About Hermes

I want to take stock of my experience with Hermes so far. I don't have a spare Mac mini, and I don't want to keep my desktop running 24/7. I'm also still paranoid and not ready to grant access to my main computer or any email account. For now, when choosing between capabilities and security, I choose the latter. That leaves its mark.

I run Hermes on a $6-a-month DigitalOcean VPS. OpenClaw used to die in that setup, so I had to keep a $12 server. OpenClaw was generally falling apart before my eyes, so it never worked out for me. Hermes is much more stable.

At first, I used the claws in Telegram, but that didn't work for me at all. It only became livable after I moved to Discord. The main reason is that every message there creates a thread inside a channel, so the entire conversation doesn't get dumped into the context as it did in Telegram. Splitting things into topics alone isn't enough.

I still don't have many useful use cases. The best ones involve having the model do something on a schedule. Summarizing news from the internet is the dumbest thing you could come up with for that. I want to give this thing some real work—something you could do manually, but why would you?

One of my recurring tasks, for example, is monitoring new posts in my Telegram channel and preparing PRs that add selected ones to my Hugo blog repository, translate them into English, and handle a bunch of other nuances. Another time, I was moving a domain from one registrar to another and asked it to check WHOIS every hour until the data changed, notify me, and then stop. Incredibly convenient.

Monitoring the schedule of pub quizzes in Almaty and sending you only certain topics from certain organizers is great. Watching Telegram channels for music festival lineup updates and sending only the ones that should interest you based on your Last.fm profile—yes. Completely ridiculous, highly individual wishes.

Without Hermes, most of these cases could be coded by hand—or vibe-coded. But some examples make it clear that these scenarios are sometimes temporary and not worth building at all. Here, the questions of how to host all of this, how to connect an OpenAI subscription, and how to deliver the result have already been solved for you once and for all. That's its main value. The key rule I've developed is not to spam me with garbage for no reason. The last thing I need is another thousand unread messages, now in Discord too.

All the use cases without cron that I trigger manually are still dumb and very boring. They could be handled by projects in the basic ChatGPT or Claude. Things like extracting key points from a stream-of-consciousness voice message or calculating how much money I owe each person from a photo of a receipt. Or, at best, creating a TickTick reminder through MCP. Absolutely nothing special.