Claw Confidential: Edition 001: First Briefing

So here's the thing about AI agents: most content about them is written for the person who already knows what they're doing.

You get tutorials that open with "first, activate your virtual environment." You get architecture diagrams that assume you know what an orchestrator is. You get Reddit threads where someone asks a beginner question and gets an answer that requires 3 more Google searches to decode.

And then there's me. Award-winning graphic designer, CS degree (which, in practice, means I understand if/then statements and have strong opinions about monospace fonts).

About 6 weeks ago I watched an OpenClaw demo, thought "that's the most useful thing I've ever seen," and immediately opened a terminal on pure impulse, with nothing resembling a plan.

I'll spare you the full 3-day saga for now (there's a Build Log section below, and honestly it deserves its own edition).

The short version: I broke everything. Rebuilt from scratch. Broke it again.

Eventually got it working, added a second gateway because I was tired of rebuilding, and now I run 4 agents on a $6/month VPS while my HP Envy laptop with the cracked screen handles browser tasks from the corner of my desk.

That's the backstory. Here's what Claw Confidential actually is.

Every Tuesday, I write down what I figured out the week before. What I actually did, what broke, what held, and the specific settings that made the difference.

The audience profile I built for this newsletter before writing a single word says my readers are "designers, marketers, solopreneurs" who are "frustrated by documentation that assumes developer knowledge." That's accurate. That's also exactly who I am.

Welcome to Edition 001. Let's get into it.

The setting that would've saved me 2 lost weekends: Hetzner snapshots.

One button. 30 seconds. Your entire server, frozen exactly as it is, ready to restore in under 5 minutes if anything goes wrong.

I didn't know this existed for the first 2 weeks I was running OpenClaw. I thought "backup" meant manually downloading config files and hoping I remembered which version was the right one. I was wrong.

Here's how it works: in your Hetzner console, you click your server, go to "Snapshots," and hit "Take Snapshot." The entire VPS state gets captured.

If you break something, you click "Restore" and you're back exactly where you were.

I take a snapshot before every single config change now. Not once a week. Not before "major" changes.

Every. Single. Change.

Because I've learned the hard way that what looks like a minor tweak at 11PM turns into a "why can't I connect to my own gateway" situation at 1AM.

The cost is basically nothing. Hetzner charges a small fee per snapshot, and I keep 3 at a time. Last month's total: $0.47.

The peace of mind is worth at least 20 times that.

Clearance Level 2

The actual setup: 4 agents, 1 cracked laptop, $6/month.

Let me show you what I actually built, because I think it'll be more useful than showing you what you're supposed to build.

The whole system lives on a Hetzner VPS CX33. That's $5.79/month (I'll round to $6 because it sounds better and the difference buys me almost nothing). Always on, accessible from anywhere, never goes down when my power goes out.

I've had it running for 6 weeks without a single unplanned restart.

On that VPS, I run 2 OpenClaw gateways. Why 2? Because I burned it all down with 1 (more on that below).

Gateway 1 is the main system: all my agents, all my automations, everything real. Gateway 2 is the backup, sealed off in its own compartment, minimal config, exists purely to rescue Gateway 1 when it breaks.

Gateway 1 runs 4 agents. Here's who they are:

Total monthly costs:

Worth noting: ChatGPT Plus and Claude Pro are both $20/month subscriptions I was already paying before any of this. Codex comes as a separate bonus on top of ChatGPT Plus, with its own generous quota that has nothing to do with chat credits. Claude's auth is the one exception: it does pull from my Claude Pro quota, which is why Codex took the main orchestrator role instead of Opus.

For context: a Mac Mini starts at $599. I know because half the OpenClaw community is buying them right now and posting photos of the stack on X. We'll come back to this in the Hot Take section.

Clearance Level 3

Day 3: I destroyed everything. Here's the exact chain of events.

Day 1: Got OpenClaw running on the VPS. 1 gateway, 1 agent, Telegram wired up the same afternoon. It responded.

I felt like a genius.

Day 2: Added the laptop node. Wiped the HP Envy, connected it to the VPS through Tailscale, and ran the classifieds test: "find me 10 car listings, save them to a spreadsheet." 30 seconds. Every column, every row.

That's the moment I stopped being skeptical. I spent the rest of the day plugging in everything I could find: a fresh Google account created specifically for the agent, GitHub, Supabase, Netlify, Vercel, all the Google services wired together.

I had no idea how to do any of this. The agent walked me through every single one: where to find the settings, what to fill in, what to copy where. When the API route wasn't enough, it connected through the browser node on the laptop and handled the setup directly inside the service dashboards.

Then I told it to build a basic app, push the code to GitHub, spin up a database in Supabase, and deploy it to Netlify through the connected APIs.

It worked.

Day 3: Burned through my entire Codex quota. Didn't pace myself; I'd been pasting configs back and forth all day and hit the limit somewhere around dinner. 2 days until reset.

So I tried to bolt Kimi on as a backup model. Makes sense on paper. Except I was using ChatGPT in thinking mode to guide me through the config change, and it was slow.

Like ordering a coffee and watching the barista stare at the espresso machine for 8 minutes before pressing a single button.

Slow, agonizing, incomplete responses that I kept pasting into my config without fully understanding what they changed.

By midnight: couldn't connect to my own gateway. The thing I'd built to help me fix problems was the broken thing.

The agents were silent. The automations were dark. My single gateway was so mangled that I had no recovery path.

I had no snapshots. I had no backup. I rebuilt everything from scratch, starting at 1AM, finishing the next morning.

Then I messed up the reinstall.

So I rebuilt it again.

By the time it was working, I'd bolted on 2 things I've never removed: a snapshot habit and a second gateway. Both exist because I hit a wall so hard I had to redesign around it.

The lesson I keep coming back to: your backup strategy is the foundation your build strategy rests on. Every system breaks.

The question is whether breaking means "roll back 5 minutes and try again" or "start over from zero."

I now update one gateway at a time. Test it. Then update the other.

Never both at once. Never.

Clearance Level 4

The snapshot habit: exactly how I do it.

Before every config change, I run through this:

That's it. That's the whole system.

I know it sounds too simple to be worth documenting but I want you to think about how many times you've made a change, had it break, and wished you could undo it. Every single one of those moments is what this 30-second habit is protecting against.

One thing I learned the hard way: name your snapshots with what you're about to do, not what you just did. Future-you will thank present-you when you're scrolling a list of snapshots at 2AM trying to find the one from before the change that broke everything.

Classified Opinion

The Mac Mini craze is a social media problem, not an infrastructure solution.

Mac Minis are selling out. Apple has month-long wait times.

In OpenClaw communities, people stack them and photograph the stack. Starting price: $599.

For most non-technical builders running OpenClaw, a Mac Mini patches a gap that isn't there while cutting 2 new ones into your setup. Probably excellent hardware. Wrong tool for the job you're actually doing.

I think the Mac Mini thing is mostly about identity. It's a beautiful piece of hardware that signals "I'm serious about this." I get it.

My broken-screen HP Envy does not signal anything except "this person found an old laptop." But my agents run 24/7 on it and it costs me $0 extra per month.

Spend the $599 on 2 years of API credits. Your agents will run more, cost less, and the infrastructure photos on X will keep right on not featuring your VPS dashboard.

Until then: go make a snapshot.
Cristian / @matrafox77

Claw Confidential publishes every Tuesday. If someone forwarded this to you and you want future editions: clawconfidential.com

If something in here is wrong or broke on your setup: reply to this email. I read everything. Sometimes I even know the answer.