I replaced Hermes... Claude Agent 2.0

Added: 2026-05-09

[00:00:00]Hermes might be the most powerful AI agent on the planet right now. It's an AI personal assistant that runs 24/7, but it has one huge limitation. And in this video, I'll show my exact Claude system that solves this limitation by taking the best systems from Hermes, Open Claw, and any AI system to turn you into a productivity superpower with an agent that runs 24/7 so you can stop wasting time and make more money. And if you don't know who I am, my name is Jack Roberts. I built and sold my last tech startup with over 60,000 customers and I build my own AI businesses and I share with you the exact stuff that works. So if you haven't already, grab that coffee and let's dive straight in. So this is the Hermes agent. As you can see, it's a very sleek, gorgeous design. And you've probably seen on YouTube, Hermes is called Open Claw. Open Claw is called NL. I've not seen this much killing since I watched Game of Thrones. It's super difficult and actually just exhausting to keep up with what's killing what. Which is why the system is so powerful because we're not basically relying on any one single framework. And you'll see why that's so important in this video. But the tilt of it is it's got 124,000 stars on GitHub. It is literally blown up. It's trending for all the right reasons. It is genuinely cool and probably the most powerful one that exists on the market right now.

[00:01:19]It's absolutely crushing it. Now, why is everybody so hyped about Hermes? We have to cover this before I can show you the system that replaces it. First of all, it dreams whilst you sleep. That's pretty cool, right? Background ga loop reflects on every session. It prunes contradictions and distills everything into memory. MD. That's cool. That is literally an AI agent that's working for you in the background constantly. Writes it own tools, its own skills. So, after five plus tool calls on a complex task, the agent also authors a skill.md capturing what worked. next time it just runs for you. So it's less if you think about like cognitive load from your perspective. You talk, it kind of learns, it listens and it does what you want to do. It actually remembers you.

[00:01:59]It's got a four layer memory system.

[00:02:00]Memory MD userd skill.md then SQLite and some episodic pluggable stuff there.

[00:02:06]Then we've got it talks on to you on every app. So again, one gateway 15 services, Telegram, Discord, etc. Real research, not just vibes and MIT license model agnostic, which is very cool. And so every time something new comes out, we have to do a new repo with new conventions, we have to recredentialize, which is a word that I I quoted. I don't know if that exists or not. Hopefully it does. But basically, you're hooking up everything all over again. Every single connection, the second that a better framework comes out, it's so time consuming. Your agent doesn't know who you are cuz again, you lose all the memories from each individual app when you move across or you have to record it. It's a complete pain in the backside. You can't fix what you don't understand. if you don't have one that's own running on your own computer or you know your environment that you actually understand it's really tough and it's always been my experience that I run into barriers or just limitations when you want to update something with these frameworks and the key thing is no matter how great the guys are it's their road map it's not yours and you're going to have a very completely different unique perspective of how you want to drive this and after months of stress testing this I pulled together a system that works for me flawlessly this is the entire thing I'll put it down below for you for free goes from telegram bot to LM keys bootstrap the environmental variables everything how the memory system works you can just copy that uh by literally clicking link down below and so when you build your own custom just like in Hermes you can do anything that you can do there so I've got mine in telegram this right here is my mission control that I run everything through this is a sample one on my new desktop but it gives you the idea now the cool thing about this is you're not constrained by any limitations in so for example I could come over to anti-gravity or basically cord code and say hey there I want you to add to my mission control an overview of all the current things that my agent is working on, make it interactive and add it to the core dashboard. And literally, that's how you do builds. You literally speak to it and it develops it. So, you can give it any command you want to.

[00:03:52]I've just said, "Hey, turn this into a billy link glider.com." There you go.

[00:03:55]All done and dusted. I click on that and does it appear? Just by magic, it literally does. So, we can basically leverage all the skills within our own gravity claw. So, the way that you can basically supercharge your Hermes agent without any of the limitations is what I would call cherry-picking the winners.

[00:04:09]So essentially anytime a new harness drops be it Hermes or anything like that all you literally do is come over to the website or the GitHub repo and copy it.

[00:04:16]Then come over to your coding environment of choice and just basically say, "Hey there, this new repo just dropped for Hermes agent. I'd like you to go through the repo, go through the website, identify and understand everything that it's offering that is different than what we currently have and what makes it unique and then give me a bolted list of those cool features and then I'm going to go select which features I want and then you're going to integrate it into our existing system and then just drop in the GitHub repo."

[00:04:39]And for context, this is the Hermes repo. One option you've got if you want to is to literally just install Hermes agent. Of course, you can do that. It's my experience though in using open core and these different systems that they invariably always hit a pressure point in the wall where you just have to bend it essentially fold it into a pretzel to do the things that you actually want.

[00:04:59]And I personally don't like the idea of being limited and stuck into just one particular system and architecture. I like to cherrypick the very best components from all of the systems, which we can do super easily. And just like that, it's come back with a full breakdown and analysis, which is awesome. It's got the full breakdown.

[00:05:16]It's got easy wins, medium effort, big lift. So, based on what we want to develop. And what's really cool here is I can now just take a look at all the different features that I want to add in that I personally think are really freaking cool. Features that Hermes has that we don't. Self-generating skills, procedural memory, sub aent delegation, context files. Really freaking cool. So, if I come down here to the bottom, what it's got here is a top recommendations for self-generating skills again, and you're just adding in the things that you pretty much like. So, all I'd say is, for example, I'd copy this and I say, "Sounds great. Go ahead and add in self-generating skills. Let me know if there's anything you need for me." And just confirm for me when that is complete. And the cool thing is it's got full access to the ripper and the frameworks, which makes it super easy.

[00:05:55]And so, when I built up the memory system for Gravity Claw, which I found to be extremely effective, I have three levels to it. Level one is the core memory, which effectively is just the core things that it needs to know. Hey, my name is Jack. I like this. It's like your soul.md, the personality of how gravity claw communicates your your key facts, your key information. Then we have this conversational buffer, which is a rolling um stack of the last X messages, the last 100 messages, 50 messages, etc. And then you have tier three, which is a semantic tier, which basically indexes everything in Pine Cone. The other cool thing I have in my operating system, if I pull up here, is I can add in any deep intelligence from anything like notebook, any videos, whatever. I I can throw it in there and then I can have conversations. And these are just limitations that I find you run into with different, you know, pre-built frameworks. And then for where it actually lives, well, you've got a couple of options, right? You've got VPS, which, you know, it's like you have to use the correct. I've got full guides that help you set it up safely. But in reality, the other ones are your desktop, so it runs on your Mac Mini or Railway, which I personally use for Gravity Club because it's super easy.

[00:07:02]It's not exposed to internet and it just means that whether the laptop's closed or not, you're running Hermes. No super freaking complicated setup. You literally connect um Railway to Anti-gravity or Cool Code and it's done.

[00:07:14]Like that's how simple it is. And you don't have to have these restless nights thinking, "Damn, did I leave the port exposed? What the hell is going on?"

[00:07:20]type of situation. But obviously, you know, you can do the VPS route. I'll put off a link to the full secure guide down below so you can do that if you want to.

[00:07:27]Then the other cool feature here is going to be its skill calling, which is great. So you can pretty much just pick any skill you want to. Fire crawable, Gmail, whatever the hell you wanted to do. You can just explain it and give it that particular skill. And as you can see, it's now added self-generating skills just off one prompt. That was it.

[00:07:43]One prompt and it's completed it, which is crazy. So how it works is skills are stored over here. plain markdown files, progressive disclosure, so only skill names and descriptions go into the system prompt at 30 tokens each, auto nudge. After a complex task or five plus tool calls, the agent gets background prompt asking if the procedure is worth saving. It decides autonomously if it's a one-off, it skips it, it's reusable, it can correct the skill, and they're self-improving. So skills can be edited when the agent finds a better approach or the user corrects the workflow. So with that in mind, let's actually create a skill that's going to be super useful for a lot of the things we do, which is research. We're out and about. We're thinking about things and I always think with use cases these things I like to use things that I would actually use, right? Like building graphics and stuff looks cool, but are we really going to use it all the time, right? So one example would be in doing some decent research. I always use firecraw for research just because if you're using AI, it's 80%ish cheaper and this actually goes ahead to the website and can scrape. And I always think of it as like you're talking to the website. So all you do is add the API key in your environmental variables in anti-gravity or claude and just say hey, you know, save this API key and it stores it for you in railway. And when you've done that, so let's go ahead and test this out. So I'm going to come down here and hit the record button. Hey there. What I'd like you to do for me is create a skill using firecrawl. So when I ask you to go and research something, I want you to use firecrawl and I want you to look at the top three to five credible resources and come back with a punchy, beautiful, punctuated summary of what you found. Now, first of all, what I'd like you to do is go ahead on to Product Hunt and find some custom reviews for speech dictation software and let me know what the three star reviews are saying. So, in other words, how people who like the product but want to be better. What are the things that they're saying? Just let me know so I can get some ideas to improve our product, right? I'm going to send that one straight off and then let our new system basically go and do that and it will use the same fire cool thing. And even this ability here to transcribe and send voice messages is something you can add in. So it just got what to do. That's how cool it is. Especially when you combine it with Hermes. Beautiful. And just like that, it's completed. Research completed and skill created, which is awesome. So now we got that brand new skill. What users like better. It's given us all of the details. Now it's ready to rock and roll. And that's how simple it is to add in any of these features. And so we can basically add in all the different skills that we want to in this user interface. And the cool thing about what model's actually talking, again, we can hot swap it for anything. My recommendation would be if you can use GPT5.5 because you can access that via your existing $20 a month chat GPT subscription at no additional cost. They let you use what's called oorthth which is just how you authenticate. Claude will not allow you to do that. Chat GPT does. Another good option of course is you've got the um 4.7 set. That's a really good one. And Deepseek is a hidden powerhouse, but just be a bit more careful of what you're typing with Deepseek. I've done a full video breaking down that because that is like insanely cheap right now to use in these models. And of course, you can use this completely for free in open code. Then once you've got that, the other thing that you want to look at is the dashboard like I showed you.

[00:10:40]Effectively, anything you want to, you can go ahead and create that dashboard.

[00:10:43]There's so many different things that you can build using Claude. I've actually got a full course that I'm releasing this week in the uh community which covers everything from foundation setup, building websites, uh memory systems, all the way down to design systems and turning it into dollar signs. Now, the next thing that we need to do is supercharge this Hemes Claude system, which we can do by building super skills.