The Mac Mini sits on a shelf next to my router. M-series chip, 64GB of RAM, one ethernet cable. From the outside it looks like a tidy little box doing nothing. From the inside, it’s a forty-job autonomous AI platform.

I call it OpenClaw.

The setup

Three native LaunchAgents do the heavy lifting:

• ai.openclaw.gateway — the brain. WebSocket API on port 18789. Manages every agent session.
• ai.openclaw.node — the runtime that hosts the agents themselves.
• ai.openclaw.mlx-server — a llama-server process serving Qwen3.6-35B-A3B (Q4_K_XL) on 127.0.0.1:8080. This is the local brain.

Two agents run on top: Roger, my primary assistant, who talks to me through Telegram and Discord. And Hermes, a coding-focused agent I pair with for development.

Forty cron jobs run on a schedule. The morning brief at 7am. Markets at 9pm. GitHub trending at 8:30am. A weekly security audit Saturday morning. An obsidian vault gardener Sunday at 2am. All of it routed through Qwen3.6 locally — no API calls, no cloud bills.

Why local-first

People ask why I don’t just use OpenAI’s Assistants API or Anthropic’s managed agents. Three reasons:

Cost. Forty cron jobs a day across forty agents would run me hundreds of dollars a month on cloud APIs. On the Mac Mini, the marginal cost is electricity. Maybe a pound a day.

Control. When something breaks at 3am during the memory-dreaming job, I want to read the logs locally, not page through a cloud dashboard. When I want to add a new cron job, it’s a openclaw cron add away.

Privacy. A lot of what Roger sees is personal. My emails. My calendar. My positions. My conversations with Riri. None of that should ever leave the box.

The escape valve

That said, I’m not a purist. When a job needs the big guns — deep research, complex reasoning, code review — OpenClaw routes to Codex via free ChatGPT OAuth as the primary, and Claude Opus 4.6 as a fallback. Both authenticated through OAuth, both effectively free for my volume.

Local-first doesn’t mean local-only. It means local-by-default, with sharp tools when you need them.

What it’s like to live with

Honestly, it’s changed how I work. The morning brief at 7am means I never start a day cold. The 9pm market watchlist catches mover stories I’d miss. The weekly retro every Sunday tells me what I actually shipped, not what I think I shipped. The home network scan at noon catches new devices.

Some of it is automation that would be possible without AI. Most of it isn’t. The agents read articles, summarise them, decide what’s worth pinging me about, and write the message in my voice. That requires judgement, and judgement is what these models bring.

What’s next

I’ll write up specific pieces over the next few months — the cron architecture, the memory store, the Discord channels, the way Hermes pairs with Claude Code. If you’re building something similar, find me on GitHub and tell me what you’re working on.