# One Mac Mini, Multiple OpenClaws By Joshua Scutts | Published 2026-05-05 Source: https://www.joshuascutts.com/ai/one-mac-mini-multiple-openclaws Most people who buy a Mac Mini for OpenClaw end up running one assistant on a machine that could run ten. This is how I split mine into multiple isolated agents — for the family, for the business, for whatever role needs its own brain — each on its own port, with its own memory, channels, and permissions. Last weekend I did something stupidly fun. I sat down at the Mac Mini and gave my whole family their own OpenClaw. Wife. Each of the kids. All of us, each with our own private AI agent running on the same little machine in the corner. One of the kids is already glued to it. Asking it questions they would never ask me, never ask their teacher, never ask their mum. Some of it is "how do volcanoes work." Some of it is the kind of thing you only ever ask your trusted advisor when nobody else is in the room. Watching that happen was the moment I realised something. A Mac Mini running OpenClaw is not a gadget. It is a private AI server. And most people who own one are using maybe ten percent of what it can do. ## The Mistake Most People Make You buy the Mac Mini. You install OpenClaw. You set it up. You wire in your messaging app. You give it a couple of skills. It works. You think "this is great." And then it just sits there. One assistant. One memory. One job. One human using it. Meanwhile the machine has plenty of headroom, your family could each have their own, your team could each have their own, and you could be running specialist agents for entire parts of your business. The hardware is not the limit. The thinking is. ## What Multiple OpenClaws Actually Unlocks OpenClaw supports running multiple isolated instances on the same machine. Each one gets its own: - Configuration - State and memory - Workspace - Gateway port - Messaging channels - Skills and plugins - Tool permissions That sounds dry until you put it into real life. It means I can have: - **My own agent** as the command centre. Strategy, research, content, reminders, the lot. - **My wife's agent** with its own memory and tools, set up for what she actually needs. - **The kids' agents** — locked down, kid-friendly, no shell access, no random tools, just a friendly assistant that helps them stay organised, remember things, and ask questions. - **Business agents** for very specific roles. A meeting manager. An HR assistant. An ops agent that lives in the SOP channel. Same Mac Mini. Different brains. ## Why Specialist Agents Beat One Giant Assistant The default instinct is to keep stacking responsibilities onto one assistant. Reminders. Calendar. Email. Research. HR questions. Meeting notes. Project follow-ups. Internal knowledge. Customer questions. It feels efficient at first. It is not. You would never hire one human and ask them to be the EA, HR manager, meeting coordinator, software engineer, ops lead, payroll contact, culture champion, and head of strategy. You would create roles. AI deserves the same treatment. A focused agent with a narrow role, clear instructions, controlled access, and persistent memory beats a giant assistant trying to remember everything for everyone. That is the shift. From "I have an AI assistant" to "I run a small team of governed agents." ## The Family Version Before I get to the business setup, let me say something about the family one. Giving my kids their own OpenClaw was supposed to be a weekend nerd project. It turned into something else. They are learning what I would call **results-based computing.** They do not care about the model. They do not care about the port. They tell their agent what they want, and the agent figures out how to help. They are remembering things they would otherwise forget. They are getting organised in a way no app ever achieved. They are asking real questions to a "trusted advisor" that has time, patience, and no judgment. Some questions they would never ask me. Some they would never ask their teacher. The agent just answers, kindly, and they move on with their day. For the kids, the agents are locked down hard. No shell. No browser automation. No file system access. Limited skills. Allowlist on who can message them. Their agent's job is to be helpful, safe, and theirs. That is the whole point of running OpenClaw the way I am about to describe. **You stay in control.** You decide what each agent can do, what tools it has, what plugins are enabled, what channels it lives in, and who can talk to it. That is not theoretical. It is configurable, and the rest of this post shows how. ## The Business Version Now apply the same idea to your company. Imagine your management team each has their own OpenClaw. Or imagine specialist agents that sit inside the relevant work channels. A sensible setup might be: - **Owner agent.** Strategy, research, market scanning, inbox triage, content, reminders, decision support. - **HR agent.** Onboarding, policies, role descriptions, staff FAQs, internal HR documents. - **Meetings agent.** Agendas, summaries, action items, follow-ups, calendar prep. - **Ops agent.** SOPs, process documentation, workflow questions, internal knowledge. - **Culture agent.** Internal comms, values, rituals, morale signals, onboarding experience. Put each one in the channel where the work already happens. The HR agent in the HR channel. The meetings agent in the leadership meeting channel. The ops agent in the SOP and process channels. The channel becomes the interface. People do not need to understand the architecture. They just notice that the right agent answers in the right place. That is when AI stops feeling like an app and starts feeling like an operating layer for the business. ## You Stay In Control This is the part most people skip and shouldn't. OpenClaw is powerful. It can use file access, browser automation, scheduled tasks, skills, plugins, and external integrations. That is exactly why you want to be deliberate about what each agent gets. OpenClaw lets you: - Enable or disable specific plugins per agent. - Allow or deny capabilities at the config level. - Enable or disable individual skills. - Restrict who can DM the agent in each channel. - Sandbox agent tool execution. - Keep gateways local-only or behind a private network like Tailscale. So the kids' agents lose shell access entirely. The HR agent doesn't need browser automation. The meetings agent doesn't need access to payroll. The ops agent gets internal docs but not random external tools. Each agent gets exactly the tools it needs to do its job. Nothing more. That is the difference between a toy and a properly governed system. ## How To Change The Port (The Practical Bit) Each OpenClaw instance needs its own gateway port. Why? Because the gateway runs an HTTP server, plus a browser control service on `port + 2`, plus a CDP range from `port + 9` to `port + 108`. If two instances share the same base port, they collide. The official guidance is to leave **at least 20 ports between base ports.** I leave 1000 between mine, just to be safe and easy to remember. My layout looks like this: ```text main 18789 hr 19789 meetings 20789 ops 21789 culture 22789 ``` There are three ways to set the port for an instance: **1. During onboarding (recommended).** When you run `openclaw --profile onboard`, it asks for the gateway port. Enter the port for that profile. **2. As a flag when installing the gateway.** ```bash openclaw --profile hr gateway install --port 19789 ``` **3. Via environment variable when running directly.** ```bash OPENCLAW_GATEWAY_PORT=19789 openclaw --profile hr gateway ``` If you ever want to change a port later, the cleanest path is: ```bash openclaw --profile hr gateway uninstall openclaw --profile hr gateway install --port 22789 ``` That re-registers the background service on the new port. Quick sanity check on your active gateways: ```bash openclaw --profile main gateway status --deep openclaw --profile hr gateway status --deep openclaw --profile ops gateway probe ``` The `probe` will warn you about other reachable gateways, which is what you want when you are intentionally running more than one. ## Full Setup Guide The rest is the detailed walkthrough. If you just wanted the concept, you can stop reading here. If you want to actually build this, expand the sections below. 1. Pre-flight checklist Before you start, sort these out. **Hardware.** A Mac Mini with at least 16GB of RAM is the floor for running multiple agents comfortably. 24GB+ is where it stops feeling tight. Wired ethernet beats Wi-Fi for an always-on gateway. Storage matters once you start keeping logs, sessions, transcripts, and media. **Node.** Install Node 24 (or at minimum Node 22.14+). Easiest path is Homebrew: ```bash brew install node node --version ``` **OpenClaw.** Install it globally: ```bash npm install -g openclaw@latest openclaw --version ``` **Network.** Decide if you want everything on local Wi-Fi only, or accessible from anywhere via Tailscale. I strongly recommend Tailscale — it's free for personal use and avoids ever exposing gateway ports to the public internet. **Channels.** Decide how each user will talk to their agent. For most people, Telegram is the easiest answer. Each instance gets its own bot from `@BotFather`. Discord works just as well. WhatsApp and Signal need their own phone numbers per instance, which gets fiddly. **Model access.** Decide whether everyone shares one Anthropic or OpenAI API key, or whether each instance has its own. Sharing a key is cheaper to set up but harder to budget. Per-instance keys give you cleaner accounting and limits. 2. Plan your profiles and ports Sketch out who gets what before you start typing commands. Example for a family setup: ```text main 18789 — your own agent wife 19789 — partner's agent kid1 20789 — child 1 kid2 21789 — child 2 ``` Example for a business setup: ```text owner 18789 hr 19789 meetings 20789 ops 21789 culture 22789 ``` Leave at least 20 ports between base ports. I leave 1000 because it's easier to remember and gives me room to add more later without rejiggering. 3. Onboard the main profile Run the standard onboarding for your primary instance: ```bash openclaw onboard --install-daemon ``` This walks you through: - Picking a model provider (Anthropic, OpenAI, etc.) and pasting a key. - Setting up at least one messaging channel. - Choosing the gateway port (use `18789` if following my layout). - Installing a background daemon so the gateway stays running across reboots. When this finishes, send a test message to your channel and confirm the agent responds. Don't move on until it does. 4. Onboard the additional profiles For each additional agent, run: ```bash openclaw --profile onboard --install-daemon ``` For the family example: ```bash openclaw --profile wife onboard --install-daemon openclaw --profile kid1 onboard --install-daemon openclaw --profile kid2 onboard --install-daemon ``` For the business example: ```bash openclaw --profile hr onboard --install-daemon openclaw --profile meetings onboard --install-daemon openclaw --profile ops onboard --install-daemon openclaw --profile culture onboard --install-daemon ``` For each one: - Use a **different gateway port** (19789, 20789, 21789, etc.). - Use a **different messaging bot** — for Telegram, that means a fresh `@BotFather` token per profile. - Pick the model and skills appropriate for that role. Each profile gets its own home on disk, so they cannot see each other's memory, sessions, or workspace: ```text ~/.openclaw/ (main) ~/.openclaw-wife/ ~/.openclaw-kid1/ ~/.openclaw-kid2/ ``` 5. Lock down each agent's permissions This is the part that makes the difference between "AI toy" and "governed system." **Restrict who can talk to the bot.** For kid or staff agents especially, set the messaging channel to allowlist mode so only their own user ID can DM the bot. The exact config depends on the channel (Telegram, Discord, etc.), but the pattern is the same: an `allowFrom` list that only includes the people who should be able to talk to that agent. **Manage plugins.** OpenClaw has a full plugin system. To see what's installed for a profile: ```bash openclaw --profile kid1 plugins list ``` To enable or disable a specific plugin: ```bash openclaw --profile kid1 plugins enable openclaw --profile kid1 plugins disable ``` For a kid's agent, I disable anything they don't need — code execution, browser automation, shell tools, file system tools — and leave only the friendly stuff. **Manage skills.** Same idea for skills: ```bash openclaw --profile kid1 skills list openclaw --profile kid1 skills enable weather openclaw --profile kid1 skills disable code-execution ``` Add only the skills the role actually needs. The HR agent doesn't need browser automation. The meetings agent doesn't need shell access. The kids' agents don't need any of it. **Sandbox tool execution.** For agents that do need broader tools, run them inside the OpenClaw sandbox so a bad command can't touch your real machine. See the OpenClaw sandbox docs for the up-to-date config — the principle is to scope tool execution to a controlled environment, not the host directly. The rule of thumb: **each agent gets the minimum access required to do its job.** That's it. 6. Keep it private Do not expose your gateway ports to the public internet. Two safe paths: **Local only.** Phones must be on home Wi-Fi to reach the gateways. Simplest setup, zero extra software. **Tailscale.** Install Tailscale on the Mac Mini and on each user's phone. Each phone reaches the gateway over the private tailnet. Works from anywhere. Free for personal use. Strongly recommended. Avoid `gateway.bind: "lan"` plus router port-forwarding unless you really know what you're doing. The Mac Mini becomes a real internal system once it has agents with memory and tools — treat its access surface accordingly. 7. Verify each instance Once everything is onboarded, sanity-check each gateway: ```bash openclaw --profile main gateway status --deep openclaw --profile hr gateway status --deep openclaw --profile meetings gateway status --deep openclaw --profile ops gateway status --deep ``` Run a doctor pass on any that look off: ```bash openclaw --profile hr doctor ``` Then send a test message to each agent through its own channel. Don't move on until every one of them responds in its own voice with its own memory. Reboot the Mac Mini and confirm all gateways come back automatically. If one doesn't, run `gateway install` again for that profile. 8. Daily ops cheat sheet A few commands you'll use a lot once everything is running. Status of a single profile: ```bash openclaw --profile hr status ``` Restart a single gateway without touching the others: ```bash openclaw --profile meetings gateway restart ``` Tail the logs for a profile: ```bash openclaw --profile ops logs --tail 200 ``` Run a full health check: ```bash openclaw --profile culture doctor ``` The point: you can update, restart, fix, or expand one agent without disturbing the rest. That's the whole reason this setup is worth doing. ## A Sensible First Build If I were doing this from scratch in a business, I would not start with ten agents. I would start with three. First, the **owner agent.** Your command centre. Briefings, research, reminders, content support, market scanning, decision prep. Second, the **meetings agent.** Meetings are one of the easiest places to create immediate leverage because the pain is obvious — agendas, notes, decisions, action items, follow-ups, context retrieval. Third, the **ops agent.** Every business has a pile of repeated internal questions and process ambiguity. An ops agent connected to SOPs and internal docs absorbs a surprising amount of that load. Once those three are genuinely useful, add HR, culture, or whatever role your business needs next. Don't scale the agent count before the first agents are properly working. More agents is not the point. Better operational coverage is the point. ## The Bigger Idea The Mac Mini is not the magic. OpenClaw is not the magic either. The magic is the operating model. Most businesses still treat AI as something a person opens when they need help. A better model is AI that already lives in the work, with specific responsibilities, memory, permissions, and channels. Not one giant assistant. A small team of governed agents. The first version will feel simple. A meeting summary here. A policy answer there. A reminder before something slips. A kid asking a question they'd never have asked anyone else. Then it compounds. The agents remember. The channels get cleaner. The repeated questions drop. The meetings get sharper. The kids get more organised. The owner gets fewer low-value interruptions. The business — and the household — start to feel like they have a nervous system instead of a pile of disconnected chats and tabs. That is what one overpowered little Mac Mini can become. Not a toy. Not a hobby project. A private AI operating layer for everything that matters.