Transcript
00:00:00You've probably already heard by now that the Hermes agent is the most powerful personal agent around, and that's not wrong.
00:00:06It actually has numerous features that make it so much better than OpenClaw.
00:00:09But what happens when you connect it to one of the most powerful coding agents out there, ClaudeCode?
00:00:14On its own, ClaudeCode is great, but it's missing one crucial part from the Hermes agent, the self-evolving skill system.
00:00:20Paired with that, you can create workflows that are so much more autonomous, even for things that we didn't think could be automated.
00:00:26But other than as a personal agent, you can set it for any business that wants to automate their processes, and it's really simple.
00:00:33But since a lot of you might be new to the Hermes agent, you don't need to worry, as we're going to guide you through it.
00:00:38But before we get into the setup, let's start with why the self-evolving skill system matters in the first place.
00:00:44When we came across it, we figured it might actually be better than OpenClaw.
00:00:48It wasn't just some random project, it's actually built by Noose Research, one of the leading labs in open source AI,
00:00:54and it has become one of their most popular projects.
00:00:57Also, here's something interesting.
00:00:58The Hermes agent was actually built before OpenClaw.
00:01:01It just didn't get much hype at first.
00:01:04The people over at Noose Research also tried OpenClaw, but they ran into issues with it, so they switched to their own setup.
00:01:10They saw its problems firsthand and open sourced their solution.
00:01:13Most of the features Hermes has are the same as OpenClaw's.
00:01:16Just like OpenClaw, you can connect it to multiple platforms.
00:01:19But there are two things that make it so much better.
00:01:22The first is persistent memory, and the second is self-improving skills.
00:01:26OpenClaw already has persistent memory which lets it remember information about you and shape its answers to what you like, but that has its limits too.
00:01:33Hermes goes further.
00:01:35It saves those memories, and whenever it finds a reusable workflow in your chats, it turns it into a skill.
00:01:40Hermes' persistent memory is built on a really smart setup.
00:01:43Hermes puts a limit on how large the user.md and memory.md files can be.
00:01:49As you're chatting with the Hermes agent, it keeps updating those files after each run.
00:01:53Now, why does that limit matter?
00:01:55It's because of how models work.
00:01:56Just like you, a model also has a really bad attention span.
00:01:59It can only focus on a limited amount of information at a time, and it gets confused when it's given a lot of information.
00:02:05And all that information, from the prompts and tools to the system instructions, and on top of that, your own files, is fighting for the model's attention in the context window.
00:02:15So the more you fit into that context, the more the model loses focus on the actual task, because all the extra information becomes noise to the agent.
00:02:23So that token limit is there to prevent this from happening.
00:02:26Once Hermes hits the token limit on the files, the model goes through them and cuts out anything that isn't useful.
00:02:32It holds the newest information in memory, so the agent isn't distracted by old details you don't need anymore.
00:02:38OpenClaw doesn't do any of this.
00:02:40It just lets the memory keep growing.
00:02:42There's another issue we faced with OpenClaw.
00:02:44To secure it, we had to sandbox the agent ourselves.
00:02:47Hermes runs in a sandbox on its own.
00:02:49That means it runs in an isolated environment where it can't reach things it shouldn't, or accidentally do something it isn't supposed to.
00:02:56So it gets rid of most of the security problems OpenClaw had.
00:02:59And if you want to run Hermes with a Claude code setup, now's the best time to do it because greedy little Dario discovered another way to make money off Claude by starting to charge for using your Claude subscription with third-party applications.
00:03:12After June 15th, you won't be able to use your Claude code subscription to run agents like Hermes for free.
00:03:18You'll have to pay Anthropic Extra.
00:03:19Your plan will include a monthly agent SDK credit, and that credit gets spent whenever you connect a third-party app through your subscription.
00:03:27The same limit applies to running Claude in non-interactive mode, which is the mode a lot of agents use to run Claude code in the background without needing any permission prompts.
00:03:36So until June 15th, you can keep running the Hermes agent without the ridiculous API costs.
00:03:41Now, setting up the Hermes agent is actually pretty simple.
00:03:44You just copy the install command and run it in your terminal.
00:03:47It first installs all the dependencies it needs, then runs the installer in interactive mode.
00:03:52If you want to set the agent up on the noose plan where you get their models and built-in tools, you can go for it.
00:03:58But we wanted our own setup, so we went with the manual option.
00:04:01You can also reconfigure the agent later on using the Hermes setup command.
00:04:05This step is where you set up everything the agent needs.
00:04:08Hermes can import from your previous OpenClaw settings, so it first asks whether you want to bring those over.
00:04:14You can check for yourself exactly what gets brought over, which covers your user profile and credentials, along with your skills and your soul file, which is basically the agent's personality and instructions.
00:04:24But just like how your bloodline's been passing down that amazing height for generations, inheriting from one agent to another comes with its own issues.
00:04:31The login details you bring over still point to the same channels your OpenClaw agent used, and the files OpenClaw relied on don't carry over cleanly, because those instructions were written specifically for OpenClaw.
00:04:43So importing them just causes problems, and that's why we chose not to import ours.
00:04:47After that, you choose which model Hermes uses.
00:04:50We wanted it on clawed models through the Anthropic subscription.
00:04:53But when we tried it, we couldn't actually use the clawed models and got an error.
00:04:57Turns out, Dario was already asking us to set up that extra usage, even though it's not June 15th yet.
00:05:03So that policy might already be rolling out gradually, but it might still work for you.
00:05:07Either way, we could still use clawed code in non-interactive mode right now, which is what we'll be using for most of our tasks anyway.
00:05:14And you can change your model provider anytime later on.
00:05:17Once the model is set, it asks where the agent will actually run, whether that's on hosting or a VPS you've set up.
00:05:24And for those of you who don't know, a VPS is basically a server you rent and run yourself.
00:05:29But since we have Mac Minis running entirely for this, we went with the local option.
00:05:33And no, we weren't the ones who caused the Mac Mini shortage, because unfortunately, just like you, our AI B2B SaaS business actually ran out of funding.
00:05:40After that, it asks you to connect whichever messaging platform you want.
00:05:44We chose Discord, but you can connect any of them.
00:05:46We won't walk through the Discord bot setup here, but you'll find the full instructions in our community, AI Labs Pro.
00:05:53Once that's done, it asks a few more questions, and your agent is ready.
00:05:56You just type Hermes, and once the UI loads, you can start chatting with the agent right there.
00:06:01In order to tailor itself to what you actually need, it needs information about you.
00:06:05So you can either keep using it for a month and let it figure you out on its own, or just tell it who you are up front before it touches any other task.
00:06:13If you want to set it up as a personal agent, you can either give all the information about yourself in the chat, or if you'd rather not type it all out, link it to your second brain vault instead.
00:06:22Just give it the path to your second brain and tell it to onboard itself from there, and it learns everything about you that way.
00:06:28If you want to set it up for a specific automation use case, just provide the docs of the use case or the general info about the company that it's being set up for.
00:06:37But before we move forward, let's have a word by our sponsor.
00:06:39So if you're building with AI tools every day, but don't fully understand how LLMs, agents, or protocols like MCP actually work under the hood, that gap catches up fast.
00:06:49That's where Educative comes in as an interactive platform used by over 3 million developers with 2,300 plus courses where you code in the browser with no setup and get AI-powered feedback on every submission.
00:07:01I'd start with their Essentials of Large Language Models course.
00:07:04In two hours, it breaks down how LLMs work, from tokenization to attention mechanisms to RAG.
00:07:09You'll finish with a real mental model of what's happening under the hood.
00:07:12From there, MCP Fundamentals teaches you to build AI agents using the model context protocol.
00:07:18Agentic system design goes further, multi-agent systems that reason, plan, and act autonomously.
00:07:23Then the 16-hour LLM bootcamp adds hands-on AWS labs in Bedrock, SageMaker, and LangGraph.
00:07:29You'll fine-tune a model, build multi-agent systems, and ship a RAG chatbot.
00:07:34You're writing real code from the very first lesson, not sitting through tutorials watching someone else build.
00:07:39Educative has helped more than 10,000 developers land jobs at top tech companies.
00:07:43Try it free. Link in the description.
00:07:45You can build a collection of skills for your Hermes agent from the Skill Hub.
00:07:49That's their official marketplace for skills, and it has skills for all kinds of use cases.
00:07:54Hermes also comes with 90 skills installed by default.
00:07:57Those pre-installed skills are actually secure because they are maintained by the organization itself.
00:08:03That's a real difference from OpenClaw skills, which we covered in our previous video.
00:08:07A huge number of those weren't safe at all, with security issues like dangerous prompts and scripts that can literally transfer your data off to some server.
00:08:15And the Skill Hub actually runs a security scan on each skill and watches for these issues.
00:08:20That way, you can add the skills you want without the same risk.
00:08:23Just like any other agent, you can connect any MCP you want to Hermes.
00:08:27But here's what separates Hermes from the rest.
00:08:29You can run your own Hermes setup as an MCP server itself and connect it to your other agents, letting them reach Hermes through tools so the communication goes both ways.
00:08:38Connecting Hermes to other agents this way fills in what those agents are missing.
00:08:42An agent like Claude Code on its own doesn't remember anything about you, and its skills don't fix or improve themselves.
00:08:49But through this MCP connection, you can give it access to everything Hermes can do.
00:08:53It also means you reach every app you've already connected to Hermes without wiring each agent up to each app separately.
00:09:00They just use those apps through your Hermes setup instead.
00:09:02To run Hermes as an MCP, you run the Hermes MCP serve command.
00:09:07There's no output on the terminal saying the server is up, but it's actually started running as the MCP server.
00:09:12To connect it to your agent, you add the Hermes MCP to the .mcp.json file, and then it's usable.
00:09:18You can set it at project scope, which means only the project you're working on gets access.
00:09:22Or you can add the config to the root.claude folder, and then the Hermes MCP is available across all your projects.
00:09:29And speaking of skills, Hermes comes bundled with a Claude Code skill, which has guidance on how to use Claude Code through the agent.
00:09:36So combined with the Hermes setup running as an MCP, this unlocks a lot for us.
00:09:41The Hermes agent and Claude Code together open up a lot of use cases,
00:09:44especially in businesses where multiple automations can be set up to handle repeatable processes.
00:09:49One of those is connecting Claude Code to your team's Slack workspace.
00:09:53This works really well because Hermes is basically an always running agent,
00:09:57while Claude Code is where the actual development happens.
00:09:59So we use the Hermes agent to access the team workspace.
00:10:02Just like with the Discord setup, we won't walk through the Slack connection here either,
00:10:06but you'll find the full guide in our community.
00:10:08In most workspaces, you could have a dedicated channel for a project,
00:10:12where the whole team discusses different points about it.
00:10:14What you can do with the Hermes agent is ask it to create a cron job that monitors that specific channel.
00:10:20From the requirements being discussed there,
00:10:22it builds a PRD skill that evolves as those requirements change.
00:10:26Having the PRD as a skill is really helpful,
00:10:29especially during the sessions where you're actually developing the product.
00:10:32Whenever it's needed, it pulls the relevant parts of the PRD into the context,
00:10:37so the project stays aligned with the original requirements.
00:10:39The PRD on its own might work too, but for the same reason we talked about earlier,
00:10:44the agent sometimes gets confused about what it needs to pay attention to.
00:10:48A skill gets called whenever it's needed and stays in the fresh part of the context window,
00:10:52where the model is actually paying attention.
00:10:54So Hermes creates the skill the way you instructed and runs it every 30 minutes as a cron job.
00:10:59This way, whenever a requirement change gets discussed in the channel,
00:11:02it updates the PRD, and the Hermes agent makes sure those changes flow both ways,
00:11:07so the skill created inside your project stays updated too.
00:11:10At this point, you might be thinking,
00:11:12since we already have an MCP connected for the Hermes agent,
00:11:15why not just use a tool to pull the information from that Slack channel and have the agent act on it?
00:11:20The reason is that the Slack MCP has a limitation.
00:11:23It can't read the entire conversation history by default.
00:11:26It only reads the messages it's tagged in,
00:11:28and it won't pull the full history unless the tagged message specifically needs that context.
00:11:33So setting it up through the Hermes agent is the better route,
00:11:36because it can sync the information directly from there.
00:11:38From there, you can also ask it to implement any feature using Claude Code in non-interactive mode,
00:11:44directly through the Hermes agent channels.
00:11:46It loads that Claude Code skill we talked about earlier,
00:11:49then launches Claude Code and uses it to build the feature.
00:11:52Also, if you are enjoying our content, consider pressing the high button,
00:11:56because it helps us create more content like this and reach out to more people.
00:11:59You can also bolt Hermes onto a deployed app, whether you're building it for yourself or for a client.
00:12:05So if you have a deployed app built with Claude Code,
00:12:08you can create skills for monitoring and health checks that guide the agent on how to monitor the running app.
00:12:13Because Claude Code has the best context on what the app actually needs.
00:12:17Then you import those skills into your Hermes agent.
00:12:20You can set up a cron job for that,
00:12:21basically a task that runs on its own on a schedule,
00:12:24and let the agent monitor both the hosted app and the code.
00:12:27We also told it that if it finds an issue while running the skill and updates it,
00:12:31it should sync those skills back to the local project,
00:12:34so Claude Code has context on them too.
00:12:36So this is how its self-evolving skills help in setting up a continuous health check
00:12:40that gets better every time it runs.
00:12:42So once you give Hermes the prompt, it sets up the cron job for you.
00:12:46You can test run it to see if it's configured properly.
00:12:49It gives you a report in whichever channel you set up,
00:12:51and in our case, it reported in Discord.
00:12:53And with the MCP configured, you can get those reports right inside Claude Code,
00:12:57along with all the suggested fixes from other team members,
00:13:01and implement them directly in your project.
00:13:03Or you can push those fixes yourself,
00:13:05or even set up the Hermes agent to fix the issues it found using Claude Code.
00:13:09If you want to found the next big AI B2B SaaS company
00:13:12and automate everything like we did with Hermes,
00:13:14you should be in AI Labs Pro.
00:13:16That's where you'll find the setup guides from this video,
00:13:19along with all the other resources and goodies we've put together.
00:13:22You'll also get to meet a bunch of like-minded nerds,
00:13:25including our team.
00:13:26The link's in the description, and you can check that out.
00:13:28That brings us to the end of this video.
00:13:30If you'd like to support the channel and help us keep making videos like this,
00:13:34you can do so by using the super thanks button below.
00:13:36As always, thank you for watching, and I'll see you in the next one.
Community Posts
No posts yet. Be the first to write about this video!
Write about this video