Hermes agent might be the best open source agent for local models right now

Posted by virtualunc@reddit | LocalLLaMA | View on Reddit | 31 comments

been running hermes agent by nous research for a bit now and the local model support is genuinely better than anything else ive tried in this space the thing that sold me: it has per-model tool call parsers built in, so it actually handles tool calling properly on 30B class models where openclaw and most other frameworks just fall apart. multiple people on here have confirmed its way less token hungry too the self improving skills thing is real but honcho (the learning engine) is off by default which confused me for like 2 days before I figured it out.. once you enable it in config.yaml the difference is noticeable within a few sessions some stuff worth knowing: one command install, handles python and node and everything. supports ollama, vllm, sglang out of the box. six terminal backends including modal and daytona for serverless (basically free when idle). gateway connects telegram discord slack whatsapp signal from one process. built in openclaw migration if youre switching the team behind it (nous research) are the people who actually train the hermes model family so the integration between agent and model is tighter than anything a third party wrapper can do 22k stars, MIT license, no telemetry. v0.6.0 just dropped with profiles for multi-instance setups wrote up a full deep dive with install paths, real costs, security breakdown, comparison vs openclaw/cowork/creao: [here](https://virtualuncle.com/hermes-agent-complete-guide-2026/) curious what models people here are running with it.. ive been on mistral 30b via ollama and its surprisingly capable for the size

Reply to Post

31 Comments

[-]

Daemontatox@reddit

Cant wait for the era of clawdbot and its clones to die out like "RAG-Memory" , this whole agentic bot hybe train is nothing but security catastrophes waiting to happen.

[-]

EuphoricPenguin22@reddit

Not really sure why I would want an agent running all of the time when I can just fire up Cline and use it when I actually need to extend an LLM with a few extra tools to more efficiently interface with a codebase. There have already been a few supply chain attacks that have happened through malicious skills installed on OpenClaw instances, and these agents that are running 24/7 are milliseconds away from dumping any and all secrets they have access to on the open web.

[-]

Pleasant-Shallot-707@reddit

Hermes isn’t an OpenClaw clone.

[-]

last_llm_standing@reddit

in terms of security its not better than openclaw, and all these posts and comments seems to be bot made, so forgive me if im sceptical

[-]

Small-Fall-6500@reddit

Post literally ends with "curious what..." And other posts this user has made follow the same AI Slop format.

[-]

Yu2sama@reddit

I mean right after that there is a grammar error so... 😭

[-]

deepspace86@reddit

Yeah I think we're starting to see the reverse bleed effect now where people consuming a lot of AI generated text are now emulating that patter with their own thought and speech patterns.

[-]

fastheadcrab@reddit

Generating AI slop as a person may not be any better than using a computer

[-]

Yu2sama@reddit

This is very true, the more you consume it the more you are influenced by it.

[-]

Small-Fall-6500@reddit

Removing capitalization and apostrophes is a fairly low bar. I'm guessing most LLMs can generate that style with a single sentence added to the prompt.

[-]

virtualunc@reddit (OP)

lol I mean I am genuinely curious what other people are doing/using, not really sure how else to phrase that but fair point I guess, sorry

[-]

DinoAmino@reddit

Not a single bot chooses to say it differently, like "Wondering what others..."

[-]

virtualunc@reddit (OP)

fair skepticism tbh.. im not going to pretend the security is dramatically better than openclaw, the docker container still runs as root by default and the execute\_code sandbox has a known bypass through the terminal tool. both are documented on github as for the bot thing, hard to argue that or convince you I'm not heh so not sure what to say

[-]

x0xxin@reddit

What are your security concerns? I run mine on a dedicated, cheap mini PC that's connected to the same VLAN as all the iot (Internet of shit) stuff that I don't trust. My agent has its own accounts for things like my local git repo and kubernetes cluster. For my fairly vanilla use cases, Hermes has been a massive upgrade over the various claws I have experimented with. Regardless of whether you're running a claw, Hermes, crew AI or a simple tool calling setup that does web requests and fetches you are still susceptible to prompt injection. From my perspective it's all about limiting blast radius. I'm not sure that using an agent with the ability to execute its own code makes much of a difference given explicit role-based access control.

[-]

kkingsbe@reddit

I agree

[-]

No_Afternoon_4260@reddit

Really fan of nous research for the bloke era. Their agent is beautiful and move really fast (like a release every 3 days with 100+pr per release) afaik we are at 0.7.0 Really impressive project to watch. Just note that this is really fresh they haven't got time to build dataset and finetune models on it (yet, afaik)

[-]

EffectiveCeilingFan@reddit

I've gotta say, that release cadence makes me extremely worried for code quality. It's just physically not possible to properly review and merge 100 PRs every three days, even if you had a fully funded, dedicated team. Not to mention the time wasted on PRs that eventually get rejected.

[-]

Ok_Mammoth589@reddit

Not really? Impossible for you maybe. I certainly wouldn't do it if you paid me. But back before we lost our education we had words like verve and impassioned to describe these weirdos that actually got shit done

[-]

EffectiveCeilingFan@reddit

Upon further research, the author is also the one who writes the majority of the PRs. This is, obviously, stupid. However, even aside from that, the pace of changes is rediculous. teknium1 (the author), personally made 131 commits between the most recent two releases, with an average length of 266.3 lines/commit. So, not small changes by any stretch of the imagination. The total lines changed by teknium alone between releases was 37,800 (none of which were generated files as far as I can tell from some thrown together Bash scripts). The time between releases was 4 full days. That's 19 lines of code written every single minute of an entire 8 hour work day for 4 days. A bunch of the code has probably never even been seen by human eyes, let alone reviewed to any meaningful degree.

[-]

virtualunc@reddit (OP)

same and youre right, v0.7.0 just dropped yesterday.. 168 PRs in that release alone. the pace is genuinely hard to keep up with tbh. the finetuning point is fair too.. the atropos integration is there for generating training trajectories but afaik nobody has published a model specifically finetuned on hermes agent task completions yet. thats the endgame though, agent that trains its own successor model from its own successful runs

[-]

dampflokfreund@reddit

Teknium is a machine. If you dislike something or it doesn't work, he will likely fix it in a few hours if not less.

[-]

No_Afternoon_4260@reddit

Thinking that low cost version of that could eventually be a A4B moe on cpu, that's wild!

[-]

DerDave@reddit

I have Honcho disabled currently and it does learn. What exactly does noticably get better, when you activate Honcho?

[-]

virtualunc@reddit (OP)

honestly thats a good question and I might be overstating the honcho dependency.. the base skill creation and memory persistence works without it, so youre right good catch honcho adds the dialectic user modeling layer on top, so its less about the skills learning and more about hermes building a deeper model of how you specifically work and communicate over time.. the difference is more subtle than the skill improvement stuff appreciate the correction though, gonna look into this more

[-]

constructrurl@reddit

Hermes actually hits different for local deployments - the function calling alone saves you hours of prompt engineering vs the usual suspects

[-]

Pleasant-Shallot-707@reddit

Seriously. It’s really great and the harness keeps it from doing dumb shit

[-]

Pleasant-Shallot-707@reddit

It’s the best agent, period

[-]

Willdudes@reddit

From a security perspective, there is a lot that can be improved. Kata containers or fully VM. Lock down all egress except specific URLs/APIs you need. If any egress attempts are made to other destinations, shut it down. All skills packages must be preinstalled. No destructive permissions should be available for the agent. If possible, include QWEN guard on all data to be processed. Scan all outputs for PII using Qwen Guard. Make sure you convert all data from UTF8 and remove all HTML and metadata. Figure out the agent blast radius what can happen if it hallucinates or is compromised. Not a full list but strongly recommended if you are going to production.

[-]

Crampappydime@reddit

Tbh it may just be a preference thing, but yeah I like hermes more. But I’ve had a lot of fun with openclaw too, comes down to what you want out of it. Hermes feels a bit more principled at the moment, when it comes to safety i just assume using an agent you kinda accept the risk that comes with smart toddler inside your PC.

[-]

defervenkat@reddit

Honestly either is just fine. Whatever you started with won’t like an upgrade once you have significantly invested and figured out the tools. Meh for me.

[-]

JohnMason6504@reddit

Per-model tool call parsers is the right call. Generic wrappers waste tokens on malformed retries. What quant are you running? Q4_K_M gives solid t/s on 30B class models without noticeable quality loss for agentic workloads.