Hi does anyone use this? https://hermes-agent.nousresearch.com/
I had a look and I can't see any posts on here about it.
pod
I had a look and I can't see any posts on here about it.
pod
Hermes Agent - AI agent
- Thread starter jpod
- Start date
It was created by Nous Research Inc. Let me google that for you: Nous Research are the developer of the Hermes series of language models which has widespread popularity in the open-source AI community and surpassed 50 million downloads.
Thanks for posting MG. Is it worth spending time on? I gather setting it up on a separate PC is a good idea.One of my clients uses it. It's like OpenClaw but a bit less ****.
Still an absolute horror show of security and privacy concerns
I am thinking of using it to download pdf (info) from websites, reminders, researching and writing.
I have been playing with it, it can be quite good but you need to give it a lot of access to be useful, well that is the same with a lot of AI things, security risk is high but man is it useful when it is doing its thing.
And boy can it burn through tokens, I run most stuff on my local AI units (Strix Halo machines) only branching out to frontier models for structure or to essentially lint check.
These units aren't the fastest but they can do useful things and not burn a lot of power at home.
And boy can it burn through tokens, I run most stuff on my local AI units (Strix Halo machines) only branching out to frontier models for structure or to essentially lint check.
These units aren't the fastest but they can do useful things and not burn a lot of power at home.
Last edited:
Client has configured it to essentially be an autonomous programming team. You can just invoke it in Slack and ask it to do something, it'll create a linear ticket with the requirements and clarify anything through slack. Once it's in linear another agent will pickup the task and set a claude code agent off planning and executing the work, which will end with a PR. That'll go through the usual array of checks and reviews and then is ready to merge.
I mean, I don't yet quite trust it enough to be that autonomous with my company, but it's genuinely useful for client's startup that is just having to pivot rapidly to demo to potential clients with slightly differing use cases.
Sandys thanks for posting. I had never heard of 'lint check' - I googled it and I learnt a new phrase!
Do you run Hermes on local AI units? Or another AI?
Thanks that is really interesting. I have been watching You Tube videos about it and when I searched on OC there were no threads which made me worry!
I guess what I am trying to find out is is it worth finding out more about it or focusing on another or other AI.
I am a rookie PC user - I don't code - I write and research and I am interested in marketing tools.
Hermes is not an AI, it's just a scaffolding for enabling AI to interact with your PC and other services through tools.
You still need to pay for the credits you're burning on the actual model underneath (or self-host, but that needs good hardware and is much more limited than frontier models).
It's not an Ai, it is an agent or a harness that uses AI and stores things along the way and you can use it in many ways, I have configured a few agents, one for coding where I have generated a lot of reference 'memory' basically text files which contain descriptors of what it can use, what exists, how I like this structured etc, etc, and as I use it, if I see something I like/dislike I let it know, it makes a note in its database and evolves it's ability for next time.
Similarly I have a travel agent uum agent and a personal asssitant, all set up with their own set of skills and descriptions of what I want from them.
As mentioned in my post I do this on my own local AI, this doesn't need to be the worlds most power machine, you don't need 4x 5090s etc, I started doing this on a gaming handheld, a Powerful Strix Halo one mind with unified RAM
but it could be any CPU and RAM setup, if you are just playing, just keep to usable models to experiment, whilst the results are far better and faster tied in to a Claude sub the burn rate for tokens on this agent is high, so local is the way for grunt work, using subs when you want higher quality so you don't end up poor.
I should mention, I don't really know what I am doing, rookie too, learning on the fly, but getting results.
Similarly I have a travel agent uum agent and a personal asssitant, all set up with their own set of skills and descriptions of what I want from them.
As mentioned in my post I do this on my own local AI, this doesn't need to be the worlds most power machine, you don't need 4x 5090s etc, I started doing this on a gaming handheld, a Powerful Strix Halo one mind with unified RAM
but it could be any CPU and RAM setup, if you are just playing, just keep to usable models to experiment, whilst the results are far better and faster tied in to a Claude sub the burn rate for tokens on this agent is high, so local is the way for grunt work, using subs when you want higher quality so you don't end up poor.I should mention, I don't really know what I am doing, rookie too, learning on the fly, but getting results.
Last edited:
Thanks for posting - very interesting to read your insight.It's not an Ai, it is an agent or a harness that uses AI and stores things along the way and you can use it in many ways, I have configured a few agents, one for coding where I have generated a lot of reference 'memory' basically text files which contain descriptors of what it can use, what exists, how I like this structured etc, etc, and as I use it, if I see something I like/dislike I let it know, it makes a note in its database and evolves it's ability for next time.
Similarly I have a travel agent uum agent and a personal asssitant, all set up with their own set of skills and descriptions of what I want from them.
As mentioned in my post I do this on my own local AI, this doesn't need to be the worlds most power machine, you don't need 4x 5090s etc, I started doing this on a gaming handheld, a Powerful Strix Halo one mind
I should mention, I don't really know what I am doing, rookie too, learning on the fly, but getting results.
Hard to know what you are looking to do but your suggestion for downloading sites and compiling info is something it can be quite good at, tied in with something like obsidian or llmWiki, research these, good for the memory,
Agents can be setup so you just drop reference data you found that looks interesting in an area, it will then look in it, compare it with its memory (wiki) update, collate and report back to you findings etc, great for travel itinerary stuff and keeping thing current.
Agents can be setup so you just drop reference data you found that looks interesting in an area, it will then look in it, compare it with its memory (wiki) update, collate and report back to you findings etc, great for travel itinerary stuff and keeping thing current.
Last edited:
I started with OpenClaw and then I got involved with Hermes and then Claude code. So now I've got three harnesses I use and then I purchased a Nvidia DGX Spark which I use for local AI and I started to go down the rabbit hole of memory and I've created an audit trail and then a wiki and then open knowledge format and now I'm doing vector on top of that. I basically use scripts to pull info out of Obsidian vault in a router type configuration which I have designed. It speeds up look ups and keeps costs down. Agents only get a context window which when compacted they start to lose context of what they are doing. If you start a new session unless you configure handover scripts they start all a fresh. So it is a lot of hand holding and token costs. Hermes is better equipped than Openclaw and Claude to reduce this
This is a long learning process and especially with the spark trying to get the inference up and running with the models and so forth but it's a good hobby but it's a really quite expensive hobby. I find the best models to use externally cloud based I have found so far to use which is more of a daily workhorse is Deepseek 4-flash. It's very cheap but it gives good results. I've tried lots of free models, I've had subscriptions with Swiss companies, I've had subscriptions with Olamma and tried free models with Openrouter ,
I've tried different methods of actually supporting what I'm trying to do but to get the right balance I'm hoping that my spark will be able to get the right balance at some point when I get the model configured to the correct model that I want to run as a daily workhorse. At the present moment I'm just running GPT OSS 120 as my Claude library assistant in the Obsidian Vault. I've got other models as well I can run as well but I find that to be the most resilient of the way I'm running the actual models from the spark.I get it to produce snippets and summaries of the sessions. I might go down the route to try and get Deepseek Flash 4 which can run on a single spark highly quantised. People are having quite good results with that but I haven't gone down the route to try to do that. As far as learning process. If you're trying to figure out a spark to do things the way you want using Vllm wise compared to other methods it can be quite complex to say the least.
I have not as yet got anything substantial out of this as far as work goes. My original idea was to research stocks or use trading view and strategies like quants using highly mathematical methods. Different ideas to create pine scripts run back test that sort of project type projects. Without memory and the research i want to do I didn't consider AI ready to do that without a good memory behind it. I want eventually to have memory provide the AI paths to quick solutions through knowledge of knowing which paths didn’t work so it can learn to get to solutions faster or find solutions where it would otherwise not have. You can’t improve the model AI unless you train it which you can also do with Spark, but the harness like Hermes and the memory can improve the results. It like us as humans our harness is our environment we have books notes TV these are all our external harness. So i am trying to develop that for AI.
This is a long learning process and especially with the spark trying to get the inference up and running with the models and so forth but it's a good hobby but it's a really quite expensive hobby. I find the best models to use externally cloud based I have found so far to use which is more of a daily workhorse is Deepseek 4-flash. It's very cheap but it gives good results. I've tried lots of free models, I've had subscriptions with Swiss companies, I've had subscriptions with Olamma and tried free models with Openrouter ,
I've tried different methods of actually supporting what I'm trying to do but to get the right balance I'm hoping that my spark will be able to get the right balance at some point when I get the model configured to the correct model that I want to run as a daily workhorse. At the present moment I'm just running GPT OSS 120 as my Claude library assistant in the Obsidian Vault. I've got other models as well I can run as well but I find that to be the most resilient of the way I'm running the actual models from the spark.I get it to produce snippets and summaries of the sessions. I might go down the route to try and get Deepseek Flash 4 which can run on a single spark highly quantised. People are having quite good results with that but I haven't gone down the route to try to do that. As far as learning process. If you're trying to figure out a spark to do things the way you want using Vllm wise compared to other methods it can be quite complex to say the least.
I have not as yet got anything substantial out of this as far as work goes. My original idea was to research stocks or use trading view and strategies like quants using highly mathematical methods. Different ideas to create pine scripts run back test that sort of project type projects. Without memory and the research i want to do I didn't consider AI ready to do that without a good memory behind it. I want eventually to have memory provide the AI paths to quick solutions through knowledge of knowing which paths didn’t work so it can learn to get to solutions faster or find solutions where it would otherwise not have. You can’t improve the model AI unless you train it which you can also do with Spark, but the harness like Hermes and the memory can improve the results. It like us as humans our harness is our environment we have books notes TV these are all our external harness. So i am trying to develop that for AI.
Wow TB you invested in some serious hardware there! Nvidia DGX Spark.
And Sandys too! Strix Halo machine!
Crivens I still use a 2500k intel machine - I love it! I have a part build Ryzen machine which I need to get up and running. Right before the pandemic I purchased a rtx 3900 with loads of onboard memory (24GB) I hope to use this for AI on this old DDR4 rig when I retire it.
Fascinating read. I like Claude - I use it to tidy my writing up - its great. I like Notebook llm too - I can drag a load of scientific papers, pdfs, into it and ask it questions - often it predicts the question I intend to ask.
I am interested in trading - I read a lot of older stock market text books - I read about ADX, RSI, SRSI, MA, and their conclusion was trade in trends (up down) and when the ADX is over 20. The RSI crossing the 50% marker is a good signal. RSI divergence too. Latest one I am tinkering with is Purple cloud on Trading View.
And Sandys too! Strix Halo machine!
Crivens I still use a 2500k intel machine - I love it! I have a part build Ryzen machine which I need to get up and running. Right before the pandemic I purchased a rtx 3900 with loads of onboard memory (24GB) I hope to use this for AI on this old DDR4 rig when I retire it.
Fascinating read. I like Claude - I use it to tidy my writing up - its great. I like Notebook llm too - I can drag a load of scientific papers, pdfs, into it and ask it questions - often it predicts the question I intend to ask.
I am interested in trading - I read a lot of older stock market text books - I read about ADX, RSI, SRSI, MA, and their conclusion was trade in trends (up down) and when the ADX is over 20. The RSI crossing the 50% marker is a good signal. RSI divergence too. Latest one I am tinkering with is Purple cloud on Trading View.
Get Hermes or Claude to connect to MCP server version of tradingview https://github.com/atilaahmettaner/tradingview-mcp then Claude or Hermes can produce the pine scripts all using free account you just need to download tradingview desktop version. With ai you can get around the 2 indicator limit by having Hermes or Claude produce a pine script in it of what ever you want. They can design a strategy for you if you want, warning back test first NFA
Soldato
- Joined
- 21 Jul 2005
- Posts
- 21,332
- Location
- Officially least sunny location -Ronskistats
Yeah I been using that since about Feb, how you finding it?
There was a time a 128Gb SH machine was £1500, sadly those times have gone.
Its very new to me I am a rookie with PCs just interested in AI. What are you using it for?