Turn a Mac mini into your team's private AI server

Most small teams that want private AI hit the same wall. The cloud assistants are easy, but nobody is comfortable pasting client work into them. Running models yourself sounds right, but nobody wants to babysit a GPU server. There is a middle path that fits on a desk: a Mac mini that runs your team's agents with local models, inside the same boundaries your SweetHive workspace already uses.

This guide walks through what that machine does, which hardware makes sense, how to set it up, and how to run automations on the same box.

What the machine actually does

In SweetHive every teammate can have a personal agent tied to the contexts they work in. The agent answers from the messages in those contexts, and it can never see more than the person who owns it. Something has to run the model behind each agent. You have two options, and the Mac mini can play either role.

SweetHive Agents Node is the free desktop app for one person. It ships with the Ollama runtime built in, installs models in one click and runs that person's agents. It is available for macOS (Apple Silicon and Intel) and Windows.
SweetHive Agents Server hosts the agents of many people from one machine. It has a desktop edition for macOS and Windows, made precisely for organizations that would rather use a Mac mini or a Mac Studio than a Linux box. It is part of the Organization plans.

If you want one quiet machine that serves the whole team, the Agents Server is the one you want. Each member still owns their agent: the server keeps every person in a separate, encrypted tenant, and conversations, tokens and memory are never mixed between people. The admin console shows operations, not content.

Choosing the hardware

On Apple Silicon the model lives in unified memory, so RAM is the number that decides what you can run. The model catalog in the SweetHive apps groups models by the memory they need. As a rough guide, and always leaving room for macOS and the apps themselves:

8 GB: small models such as Qwen 3.5 4B. Fine for short answers and simple lookups, not for long summaries.
16 GB: the standard tier, for example Qwen 3.5 9B or Gemma 4 12B. A reasonable floor for a team machine.
32 to 40 GB: the performance tier, for example Qwen 3.6 35B (a mixture-of-experts model that stays fast) or Gemma 4 26B. This is where agents start to feel genuinely useful on multi-step questions.
96 GB and more: server-class models. Beyond what a Mac mini offers, so look at a Mac Studio if you need them.

For a team of five to ten people, a Mac mini with 32 GB or more is a sensible target. With 16 GB it works, but expect simpler answers. Download sizes range from about 2 GB for the smallest models to over 20 GB for the performance tier, so plan some free disk space too.

One honest note on speed. The server shares one model runtime across everyone, with fair scheduling: each person gets at most one answer generating at a time, and there is a global cap on parallel work. If three people ask at the same moment on a modest machine, someone waits a few seconds. For a small team that is usually fine; for heavy use, more memory helps more than anything else.

Setting it up

Create the server in SweetHive. An organization admin opens Settings > Organizations > Agent servers, creates the server and copies the enrollment key. The installer download is linked from the same tab.
Install on the Mac mini. The first-run setup asks for a console password, the enrollment key, whether the console is reachable only from this computer or from your local network, and a default model. Keys are stored in the macOS Keychain.
Export the recovery key. Do this on day one and keep it somewhere safe. Without it, the encrypted tenant data cannot be recovered on a new computer.
Let it stay awake. The desktop edition can start at login, keep the machine awake and restart itself after a crash. Plug the Mac mini into wired network and leave it on.
Members pick the company server. When a teammate creates their agent in SweetHive, they choose the company server instead of their own computer. Nothing to install on their laptop. The agent arrives on the server automatically.

What the agents can see

Running locally does not mean running wide open. Each agent is scoped to the contexts its owner chose, and SweetHive checks that scope against the owner's current visibility on every request. If someone leaves a project group, their agent loses that project too.

Agents work from the messages in their contexts: they can list recent activity, search by keywords and, if their owner allows it, draft and post replies after asking for confirmation. With Agents Node 0.2.23 or later they can also read the files attached to those messages: the text of PDFs, scans and images through a local model that reads images, and text files such as CSV, up to 20 MB each. Office documents, audio and video are not read yet. If your team shares scans or photos, install one image-reading model on the Mac mini next to the main one. A one-line conclusion in the message still pays off: keyword search finds it, and so do your colleagues.

Automations on the same box

The Mac mini is also a good home for automations, and here you have two tools that complement each other.

SweetHive Workflows

Under My agents > Workflows you can draw a chain of steps on a canvas: generate text with a model, extract structured fields, call an HTTP endpoint, branch on a condition, post the result into a context. A workflow starts manually, on a schedule or from a webhook, and every step runs on your own machines, not in the cloud. Each workflow runs as one of your agents, so it can never read or post beyond that agent's scope.

n8n next to it

If your team already uses n8n, a self-hosted instance runs happily on the same Mac mini. There is no dedicated n8n integration, but the two can talk through plain HTTP:

From n8n into SweetHive: give a SweetHive workflow a webhook trigger. When you save it, you get its URL and a secret, shown once. An n8n HTTP Request node can then call that URL with the secret in the X-Webhook-Secret header, and the JSON body becomes the workflow's input.
From SweetHive to n8n: the HTTP request step only calls public HTTPS addresses and blocks private networks, so localhost will not work. Expose the n8n webhook through HTTPS (a reverse proxy or a tunnel) if you need this direction.

A small example: n8n watches your website form, sends each new request to a SweetHive workflow, a local model writes a two-line summary, and the workflow posts it into your Sales context for the right group. The data that reaches the model never leaves the Mac mini during inference.

POST <your workflow webhook URL>
X-Webhook-Secret: <secret shown when the trigger was created>
Content-Type: application/json

{"name": "Ada", "company": "Acme", "message": "We need a quote for 40 units"}

Workflows are new, so start with one small automation, watch it in the Jobs monitor, and grow from there. Webhooks are rate limited per workflow, which is plenty for form submissions but not meant for bulk data.

A short checklist

Memory first: 32 GB or more if the whole team will use it.
Wired network, start at login, keep awake.
Recovery key exported and stored away from the machine.
One model to start; add a second only when you know why.
Write decisions in messages, so agents can find them.

Few, but mighty teams do not need a data center to use AI privately. They need one well-chosen machine and a workspace that already knows who should see what.

Try it with your team. Start the 14-day free trial of SweetHive, set up your contexts, and connect your first agent to a model that runs on your own hardware.