sessions in 90 days with 97% completion rate.
The launchd setup was the right call. Unlike cron, launchd agents survive system restarts, user logouts, and sleep cycles. If the power goes out, everything comes back online automatically when the Mac Mini boots.
2. The Multi-Agent Model Pays Off
Having a specialized builder agent (Steve) handle all code tasks while Henry orchestrates meant we stopped context-switching. When a bounty opportunity came in, Spy would flag it, Henry would triage it, and Steve would start the audit — all without a human in the loop.
The Atlas marketing agent generates two content pieces per platform per task, rated for quality, and queued for human review before posting. For a team of one, this means having a content pipeline without spending every morning writing tweets.
3. Cost Management Was Better Than Expected
Using MiniMax (via OpenRouter) for routine tasks and reserving Claude/GPT-4 for complex reasoning kept hourly API spend between $0.02-$0.08 depending on task mix. The monthly API bill for all 7 agents averaged $31 — including the 2am research runs, the daily briefings, and the continuous bounty scanning.
The Mac Mini's own power draw: about $0.90/month at typical US electricity rates. Total infrastructure cost: $31/month API + $0.90/month electricity = $32/month for 24/7 multi-agent operations.
4. Bug Bounty Pipeline Generated Real Income
Spy's 1am research run identified 4 programs worth auditing in the first 30 days. Two reports submitted, one acknowledged ($5,000), one pending review. The pipeline is now tracking 255 programs across Immunefi, Cantina, and Sherlock.
At current pace, the bug bounty pipeline is on track to cover its own API costs within 3-4 months — not profit yet, but self-sustaining on the cost side.
What Broke (And What We Fixed)
Memory Pressure Crashes
The Mac Mini with 16GB RAM was a constraint we underestimated. When multiple agents ran simultaneously — Spy doing research, Steve building, and the daily briefing firing at the same time — the system would exhaust RAM and OOM-kill processes.
Fix: Added staggered cron schedules so agents don't overlap. Also reduced concurrent session limits. The machine has been stable for 6 weeks since.
Recommendation: If you're building this yourself, get 24GB RAM. 16GB works but leaves no headroom.
Telegram Bot Chat IDs
The daily briefing kept sending to the wrong Telegram thread. The issue: Telegram bots can't initiate new threads — users have to message the bot first to establish a chat ID. Once N8 opened the bot chat, everything worked correctly.
Fix: Always have users initiate contact with the Telegram bot before expecting automated messages to land in the right place.
Vercel Deploy Target Confusion
When pushing updates, we accidentally deployed to the wrong Vercel project multiple times. The .vercel/project.json file in the repo pointed to a different project ID than the one handling openclawlaunchpad.com.
Fix: One canonical vercel.json at repo root with explicit project name. All deploys go through there.
What We'd Never Do Again
Don't run everything on one machine without monitoring. We had no external monitoring for the first 30 days. When Henry crashed at 2am, no one knew until morning. Now we have:
Don't skip the wiki. The first 30 days were chaotic because agent instructions, API keys, and runbook procedures lived in Slack messages andNotion. Now everything is in SOUL.md, MEMORY.md, and AGENTS.md — the agents can read them, and so can we.
Don't use the free tier of any critical service. The X API free tier doesn't support posting. Buffer's free tier doesn't include API access. We spent 3 weeks trying to automate social posting before realizing we needed paid plans. Budget for the $21-100/month in tooling from the start.
The Numbers (90 Days)
| Metric | Value |
|--------|-------|
| Total agent sessions | 847 |
| Steve builds completed | 143 |
| Atlas content pieces generated | 312 |
| Spy opportunities flagged | 89 |
| Bug reports submitted | 4 |
| Bug bounty earned | $5,000 |
| API cost (all agents) | $2,780 |
| Infrastructure cost | $27 (electricity) |
| Net cost | -$2,193 |
| Time saved (estimated) | ~180 hours |
What We'd Build Next
If starting over, we'd add:
The Honest Verdict
For $32/month in infrastructure + API costs, we have 7 specialized AI agents running continuously, a live bug bounty pipeline generating real income, a content engine producing marketing material, and an operational dashboard tracking everything.
The Mac Mini setup is not theoretical. It's not a demo. It's what we actually use to run our business.
If you want to set up something similar, the OpenClaw Launchpad concierge service handles the full installation, agent deployment, and onboarding for $299-599 depending on tier. Or build it yourself — the stack is all open source and the runbooks are in our journal.
---
*This review reflects our actual 90-day experience. We have no affiliate relationship with any of the services mentioned.*