Running Hermes Agent locally with lm studio
Posted by KarezzaReporter@reddit | LocalLLaMA | View on Reddit | 12 comments
I am not a super smart guy and I'm not a tech guy. I'm not a developer but I use Claude code and Codex quite a bit. I loaded the Hermes agent and connected it with a qwen coder next on LM studio and it is pretty good. It's a way better experience than Open Claw. I got rid of Open Claw completely. I was an early adopter of Open Claw and I spent countless hours trying to get it to work right and I was just tired of it.
This Hermes agent already works way way better than Open Claw and it actually works pretty well locally. I have to be super careful about exposing this to the outside world because the model is not smart enough, probably, to catch sophisticated prompt injection attacks but it does work pretty well. I'm happy to have it and now I can talk to my Mac and tell it to do things over Telegram
Positive_Kale@reddit
So it is not useful to run the Hermes Agent inside the Docker container with LM Studio?
Also what hardware are you running this on?
I am asking because I was considering a Mac Mini Pro with 64 gb, where I would have a Hermes Agent loaded inside a Docker running Qwen 3.6 27B. I want to be able to use the Mac Mini for other things as well, like Firecrawl in a different docker container etc., so I won't give Hermes the entire Mac Mini alone.
General_Arrival_9176@reddit
honestly Hermes + qwen coder next on LM Studio is a solid combo. the LM Studio interface handles a lot of the setup pain that made OpenClaw frustrating. if you are getting decent results with it, stick with it. the prompt injection risk is real for local though - id be careful exposing it to anything beyond your local network
KarezzaReporter@reddit (OP)
it’s a huge risk. Have to be extremely careful. I’m going to audit security and lock things down even more. Thank you!
PracticlySpeaking@reddit
How high do you have to set the context for Hermes-Agent to work smoothly with Qwen-Coder?
KarezzaReporter@reddit (OP)
I’m using maximum. cuz why not. I haven’t experimented with lower lengths or context.
PracticlySpeaking@reddit
The RAM-challenged need to think about these things. 🤣
Agreeable-Shirt-5893@reddit
True! If you use turboquant even on 8GB Ram it optimized everything and ran with no gpu and low ram
TechnicianSoft4775@reddit
Lo hiciste usando Linux o windows, porque a mi usando wsl me da error 7
Big-Wear-8148@reddit
Hermes has a huge system prompt. When I try to run it with Qwen-3.5 35B it's difficult
KarezzaReporter@reddit (OP)
Did a security audit. Set up a new browser profile for my agent, locked down its folders to the extent I can, and removed the telegram key so it can’t be seized by someone. When you run your own agent you have to be extra careful. Running it in docker would be the ultimate, but that didn’t let it do anything useful so I had to stop that.
Jonathan_Rivera@reddit
Is your agent only allowed to respond to your user id on telegram?
KarezzaReporter@reddit (OP)
Yes