Messing around with Hermes Agent
Weblog notes from June 14, 2026 that were compiled and shared
Last weekend was great. It was one of those times of rest and reflection. Sometimes you just need that bit of time to sort of reset. Today I’m actually going to go golfing, which should be interesting. I hit golf balls at Top Golf throughout the year, but I rarely make it out to play 18 holes of golf. We will see how the process of actually playing golf ends up going. Right now I’m spending some time looking at OpenClaw vs. the Hermes Agent and trying to decide if now is the time to install the Hermes Agent on my MacBook. My initial evaluation involved checking out the Nous Research website and the associated GitHub code they shared for this project [1]. They have a desktop version that is compatible with my MacBook. The GitHub project has 192,356 stars right now, which is pretty impressive for an open source project [2]. People are definitely engaging with the project, and it seems to have been forked 33,535 times, which is interesting.
The initial download is pretty small, but it requires a bunch of installation and download. I don’t plan on giving it any keys or anything. During setup, the configure API keys and settings section was not really used. Right now I just want to kick the tires and see how it works running on my MacBook. I listened to the Intelligent Machines podcast this week, which made me question my donut choices, and it had Jeffrey Quesnelle as a guest [3]. You may know that name from the “YaRN: Efficient Context Window Extension of Large Language Models” research paper [4]. You may recall that back in 2023, this research helped push things from a context of about 4,000 words to 30,000 and opened the door to where we are now. That might not sound like much, but it was a huge unlock and a great open source contribution.
The download process was pretty easy, but I ended up subscribing to the base-level plan for $20 per month. We will see how fast the Hermes agent runs out of tokens, I guess. Things opened up, and I’m running Anthropic’s flagship Claude Opus 4.8 model to get things going. My first inquiry for Hermes was to find out if it can run a model locally instead of using cloud tokens. This got things started out by looking at my system and suggesting, based on my MacBook Air with 24 GB and an M3 chip setup, the Qwen2.5 14B model using Ollama to run locally. So far, things are moving along pretty quickly, and I’m starting to use the Hermes Agent without any trouble. The process has been pretty easy. At the moment, I’m just waiting for the ~9 gigabytes of the Qwen model to download. I have tried a few different models, and oh my goodness, did it make my MacBook Air run hot locally. This is the only time I have actually seen my fanless MacBook get hot.
Sadly, I spent my allotment credit during the first day. It seems like my best method to try out Hermes Agent on my MacBook is to use the cloud options and just accept that this experience is going to cost me a few extra bucks this month. I think running locally on my regular desktop computer would be fine. I have a good graphics card and 128 gigabytes of memory. I knew going into this that my MacBook Air was not the best device to test out a locally running model, but I was curious how it would end up going. My initial take is that the Hermes Agent is pretty powerful, but it really does want to have access to the machine, and so far, using the Nous portal models has made me less nervous about any key sharing. In terms of security, it seems like a better setup to reduce risk.
Footnotes:
[1] https://hermes-agent.nousresearch.com/desktop
[2] https://github.com/NousResearch/hermes-agent
[3] https://twit.tv/shows/intelligent-machines/episodes/874?autostart=false
[4] https://arxiv.org/abs/2309.00071 or https://arxiv.org/pdf/2309.00071 for the PDF