The Sovereign Workstation: Building an AI-First Setup from Scratch (Without Breaking Everything)

For a long time, I did not have a personal computer.

It sounds like a strange admission in 2026, but it was just never a priority. Then I started taking AI seriously, and that changed fast. I wanted to build agents, experiment with local models, and set up a space to vibe-code freely without any guardrails on what I could install or try. A shared or borrowed machine was not going to cut it.

The fix was simple in theory: buy a dedicated, privacy-first machine and build it entirely from scratch.

The last time I touched a Mac, I was probably playing Oregon Trail. My return to the ecosystem was not born out of nostalgia. It was born out of a specific checklist. After running the numbers on a high-powered PC versus a laptop versus a Mini, the Mac Mini M4 with 24GB of RAM was the clear answer.

This is the honest account of what that build actually looked like, including the keyboard that briefly convinced itself it was Turkish.

---

Why the Mac Mini M4?

The choice came down to how Apple handles hardware. These machines are inherently graphics heavy, which makes them work incredibly well with AI tasks.

For a budget of around 10,000 DKK, the Mac Mini M4 with 24GB of RAM was what I went for. Because of the way Apple chips are designed, the system can run local AI models much more efficiently than a standard PC where the memory is split up. I skipped the 16GB version because RAM is the ceiling for AI. Once you hit that limit, your project stalls.

---

Function Over Hype: The Application Stack

I am not interested in a default setup. I want tools that serve a specific function while keeping my data private.

• Superwhisper: Since I prefer talking over typing, this is my primary interface. It runs the transcription models locally on the M4. This means my voice never leaves the machine, which is a massive privacy win.
• Claude Desktop: This is my primary command center for blogging and vibe coding. The app includes a desktop-specific interface that is always a quick keyboard shortcut away. I actually downloaded this directly from the webpage rather than using a package manager.
• Obsidian: This is essentially a stepped-up version of OneNote. It stores everything as plain Markdown files on my own hard drive rather than in a corporate cloud.
• Raycast: The AI assistants suggested this as a replacement for the standard Mac search. I am still exploring the specific use cases to see if it earns its spot in my daily workflow.
• Homebrew: This is another tool the AI recommended for installing technical packages like Claude Code. I am still figuring out exactly how much I need it for my specific workflow.

---

The Friction: When the Machine Fought Back

This was not a 15-minute miracle. The hardware and software had their own ideas.

The Keyboard Struggle

The biggest blocker was a simple keyboard. I tried using standard PC keyboards, but the Mac insisted they were Turkish layouts. Every time I tried to type the ”@” symbol for a prompt, I got a quotation mark or a weird accented character.

I eventually gave in and bought the Apple Magic Keyboard with Touch ID. It worked instantly. The fingerprint sensor turned out to be a great quality-of-life upgrade for authorizing app installs without typing a password 20 times a day.

The iCloud Sync

Obsidian is great for privacy, but syncing it to an iPhone requires a very specific folder structure in iCloud Drive. If the files are not in a folder named exactly “Obsidian,” the phone simply will not find them. It is a minor technicality that causes unnecessary friction if you do not know the exact path upfront.

The Prompting Reality Check

I hit 75% of my weekly Claude usage limit before the setup was even finished. While this includes work from other projects, it was a clear signal. I need to be much more mindful of how I am asking things and how I am creating my prompts to avoid burning through my usage too quickly.

---

Current State

The workstation is live. I have a machine running hardware I own, using tools that keep my data on my own desk.

The next step is getting deeper into local models and seeing how far I can push this voice-driven process before it breaks. I am still figuring it out as I go.

---

The Stack at a Glance