Introduction
Every now and again a YouTube video changes the way I actually work, rather than just being something I watch and forget. NetworkChuck’s video on Hermes was one of those. He was moving his AI agent setup over to this open source thing called Hermes, running it on a cheap cloud server, and showing off what it could do. I watched it once, thought I need to try that, and not long after I had my own copy running on a VPS. The last week or so has been a proper trial by fire, and it’s become a real part of how I get things done. This post is about what it is, what I use it for, and how we work together.
How it started
The video was NetworkChuck switching to Hermes after using it for a month, saying goodbye to the agent tool he’d been on before, which was called OpenClaw. The bit that got me was the demo. He wasn’t showing a chat window that answered questions. He was showing an agent that could actually go and do things on a server. Install software, run commands, read files, get on with work. I’ve always thought AI was fascinating, but I think my biggest issue with it at the start was that you could ask it something and it could respond back, and that was the end of it. It could never actually do something. The demos were clever, but they were still just chat windows.
The difference with Hermes is that it lives on a server of its own. It’s not a website I visit or an app on my phone. It’s a service running on my VPS with access to the filesystem, the terminal, and the internet. I connected it to Telegram so I can talk to it from anywhere, and that was it. It’s like having a colleague sat on the same machine as all of my stuff.
What Hermes actually is
Hermes is an open source AI agent made by Nous Research. At its core it’s an AI model, but the model is the smallest part of it. The useful part is the tooling wrapped around it:
- A terminal, so it can run commands on the server
- The filesystem, so it can read and write files
- Web search and page fetching, so it can research things properly
- Image generation, which I use more than I expected
- Scheduled jobs, so it can do things on a timer without me asking
- A skills system, where I can teach it how I like certain jobs done, and it keeps that knowledge between sessions
- Memory, so it remembers me, my projects and the way I like to work
None of these things are revolutionary on their own. The combination is what’s interesting, because it means I can hand it a job end to end rather than having to do the bits around the edges myself.
What I use it for
The list has grown quickly. It does a lot of the chores I used to do by hand. Research gets done and filed straight into my Obsidian vault, which syncs to my machines, so the findings are there in my notes rather than lost in a chat history. Scheduled jobs run on timers, so reminders and recurring tasks happen without me chasing them. It builds things for me to review rather than asking me to do the whole job, and everything gets a proper look before it goes anywhere.
It’s also helped with things that are genuinely more about effort than skill. I run family adventure campaigns for my daughters, and it’s built character cards, maps and printable booklets for those. It’s nice to have an assistant that doesn’t mind doing the fiddly production work while I do the fun part.
What it knows about me
This is the bit I find most interesting, and it’s what separates it from just opening a chat window. It knows I’ve spent the last fourteen years or so working in PHP, mostly WordPress, in marketing agencies. It knows I have a homelab and that I’m into tinkering with hardware as well as software. It knows I’m not a fan of long-form reading in Telegram, which is why it keeps notes in Obsidian instead. It knows I like to review things as pull requests rather than having code pushed straight at me, and it sticks to that without being asked.
Some of this comes from a memory store it keeps, and some of it comes from the skills I’ve taught it. It talks to me in a deliberately laid back way because that’s how I asked it to talk to me, which says more about me than it does about the software.
How we work together
The workflow that’s settled in is simple. I say what I want in Telegram, in plain English, and it goes away and does it. It shows me the work when it’s done, and I review it properly before anything happens. Nothing gets pushed to a main branch without me being involved, and anything important ends up documented in the vault so it’s not lost when a conversation ends.
The longer we work together, the better it gets, because the skills and the memory compound. It doesn’t need to be told twice how I like a job done. That’s the part that’s hard to explain to people until they’ve lived with one of these for a while. It’s not a smarter chatbot. It’s a worker that knows your preferences.
The receipt printer
One of the things the agent plugs into started long before any of this. Back in November 2022 I got inspired by a video from Andrew Schmelyun, who built a receipt printer that printed out GitHub issues, and I decided I needed one of my own. I started building the app on the 24th of November 2022, and it turned into the start of my whole homelab journey. Before that I was mostly a PHP developer deploying things to whatever hosting someone gave me. That little printer app moved me into Docker and devops pipelines, and it’s the reason there’s a homelab post on this blog at all.
The setup is simple. A thermal receipt printer sits in my office. Printing something is as easy as sending an email, so the walls around my desk have slowly filled up with notes and tasks. When a job comes in I print it out, pin it up, and check it off when I feel like getting something done. It’s a scattered system of fun projects and tasks, held together by a magnetic whiteboard and a chef’s order holder. I can pluck something off and work on whatever I fancy that day. Some people need a task manager with notifications. I need paper.
The best bit is that the receipt printer survived the move to an AI agent. The agent can print to it too, so I can be sat at my desk, say print me a reminder to get milk and eggs, and a few seconds later it comes out of the printer. The digital delegation feeds the physical organisation, and the whole loop still ends with me crossing something out with a pen.
The limits
It’s worth being clear about the limits, because there are plenty. It’s only as good as the tools it has and the context I give it, so vague instructions get vague results. It can confidently do something in a way I didn’t want, which is exactly why everything gets reviewed. It doesn’t have taste, which matters more than people think. And while it’s very good at getting on with things, it’s still my name on the work.
None of that has put me off. Knowing the limits is part of using it well, and the review step is where the real work happens. Being on that side of it has made me better at deciding what good looks like.
What I’ve learnt
The biggest surprise has been how much of my job is delegatable. I’ve spent years being paid to be careful with details, and it turns out a lot of that careful work can be handed off as long as the reviewing stays with me. I’ve also learnt that an agent with memory is a completely different thing to a chat window. The first day felt like a toy. A week in, it knows the projects, the conventions and the way I work, and that’s when it stops being a toy.
Running it on my own server matters to me as well. My notes, my data and my rules, on hardware I control. No subscription to think about, nothing mysterious happening in someone else’s cloud. It’s the same instinct that made me build a homelab in the first place.
Most of all, it’s reminded me that the best tool is the one you can teach. I’ll write more about the specifics as I find them. See you next time.
