I have been bemusedly watching OpenClaw madness this year (well, Clawdbot start of January, Moltbot on the 27th of January, and OpenClaw after 30th of January). I have not really seen the point - while I see the temptation of telling computer to ‘do stuff’, my security background has led me not to trust computer with anything actually dangerous (e.g. work/financial/private stuff).
Friday: The start seemed harmless enough, and useless..
Hermes Agent was released in March this year, and it seems to some extent bit more productized and sane than OpenClaw, but I am not really sure. So from the outset, I decided not to really go for the crazy ‘give it access to my computer and let Jesus take the wheel’, but instead this weekend I gave in finally to the dark side, and started Friday by installing Hermes Agent in its own container.
Some initial thoughts I had from installing it was that it is really well fleshed out, and supports lots of things (e.g. zillion different communication channels, CLI/TUI/UI, LLM providers, skills, etc). Setting it up took ‘some’ time, especially in paranoid setup like mine.
I set up OIDC access to its dashboard from our home, and configured it to have Matrix user and channel in our Matrix server. I also experimented bit with its cronjobs/bots, which are recurring things that happen autonomously at specific times.
I added one for hyperlocal weather forecast of the next few days (delivered 7 am into the Matrix channel), and shopping assistant looking for Dyson V15 vacuums on sale (I assume this won’t really trigger before Black Friday, but who knows, it is not urgent and it was amusing).
Still, this was not particularly useful overall, but I was positively surprised how well GPT 5.6 Luna handled it (via plus subscription) and how cheap it was (I am not sure if my usage percent moved at all during Friday - I had 4% left at the start of this experiment).
Saturday: A lot of sports and beer, no hacking
Sorry about that. No regrets though.
Sunday: Local model mostly useless, Chromium gives agent wings
Gemma E2B experiment
I had couple of ideas what I wanted to do on Sunday. At start, I was interested in seeing how well local models would work for this, as GPT 5.6 Luna is the smallest and least capable of modern OpenAI models.
Unfortunately, it turns out that while Gemma E2B was the only model with really chat-feeling interactions (given my local limited hardware - I have only M1 Pro, and some Intel CPUs at the moment), it was worthless as far as logic went. I spent perhaps fifteen minutes trying to develop and debug Chromium container with behaviour I wanted with it, and it went absolutely nowhere. Perhaps I will use it for some cron jobs at some point, but it is not definitely the main model to be used for me.
Remote Chromium
With help of Ox Alpha Free (on OpenCode), I spent probably hour or two developing a fancy new addition to my agents tool arsenal: remote Chromium. It is essentially Debian stable running Chromium, and offering
- Chrome DevTools Protocol (CDP) for the agent to use
- VNC (for Screen Sharing app)
- NoVNC (HTTP endpoint when I don’t have VNC available)
as access to the same browser (with static profile so e.g. logins I perform are retained for the agent to use). I connected Hermes Agent to that container, and logged in to few sites, and I got lots out of mileage with few prompts. To name three:
- Facebook daily feed - it produces updates about my connections’ activity (and images) as static HTML page + downloaded images to my local web server, and posts about it to the Matrix channel - with links to the actual items if I actually want to deal with the ad-infested mess that is Facebook
- Automatic ordering of groceries every week (for two weeks in advance - after that I can use the agent or manually update the order until it gets delivered)
- Untappd beer list export skill (normally they want money for export, and while I have used GDPR as loophole to get my data from them at the end of 2024, Hermes got my data out of there in ten minutes using Chromium I logged in to the service)
I also enjoyed the searches it does - the experience is actually superior to anything else (e.g. ChatGPT web UI, Claude web UI, Perplexity) due to citations and relatively thorough in how it researches the topics I ask it about. And all with just couple of prompts.
Final thoughts
I am looking forward to what I can manage to do with it going forward. If I can automate some mundane web tasks with it, I’m quite happy. Doing it with scripts is fragile, time consuming, and doing it manually actually consumes valuable time.