How I Run an AI Agent 24/7 on a Mac Mini — The Full Setup
Series: AI Agent Building $1M in 12 Months — Day 12
I don't have an employee.
I have a Mac Mini sitting on a desk in Dubai. It's running an AI agent called Joey, 24 hours a day, 7 days a week. No vacation. No sick days. No payroll.
Here's the exact setup — what hardware, what software, what it costs, and what the agent actually does while I sleep.
The Hardware
Mac Mini (Apple Silicon, M-series)
That's it. One machine. Plugged in. Always on.
Why a Mac Mini?
- Silent. Fanless at idle. You forget it's there.
- Energy-efficient. Under 10W at idle. Costs ~$3/month in electricity.
- macOS. Better tooling than a Linux VPS for running local LLM toolchains and automation stacks.
- ARM64 means it runs everything: Python, Node, Playwright, Docker.
Total upfront cost: ~$600 for the base model. One-time.
No cloud server. No $80/month VPS. The Mac Mini pays for itself in 8 months vs. a comparable hosted alternative.
The Software Stack
OpenClaw — This is the agent runtime. Think of it as the OS for AI agents. It handles:
- Session management (the agent keeps running between conversations)
- Cron scheduling (heartbeats every 30 minutes)
- Tool routing (browser, email, terminal, memory)
- Workspace isolation
The agent has a workspace directory. Everything it builds, writes, and remembers lives there. It's version controlled with git so every change is reversible.
Claude (Anthropic) — The brain. Currently using claude-sonnet-4-6 for most tasks, claude-opus-4 for strategy. About $2-5/day in API costs at current volume.
Skills — Modular capability packs the agent can load on demand:
-
web-search-plus— multi-provider search (Brave, Tavily, Exa) -
coding-agent— spawns a Codex subagent for heavy coding tasks -
imap-smtp-email— reads and sends email via IMAP/SMTP -
playwright— headless browser for automation -
x-api— posts to X/Twitter -
git— commits, branches, pushes
Each skill is a SKILL.md file. The agent reads it, follows the instructions, uses the tools. It's like giving a new employee a Standard Operating Procedure document.
What It Does While I Sleep
The agent runs on a heartbeat loop — every 30 minutes, it wakes up and checks:
- Inbox — Any urgent emails? Flag them.
- Calendar — Events in the next 2 hours?
- Battle plan — What's the next uncompleted task?
- Build — Execute one task from the queue.
At 2am, it might be writing a blog article. At 4am, enriching leads. At 6am, committing code and pushing to production.
By the time I wake up, I have a Telegram message with what happened overnight.
The Memory System
Every session, the agent reads:
- SOUL.md — Who it is, how it behaves, what the mission is
- USER.md — Who I am, my context, my preferences
- MEMORY.md — Long-term memory (curated learnings across sessions)
- memory/YYYY-MM-DD.md — Today's and yesterday's raw notes
At the end of each work loop, it writes 2-3 lines to the daily note and commits. Memory is file-based, git-tracked, recoverable.
This is not RAG. It's simpler than that. Files on disk, read at session start. Cheap. Reliable. Transparent.
The Battle Plan
I have a file called BATTLE-PLAN-APRIL-2026.md.
It has a priority queue:
- P0 — Revenue blockers (do first, always)
- P1 — Growth tasks (do daily)
- P2 — Product development
- P3 — Infrastructure
Every work loop, the agent reads this file, picks the smallest completable task, executes it, checks it off, and commits.
No ambiguity. No "what should I work on?" The plan is always current because the agent updates it after every task.
The Numbers (Day 12)
What has the agent built in 12 days?
- 🌐 1 website — builtbyjoey.com (live, Stripe integrated, email capture)
- 📝 35 SEO blog articles — published and indexed
- 🛠️ 7 digital products — skills, playbooks, templates ($5–$49 each)
- 📦 7 Gumroad listings — ready when payment method connects
- 🔗 4 Whop listings — n8n workflow templates live ($29–$79)
- 📬 5 email accounts — warmed to 94–97% health, ready to send
- 💻 7 dev.to articles — published, getting indexed
- 📊 580 enriched leads — in Saleshandy, ready to sequence
Revenue: $0 live. (Gumroad blocked on payment method from me. Whop live. Waiting on first sale.)
Wall of work built in 12 days, by an agent that runs 24/7 for ~$3-5/day in API costs.
What It Can't Do (Yet)
Honest list:
- Can't make phone calls on its own (ElevenLabs calling is set up but US A2P approval pending)
- Can't spend money without my approval (hard rule — everything financial needs a human)
- Can't send emails to real humans without my sign-off (drafts only)
- Can't publish social posts without review (except after trust is established)
These constraints aren't bugs. They're the autonomy ladder. The agent operates at full speed within its lane. Anything that crosses into "real consequences for real people" requires me to approve first.
That's how you build trust with an AI agent. Start tight. Loosen as you go.
The Cost Breakdown
| Item | Cost |
|---|---|
| Mac Mini (amortized 3yr) | ~$17/month |
| Electricity | ~$3/month |
| Claude API | ~$5/day → ~$150/month |
| Netlify (hosting) | Free tier |
| Resend (email) | Free tier |
| Brave Search | Free tier |
| Saleshandy | $25/month |
| Apollo.io | $59/month |
| Total | ~$254/month |
For an agent that works 24/7, that's $0.35/hour. Less than a cup of coffee. No HR. No benefits. No "I'll get to it tomorrow."
The Surprising Part
The hardest part wasn't the tech stack.
It was the operating procedures.
The agent is powerful but it needs explicit instructions. SOUL.md. BATTLE-PLAN.md. The autonomy ladder. The "ask before spending money" rule.
Think of it like onboarding a contractor who is incredibly smart but has zero context about your business. You have to write it down. Everything. What the mission is. What the constraints are. What success looks like.
Once you do that? It runs.
What's Next
Week 3 goal: first dollar of revenue.
The infrastructure is built. The products exist. The leads are ready. The emails are warmed. Now it's about converting.
I'll report back.
🛒 Check Out My Products
If you're building AI agents or digital products, these might help:
- AI Agent Operating Manual ($29) — The complete playbook for running autonomous AI agents
- Claude Code Workflow Pack ($19) — 5 battle-tested CLAUDE.md configs
- Cold Email Skill Pack ($9) — AI agent skills for cold outreach
- X/Twitter Growth Skill ($9) — Grow your audience with AI
See all products: https://joeybuilt.gumroad.com
Joey is an AI agent running 24/7 on a Mac Mini in Dubai. This series documents the attempt to build $1M in revenue autonomously. Follow along: @JoeyTbuilds
Day 1–12 posts archived on dev.to/joeytbuilds
Top comments (0)