- Joined
- Apr 17, 2018
- Messages
- 26,097
Hermes Agent vs OpenClaw: what I learned
I started this thread to see how Hermes Agent compared to my OpenClaw setup. Since then, I've posted updates about different models, costs, things that worked, and things that broke
Here is the short version so you don't have to dig through all the replies.
My setup and costs
For my original Hermes test, I used a $10/month Ubuntu server on Vultr and connected it to OpenRouter. I started with MiniMax 2.5 and then switched to MiniMax 2.7. During that first month, I used about $15 in OpenRouter credits.
My OpenClaw agent was running on a Mac Mini with Opus 4.6. It was my main assistant, while Hermes was something I was using for smaller projects and testing other models.
These were different setups doing different amounts of work. This wasn't a same-model, same-task benchmark, so those costs don't prove Hermes is always cheaper than OpenClaw.
Your model, how much you use it, and the work you give your agent all affect the cost. I've also changed models since the original post, so treat the prices and model choices below as a record of that test rather than a current pricing guide.
What worked well with Hermes
Getting it running on the server was simple. Using OpenRouter also made it easy to try different models without rebuilding the whole setup.
I used Hermes and MiniMax to build an Arc Raiders website. It was a useful way to test coding and tool calls on an actual project rather than just asking the agent questions.
Later in the thread, I also shared updates about using OpenAI models and a Hermes Web UI. I haven't limited my testing to MiniMax.
What went wrong
A few hours after posting how happy I was with the results, my agent broke parts of the site when I asked it to add the JavaScript for ProPush
Another member also pointed out that some of the game information was wrong. Fair criticism. Getting an agent to build a website and getting it to publish accurate content are two different things.
The lesson I'd already learned with OpenClaw applied here too: have your agent manage code through GitHub so you can review changes and go back when something breaks. Check the content it creates too.
Which one would I use?
At the time of my original comparison, OpenClaw was my main daily assistant and Hermes was my option for smaller projects and testing models. Hermes turned out to be a viable alternative, and I've continued using it.
I wouldn't pick either one based only on a cheap model working well for someone else's project. Try it on something useful to you, check the results, and see how much fixing you need to do.
Using a frontier model is best. I am using GPT Astra on both Hermes and OpenClaw now. My OpenClaw is my main personal assistant and my Hermes agents are mainly focused on specific projects.
If you want to try Hermes for affiliate marketing, I put together a separate guide showing my server setup and how I used it to build a tracker and connect to Adcash:
How to Setup a Hermes Agent for Affiliate Marketing
The original post is below, followed by the updates and discussion. Feel free to share what you're using and how it's working for you
My original Hermes Agent test
A few weeks ago, I installed Hermes Agent on a Vultr server to give it a try so I could compare it to my OpenClaw agent:
Hermes Agent — The Agent That Grows With You
Hermes Agent — a standalone terminal app and a native application for macOS, Windows, and Linux. Install it and start a conversation with the agent that grows with you.
I am not expecting to replace my OpenClaw agent, Jeff, but I wanted to see if I could find a simpler and cheaper solution. Currently, my OpenClaw lives on a Mac Mini in my office and runs on Opus 4.6 with 1,000,000 token context window. Very powerful, but it's also expensive.
So, is a cheaper and more simple option possible? Well, it turns out it is.
The Hermes Docs do a great job of explaining all the benefits/features:
Hermes Agent Documentation | Hermes Agent
The self-improving AI agent built by Nous Research. A built-in learning loop that creates skills from experience, improves them during use, and remembers across sessions.
hermes-agent.nousresearch.com
I installed Hermes on a $10 Ubuntu server. The installation process was a few simple Terminal commands. But, instead of Claude/Opus, I decided to use OpenRouter so I could quickly test a few different cheaper models to see which one I liked best.
The winner (so far) is...
MiniMax 2.7
MiniMax M2.7 - API Pricing & Benchmarks
MiniMax-M2.7 is a next-generation large language model designed for autonomous, real-world productivity and continuous improvement. $0.21 per million input tokens, $0.84 per million output tokens. 204,800 token context window. Higher uptime with 6 providers. Includes independent benchmarks from...
openrouter.ai
Originally, I was using 2.5 but when 2.7 was released I switched my Hermes Agent, Herman, to 2.7. Minimax is a Chinese model. It's VERY good for the price.
Here is a comparison:
MiniMax M2.7 vs Claude Opus 4.6 - AI Model Comparison | OpenRouter
Compare MiniMax M2.7 from MiniMax and Claude Opus 4.6 from Anthropic on key metrics including benchmarks, price, context length, and other model features.
openrouter.ai
For what you get with 2.7, it's fantastic. Tool calls are working great with Hermes and coding has been great too. Here is an Arc Raiders website it built for me:
SperanzaGuide.com | Arc Raiders Complete Guide
The ultimate resource for Arc Raiders players. Enemy weakpoints, weapon stats, maps, crafting guides, and Speranza vendor information.
98% of that is built with MiniMax 2.5 and 2.7. The 2% leftover is some content that was created with GPT-5.4 Nano (coding with MiniMax though).
Even with all my testing and playing around (including some OpenRouter Opus calls), I have only used about $15 in credits for the past month with Hermes. I think you could probably run it on a smaller/cheaper server as well. You could also get a Token Plan directly with MiniMax and probably save some money.
So, if you are interested in trying out an AI agent but don't want the complications of OpenClaw and the cost of Opus, this might be a good solution for you to try. It's very powerful and gets more powerful the more you use it.
Give it a try and let me know if you have any questions
Last edited:
