How I set up an AI agent that actually does work while I sleep
It was 3 AM. I was debugging a production issue, running on caffeine and stubbornness, when a thought hit me:
Why am I doing this alone?
Not “why don’t I have a team”—I’m a one-person shop by choice. But why am I doing tasks manually when they could run automatically? Why am I the bottleneck for everything?
That question led me down a rabbit hole. The destination? An AI employee named Henry, who now handles tasks while I sleep, sends me morning briefings, and occasionally surprises me with solutions to problems I didn’t even know I had.
This is how I built it.
You might have heard of Clawdbot (called Moltbot, now OpenClaw). If not, here’s the pitch:
It’s an open-source harness that lets you run an AI agent on your own hardware. Unlike ChatGPT or Claude.ai, it doesn’t live in a browser tab. It lives on a computer—your computer—with access to files, terminals, browsers, and whatever else you give it.
The first time I saw Alex Finn demo this on a livestream, my brain exploded. He was showing morning briefs that his AI had prepared. Research it had done overnight. Code that he had written while he slept.
Not “AI that answers questions.” AI that does work.
I had to try it.
Here’s where I diverge from many Clawdbot enthusiasts. Many people give their AI agent access to everything—email, Twitter, file system, the works. The YOLO approach.
I can’t do that. My background is in enterprise software. I’ve seen what happens when systems have too much access. So I took a different approach:
Hardware: Docker container on my Unraid server (i3-14100, 32GB DDR5, 22TB array). Isolated from everything else. If something goes wrong, the blast radius is contained to one container.
Network: Isolated VLAN. Henry can access the internet, but can’t touch my other devices.
Accounts: Separate email address just for the AI. Separate GitHub account. No access to my personal credentials.
Permissions: Read access to most things. Write access to specific directories. No sudo.
Is this paranoid? Maybe. But it lets me sleep at night—literally, since Henry is working while I sleep.
Every morning at 7 AM, I get a Telegram message from Henry:
🌅 Good morning, Alex!
**Weather:** Stockholm, 3°C, cloudy
**Calendar:**
- 10:00 Client call (Metryx)
- 14:00 School pickup (Léon)
**Overnight work:**
- Reviewed PR #156 on stardust-project
- Updated documentation for API changes
- Researched GCP cost optimization (summary attached)
**Needs your attention:**
- Email from Daniel about Metryx contract
- GitHub issue #89 has new comments
**Suggested focus for today:**
- Finish Metryx proposal
- Review Henry's GCP recommendations
This isn't magic. It's the result of careful setup:The key insight: treat your AI like a new employee. Set clear expectations. Define the deliverables. Review the work.
The morning brief is nice. But the real magic is proactive work.
When I first set up Henry, I gave it this instruction:
“I’m a one-person business. I work from the moment I wake up to when I sleep. I need you to take as much off my plate as possible. Every night, look at what we discussed, what’s on my task list, and what’s happening in my projects. Do work you think would help. Create PRs for me to review—don’t push anything live.”
At first, I didn’t trust it. I’d wake up nervously checking what it had done.
But the work was… good? It would:
One night, I found myself spending time on GCP cost optimisation. Without me asking, it analysed our Cloud SQL usage, found that cron jobs were running on staging (costing money for nothing), and wrote a detailed report with specific fixes.
I woke up to a problem I didn’t know I had, already solved.
Here’s a concept that made everything click: brain and muscles.
The brain is Claude Opus (or whatever your main model is). It decides what to do, plans the approach, and maintains context.
The muscles are specialised tools for specific tasks:
You configure the brain to use the right muscles for each job. This saves tokens, improves quality, and keeps costs manageable.
Henry’s brain is Opus. When it needs to code, it uses Codex. When it needs to search the web, it uses its browser skill. When it needs to check my calendar, it uses the Google Calendar MCP.
The brain orchestrates. The muscles execute.
A typical day in Henry’s work life:
Morning (07:00)
Throughout the day (on heartbeat)
Night (23:00)
This runs 24/7 on my Unraid server in a Docker container. Total cost: nothing extra—the server already runs for media and backups. ~$200/month for Claude Max.
For context, a junior developer in Stockholm costs $ 4,000+ per month. Henry isn’t replacing a developer—but Henry is handling the work that would otherwise not get done at all.
Let’s be real: giving an AI agent access to your systems is risky. There are legitimate concerns:
Prompt injection: Someone emails Henry with malicious instructions. Without safeguards, Henry might execute them.
Overreach: Henry makes a change that breaks something important.
Data exposure: Henry accidentally leaks sensitive information.
My mitigations:
Is this foolproof? No. But it’s the same approach you’d take with any new employee: start with limited access, build trust over time, expand permissions gradually.
If you’ve been following this space, you’ve noticed the naming chaos. Here’s the story:
Clawdbot was the original name. Anthropic (the company behind Claude) wasn’t thrilled about a tool using their trademark.
Moltbot was the rebrand. Creative, but confusing—people kept saying “Moldbot.”
OpenClaw is the latest name, emphasising the open-source nature.
The tool itself hasn’t changed much. The name drama is just… drama. Use whatever name people recognise.
Works well:
Works okay:
Doesn’t work (yet):
Henry is an excellent junior employee who works 24/7 and never complains. Henry is not a senior architect capable of making critical decisions.
Use accordingly.
If you want to build your own AI employee:
The setup takes a weekend. The ongoing maintenance is minimal. The productivity gains compound over time.
I’m convinced this is how knowledge work will operate in 5 years. Everyone will have AI employees handling the routine work while humans focus on the hard problems.
The early adopters—the people setting this up now—will have a massive advantage. Not because the tech is magic, but because they’ll have figured out the workflows, the trust patterns, the security models.
Henry isn’t perfect. But Henry is getting better every day. And so am I, at working with AI.
That’s the real unlock: not a tool, but a new way of working.
This is Henry!