When OpenAI triumphed with its AI chatbot, a large slice of our applications migrated toward that format. Now that agents like Claude Code or Codex are in vogue, they have become the benchmark to look at. Spark, from Google, is one of the latest arrivals: the company makes it clear that Gemini must dominate the personal experience beyond the assistant. Yet I have put a better idea into practice: my own local agent.
I have been flirting with local agents for several months since OpenClaw helped popularize them. I think it’s the best solution to a double problem: save money with a solution you run on your computer and avoid sending my personal data to a cloud I don’t control. The local agents I built kept dying just as my Tamagotchis did in the golden era of virtual pets. Until I tried Hermes.
Hermes is the local agent I didn’t know I needed
I can’t praise this open-source project enough; I think it’s the best thing out there right now for building a local agent from scratch. It can run on Linux, on Windows, on a PC, on a Mac, on a phone… With cloud API or with 100% local models. It even allows configuring routing so that Hermes picks the model depending on the task. It’s precisely how I’ve set it up: the group of agents and subagents work on their own to achieve the result, spending as little as possible. Like The Matrix, but without Skynet from that other movie.
Cost is one of the keys when using agents. Companies like Google will push their autonomous tools, albeit under their own paid umbrellas. In the case of the latest Spark, Google requires a Gemini Ultra subscription. Claude Code and Codex can run on free accounts, but with extremely limited usage. Result: if you want a capable agent that won’t be throttled, set it up yourself.
Hermes is free, open-source, easy to install, 100% autonomous and somewhat dangerous, because it can figure out how to do (almost) anything. One of the possibilities could be to upend the computer, so precautions are necessary. In my case, I run it on WSL 2 with Ubuntu on my secondary Windows PC. What I use to interact with it is my phone.
Hermes carries risks. For my use, I think the advantages far outweigh the drawbacks: it has become my software creator, my task manager, my calendar, my feed with a daily summary of sources… Just ask for anything and you’ll have it minutes after it has hunted down the path to learn the skill. After testing Spark, Claude Code, Codex and other agents, I stick with Hermes for its enormous versatility. Not just for that.
Memory stays on my computer
The amount of data that companies such as Google, Anthropic and OpenAI are collecting is enormous. As agent usage becomes more widespread, that data capture will grow because these tools need a deeper understanding of the user to be conclusive. With Hermes I keep all that data locally.
The agent’s memory stays on my computer. Just like all its skills, the configuration files, the access data, the APIs, the conversation logs… With a Gemini-like agent all of that would be on Google’s servers, and would have served as training data for the agent. Keeping my knowledge locked behind a key is a guarantee for my privacy.
Yes, Hermes has risks: it could happen that someone taps into my PC and obtains the information. I try to keep it as local as possible and run security audits on it. If all that data were in the cloud of a company, I wouldn’t know what happens to it, and I’d feel more uneasy not having the key.
Once conversing with a chatbot becomes a habit, the next step will be to offer every user an agent. We’ll have choices, and also alternatives. I opt for my own agent: the existence of open-source tools like Hermes feels fortunate. And I think we should take advantage of them: I’m getting more and more value out of it every day.
Cover image | Iván Linares