Guide · 5 min read

Turn a Mac mini into an AI agent server in 5 minutes

A Mac mini is a near-perfect home for AI coding agents. It's quiet, it sips power, and it can run your agents around the clock without tying up your laptop. Here's how to get Claude Code, Codex, and OpenClaw running on one, each in its own sandbox, reachable from anywhere on your network.

What you need

• A Mac mini. An M1 from 2020 is plenty; an old Intel one works too.
• About five minutes of your attention. The installer does the slow part, downloading Docker and building the binaries, on its own while you do something else.
• A Tailscale account, so you can reach the mini from your laptop without opening any ports. Free for personal use.
• API keys for whatever agents you want to run. You bring your own; Dejima never holds them in a cloud.

You can do all of this sitting at the mini, or over SSH. After setup you won't need to SSH in again.

Step 1: install Dejima

One command. Run it on the mini.

curl -fsSL https://dejima.tech/install.sh | bash

This isn't a toy script that drops a binary and wishes you luck. It checks prerequisites and names the fix for anything missing, installs Docker if you don't have it, builds the dejima and dejimad binaries, and registers the background service as a launchd unit so it comes back after a reboot. When it finishes, the daemon is running.

Step 2: make your first island

An island is one container with its own home, credentials, and network. Your agents live inside; your Mac mini's files stay outside. Point one at a repo.

dejima init --repo git@github.com:you/your-project.git

That clones the repo into a fresh island and seeds it with a Claude Code agent. Working on something you haven't pushed yet? Use --local-copy ./path to copy it in instead.

Step 3: add the agents you want

One island can hold several agents. They share the workspace and tool logins, but each gets its own git worktree, so they don't step on each other. Add a couple.

# a second terminal agent: Codex
dejima agent add your-project --type codex

# a headless one: OpenClaw, running on its own
dejima agent add your-project --type openclaw

Terminal agents are the ones you attach to and watch: Claude Code, Codex, or a plain shell. Headless runtimes like OpenClaw, Letta, and Hermes run on their own and stream output to dejima logs. Mix them however you want, and bring your own keys for each.

Step 4: run it

Start the TUI. It's the dashboard for everything on the box.

dejima

You get one screen listing every island and every agent, with status, memory, and CPU. From there you start, stop, hibernate, or attach to any of them. To drop into an agent's live session:

dejima connect your-project   # Ctrl-b d detaches and leaves it running

Step 5: reach it from your laptop

This is where the Mac mini earns its keep. With both machines on your tailnet, point any Dejima command at the mini by setting one variable.

DEJIMA_HOST=mac-mini.your-tailnet.ts.net:7273 dejima connect your-project

Close the laptop, walk away, open it later, reconnect. The session was running on the mini the whole time. The same works from a second device, and two devices can attach to one agent at once.

Leaving it running

The daemon is already a launchd service, so it survives reboots on its own. A few commands worth knowing for day-to-day life:

• dejima hibernate your-project frees resources when an island is idle; dejima wake brings it back.
• dejima logs your-project tails what a headless agent is doing.
• dejima purge your-project destroys the island and its volumes when you're done. Push your work out first, since the island's volume goes with it.

That's the whole thing. A quiet box in the corner, a fleet of agents that can't wander off it, and one command from your laptop to check in.

Common questions