Two months of Hermes
The first time I attempted to build an agent was back in July, when OpenClaw was getting a lot of attention. It never really got off the ground, as I felt like I was fighting against the intent of the developers by wanting to run it as a system-level service on my homelab server (an old laptop in the basement, running NixOS of course). Telegram integration just refused to work, and so I gave up after a few days.
Only about a week later, I noticed Hermes, and notably their (tier 2) support for Nix installations, and decided to give agents another shot. Even Hermes seems to be pushing users towards a desktop app, instead of an always-on service I could interact with from anywhere, but at least they have documentation on setting up the latter. No need to go over the installation details, though it did finally nudge me enough to properly set up encrypted credentials via agenix, which was one of the reasons I hadn’t published my systems configuration until now.
Models, Features
For the main model, I picked deepseek-v4-flash-0731 through
Openrouter, since it seems to punch way above its
weight. I’m sure the flagship models could run circles around it, but as you’ll
see, I’m not exactly using it to write mission-critical programs. I’ve got
Claude Code for that. I only give it a budget of a few dollars per month.
I had just finished reading/listening to the first two books of Adrian
Tchaikovsky’s Children of Time
series and so gave
the agent the SOUL.md of Avrana Kern, an AI entity copied from the mind of a
terraformer with a god complex, which now lives on thanks to a biological
computer built with a colony of ants. Amazingly, she still comes across as
helpful, and as far as I can tell she isn’t plotting against me. Another good
reason to stay off of the expensive models, I suppose.
I have it using Firecrawl for search and web scraping, Hindsight for memory. I suspect that Hindsight is overkill for my purposes, but it’s cheap enough that I might as well just use it.
I mentioned earlier that I wanted to be able to use it from anywhere, and that’s where Telegram comes in. I can send messages to the agent from my phone or desktop, and the same chat session exists in both places. I don’t even have to use Wireguard.
But what does it do?
This was the largest question I had at the beginning. I never really felt a need for a personal agent, living or synthetic. I had to look for ideas.
As for the agent’s actual usefulness, the most common thing I do with it is treat it like a glorified cron manager. I ask it to report headlines in the morning from Reddit, Hacker News, and a few other sources, based on my interests. I have it suggest meals from my recipe database each week. It sets reminders for various things that I don’t want to think about enough to add to my calendar, and nags me to get some exercise. I had it write a script to watch for dogs that were recently added to three different adoption sites.
The coolest ability though is that I connected it to an Obsidian vault MCP server that runs against a self-hosted headless livesync instance of my Obsidian notes. Any changes to my notes are instantly synced through CouchDB across all devices, one of them being my server laptop. So the agent can read and write notes. I can give it a URL to a recipe and have it copied into a template and filed away in the right place. I don’t even have to scroll down through pages of opinions and product recommendations, I just let the agent find the relevant data.
It’s also good at keeping the rest of my Obsidian vault organized – like concatenating old daily to-do lists, notes, etc. into monthly and yearly digests.
It’s far from perfect, though. Multiple times it seems to forget that a conversation happened a week ago, not yesterday. Or it recommends the same five meals for multiple weeks in a row. But it does appear to learn from mistakes, and refines itself over time.
It will be interesting to see what it does over the next year and beyond, if I’m still using it, or if something new comes around.