I’ve neglected this blog for a while. Sue me.
I had also neglected my receipts, which turned out to be more relevant.
For a long time, I had been interested in AI agents in the abstract. I understood the general idea: unlike a chatbot, an agent can use tools, interact with files, call services, and carry out a sequence of actions.
But understanding the idea and actually setting one up are different things.
So I decided to build my first one using Hermes, by Nous Research.
The plan was modest. I wanted to run it on a VPS, connect it to some of the tools I already use, and see whether it could help with a small amount of recurring admin.
I expected the difficult part to be designing the workflow.
I was wrong.
The first setup
I chose a VPS in Europe (s/o to Contabo) and spent a few hours setting everything up. Docker was involved. There were installation commands, configuration files, authentication steps, and the usual feeling that I was one typo away from having no idea what was happening.
Eventually, I got Hermes running with DeepSeek V4 Flash through OpenRouter. It could talk, take instructions, and use tools.
Then I hit a dumber problem. Europe was slow to get the newest models. I wanted Grok 4.5, and that basically meant starting over on a US server: reinstall the stack, redo the config, move the setup I’d just finished. Annoying. Also very on-brand for “the hard part is not the workflow.”
Once the US box was up, it was good enough to stop theorising and give it a real job.
The first real job
That job was the receipts I’d been ignoring.
I pointed Hermes at a folder of messy PDFs and photos and asked for something simple: merchant, date, amount, currency, and a clear flag when something looked off. No full accounting system. Just enough structure that I could review a list instead of opening files one by one.
It worked better than I expected. Most of it was seamless. Foreign currency was fine. The boring extraction work, the part I thought would be fiddly, was not the hard part.
Setup was.
Getting Hermes onto a VPS, picking the right region for the models I actually wanted, wiring tools and permissions, then doing it again when Europe wasn’t enough, took longer than the receipt automation itself. Once that was in place, the workflow was almost straightforward. It can also run on a schedule via cron, so the boring pass doesn’t even need me to kick it off each time.
I’m still careful about what I hand over. But this was enough to get me hooked. The install was the tax. The next builds are the fun part, and I’m very ready to find out what else is worth automating.