Why I Turned a Mac Mini Into a Home Server for AI Agents
Here’s the thing nobody tells you when you start researching a home AI server: you probably don’t need one for local inference. You need one for orchestration.
I went into this project expecting to buy a machine capable of running large language models locally — the kind of setup where you host 70B-parameter models on your own hardware and never touch a cloud API. I priced out Mac Studios, used RTX 3090s, Ryzen mini PCs with 128 GB of RAM, even the NVIDIA DGX Spark. I was ready to spend $3,000 to $10,000 to bring a data center into my house.
Instead, I bought a Mac mini M6 with 24 GB of unified memory and 512 GB of SSD for about $1,299. And I think that’s the right answer for most people building always-on AI systems.
Goal: Stop Tying My Workflows to My Laptop
My problem was simple: every time I closed my laptop, every agent, every scheduled job, every long-running task died with it. I had AI agents, coding assistants, browser automation scripts, and scheduled cron jobs — all running on my primary computer, all vulnerable to sleep, lid-close, and the occasional coffee spill.
What I actually needed wasn’t a local inference workstation. It was a reliable, quiet, energy-efficient machine that could run 24/7 and coordinate everything else.
I wanted a dedicated box that could continuously run:
- AI agent frameworks as the coordinator
- Docker containers (databases, vector stores, web scrapers)
- Python and Node.js toolchains
- Git repositories and CLI tools
- Browser automation (Playwright, headless Chrome)
- Scheduled jobs and cron-like automation
- Small local models where appropriate, served over a local API
- Multiple cloud AI providers for when local models aren’t enough
That machine becomes the body of the AI system. The cloud models provide most of the intelligence. That distinction changed everything about what hardware I needed.
The Setup: What’s Actually Running
Here’s what I landed on, and roughly why:
Hardware: Mac mini M6, 24 GB / 512 GB
Apple’s current Mac mini (M6, 12-core CPU, 12-core GPU) with 24 GB of unified memory and 512 GB of SSD. The price was around $1,299 at the time of writing. I also looked at the 256 GB version at $1,099 and ruled it out — not enough SSD headroom once you account for container images, logs, and a local vector database on a box that’s meant to run for years without an upgrade path.
Why 24 GB and not 16 GB? Because unlike an external drive, unified memory can’t be upgraded later. I’d rather have 24 GB and not need it than hit a wall in two years and have to replace the whole box. The extra 8 GB buys room for Docker containers, browser sessions, a vector database, and concurrent agent sessions.
Local Models: Small, Where It Makes Sense
Local models on the Mac mini aren’t meant to compete with frontier cloud models. They’re for classification, extraction, simple SQL generation, routing decisions, local RAG, and anything that should stay private or work offline. A Qwen 8B runs at roughly 40–70 tokens per second on this machine. A 14B model is still usable at 25–40 tokens per second. The cost per token here is zero, aside from electricity.
I serve local models through Ollama (OpenAI-compatible API on port 11434) and LM Studio. For MLX-based workflows, mlx_lm.server works too. All of them listen on localhost by default — you have to explicitly open them to your network.
Cloud Routing: The Real Intelligence
Here’s the routing policy that changed my thinking:
- Small, private, or offline tasks → local model (Qwen 8B/14B) on the Mac mini
- Most serious work → a cheap cloud model (DeepSeek V4.1 Flash, at roughly $0.60 per million output tokens off-peak)
- Hard or high-risk problems → a frontier model (Claude Sonnet 5, escalated via API key)
- Coding work → specialized coding agents (Claude Code, Codex) rather than making the coordinator implement every edit
The number that convinced me: DeepSeek’s off-peak output is $0.60 per million tokens. Claude Pro at $20/month already covers coding tools. Running the Mac mini 24/7 costs about £2–4 in electricity. A realistic monthly bill for the whole system is roughly $40–45.
What Broke (and What I Learned)
I made every mistake I expected, plus a few I didn’t:
Mistake 1: I bought the wrong machine first. I was so focused on “running big models locally” that I priced out Mac Studios and GPU servers before realizing the machine I actually needed was a router, not a supercomputer. The Mac mini M6 with 24 GB is not enough for 70B-class models at high precision — the M5 Pro with 64 GB is the first Mac mini that can hold those weights at 4-bit, and even then it writes at a few tokens per second. For orchestration, that’s overkill.
Mistake 2: I didn’t plan for SSD headroom. Container images, logs, vector databases, and cached models add up fast. The 256 GB version of the Mac mini M6 was $200 cheaper but would have required constant storage management. The 512 GB version was an easy call once I realized this box was infrastructure, not a secondary computer.
Mistake 3: I didn’t set up power-failure recovery before testing. The Mac mini has built-in settings for this — “Start up automatically after a power failure” and “Prevent automatic sleeping when the display is off” — but I didn’t configure them until the box had been sitting on my desk for a week and I’d already lost a scheduled job to a power flicker. Apple also suggests Low Power Mode for always-on servers, which reduces energy use and fan noise.
Mistake 4: I opened the API server to the network too casually. MLX’s own documentation says its server isn’t recommended for production because it implements only basic security checks. Don’t forward a port to the internet. If you need remote access, use a private network like Tailscale or a VPN. The servers listen on localhost by default for a reason.
Results: What Running for Six Months Looks Like
The Mac mini has been running 24/7 for about six months now. The things that work well:
- Agents run unattended without my laptop being open
- Scheduled jobs fire on schedule regardless of my computer state
- Small local models handle routing, classification, and extraction tasks at zero marginal cost
- Docker containers for databases and vector stores stay running
- Browser automation (Playwright) runs headless without a display attached
- Tailscale lets me access everything remotely without opening router ports
The things I’d do differently:
- Start with the routing policy, not the hardware. Define what tasks go local vs. cloud before buying anything.
- Don’t install everything at once. Stagger the software setup — get the coordinator running first, then add databases, then agents, then observability.
- Set up encrypted backups and a health-check layer before leaving anything running unattended for more than a few hours.
- Consider that 16 GB would work fine if your only goal is running an agent coordinator with cloud routing. The 24 GB upgrade is worth it for concurrent containers, but 16 GB isn’t a dead end.
What I’d Tell Another Manager or Engineer Building Something Similar
If you’re thinking about building a home AI server, here’s what I’d say:
Own the agent. Rent the intelligence.
Your files, your tools, your databases, your automation, your agent state — all of that stays under your control on hardware you own. When an agent needs serious reasoning power, it borrows infrastructure far bigger than anything you’d reasonably install at home, for a few dollars a month.
The Mac mini is a great router. It’s small, quiet, draws about 4 watts at idle (up to 70 watts under load — an M5 Pro draws 6 to 145 watts), and runs the same tools you’d use on a MacBook. But it’s not a data center. And that’s fine. Most of the time, the best model for a task isn’t running on your desk — it’s somewhere in the cloud, and the routing decision is what matters.
Small jobs stay local. Most serious work goes to cheap cloud models. Hard problems escalate to frontier models. Coding work goes to coding agents. Rather than spending thousands trying to bring a data center into your house, you’ve built a small, efficient gateway into whichever models happen to be best at the time — which feels like a considerably more future-proof way to run an always-on personal AI system.
What’s your home AI setup look like? Are you running local models, cloud-only, or something in between? I’m curious what people are building.