James Read's Code, Containers & Cloud blog

This is a brief snapshot post — mostly for future-me to look back on — about my early success with Hermes Agent. I’ve only been running it for 9 days, so treat everything here as first impressions, not a long-term review.

I tried OpenClaw in the past. It was interesting — I even mentioned it when writing about AI autonomy levels — but it never quite clicked for me. Hermes feels different so far: it seems to understand what I’m asking for more often, with less re-explaining and less “yes, but that’s not what I meant.”

I’d like to add that I first learned about Hermes from a colleague - Stephan Michard - he’s got a got a really interesting blog that I’ve started following, please do check it out.

How I’m running it

I’m talking to Hermes over Telegram, which is exactly how I want a personal agent to work — ping it from my phone, get a useful answer back, move on with my day. The gateway runs as a background service on one of my machines, installed with Hermes’s built-in systemd integration (hermes gateway install, then the usual systemctl / journalctl workflow). That matters to me. For something I expect to rely on daily, I want the same operational model as the rest of my homelab: enable on boot, check status with standard tools, read logs when something breaks.

The model backend is ChatGPT via a paid platform.openai.com subscription — the model is just 5.5, but that’s not necessarily going to stay like that. What is stable is the bill: costs are pretty expensive so far. Useful, but not cheap. I’ll be watching usage closely and may look at routing some workloads to cheaper models later; Hermes supports switching providers without rewriting everything (hermes model), which is reassuring.

Another option is Cursor, which I’m paying for now. It seems to legitimately support API usage like this as part of the plan - so far all my Cursor usage is via the IDE and CLI Agent. I need to check this a little more as I don’t want to run afoul of cursor’s terms of service - much like many people were doing with OpenClaw and similar AI IDEs.

Use cases so far

Nothing here is production-critical yet — it’s a toy that might grow into a tool — but the early wins are real:

• Task tracking with WackyTracky — asking Hermes query, check and summarise what’s on my plate instead of opening another app.
• Calendar with SickRock — checking what’s coming up, and nudging events around when plans shift.
• Managing both together — “what’s due this week and what clashes with the school run?” is the kind of cross-system question I actually ask in natural language.
• Website suggestions for lanlist.info and technowax.net — layout tweaks, content gaps, things I’ve been putting off because I’m too close to the sites.
• Daily tech news, tailored to me — not a generic firehose; prompts shaped around my stack and interests (OliveTin, self-hosting, Fedora, the usual).
• LAN events for lanlist.info — hunting for upcoming LAN parties and community events worth listing.
• Triaging GitHub Security Advisories for OliveTin — with a custom skill wired up similarly to how I fetch GHSAs elsewhere; early tests look promising for sorting noise from things that actually need a patch this week.

What I’m watching

Hermes is agentic in the real sense — terminal access, tools, skills, scheduled jobs — so the usual cautions apply. It’s mostly got read access to my calendar/tasks for now, which is enough to make it useful. It’s living on it’s own VM that I can kill off quickly if I want to, as well.

The cost curve is the other thing on my mind. A capable model on every “quick question” adds up fast. I’ll report back if I find a sustainable split between “cheap model for triage” and “smarter model for multi-step work.”

Summary

Too early to declare victory, but Hermes + Telegram + systemd already feels like a better personal agent setup than OpenClaw did for me — mostly because it seems to grasp intent without as much hand-holding. The OpenAI platform bill is the main downside so far.

If you’re curious, start at the Hermes docs and the messaging gateway guide. And if you’ve migrated from OpenClaw yourself, Hermes even ships hermes claw migrate — I didn’t need it, but it’s a nice touch.

I’ll post again once I’ve had a few months of real usage — or once the API costs make me rethink my life choices :-).