I've been using Hermes a lot lately. Especially after reading Duarte's post

At first it was interesting mostly as a tool. Another AI agent, another interface, another way to ask models to do things.

I've tried Openclaw, Nanoclaw and many others. But none of them clicked with me. I couldn't do anything really useful with it.

But recently I've realised how useful it can be when I asked it to help me with some maintenance and updates on my Raspberry Pi server, where I run several apps, including Immich.

Recently I asked Hermes to help me back up all my photos from Immich to a storage box and then help me update Immich to version 3. That is not a toy task. That is my real data, my real server and something I did not want to mess up.

And it worked amazingly, surprisingly, well. With no hiccups, nothing wrong. Of course, I also felt safer because it could stay with me through the whole process. Checking things. Preparing steps. Watching long-running tasks. Helping me think about risk before doing anything dangerous. That was the moment where Hermes stopped feeling like a demo and started feeling like something I could actually rely on.

I've mostly used it with Codex. But I've also used it with local LLMs and occasionally with other models through OpenRouter. What matters less to me is the exact model and more that Hermes gives me a practical way to work with them. It lets me bring the same kind of help into different parts of my digital life.

I've also used Hermes for software and infrastructure work. To build things for Town Square and my personal website, but also to help with operational work around them. Deploying to the Hetzner server, investigating issues and checking what is happening directly on live data.

Another important aspect is writing, notes and authorship. Hermes sometimes review my blog posts and point out weak or confusing parts. It also has access to my Obsidian vault with my personal notes, planning, projects, etc.

But controlled authorship is an important point.

I still do 90% of the changes in my notes myself. I do not want an AI silently rewriting my thoughts and mixing its voice with mine. But sometimes I do ask Hermes to change a few things. Maybe clean up grammar. Maybe reorganize a paragraph. Maybe review a draft and tell me where the weak points are.

Whenever it touches a note, it adds the tag AI_coauthored. It makes the collaboration explicit. The note is still mine but clearly shows when AI touches something.

I've been thinking that maybe this should have more levels in the future. Not just one tag, but maybe two or three. One for small grammar fixes. Another for suggestions or review. Maybe another for cases where the AI had a more substantial co-authoring role. I have not defined those tags yet. But that is the direction I want to go. Not all AI help is the same and I think it is useful to make those differences visible.

That is probably the main thing I like about using Hermes. It is not just automation, it is negotiated collaboration. I decide where I want help and how much. And I can keep a clear line between my thinking and the machine's contribution.

So for me, Hermes became real not when it wrote text or generated code. It became real when I trusted it with things that mattered and when I found a way to include it in my workflow without giving up authorship.

That balance is what makes it useful. Not full automation, no blind trust. Just a very practical kind of collaboration.