Speaker A
Welcome to "The Ultimate Beginner's Guide to Hermes Agent." In this video, you're going to learn how to set up your own AI agent from scratch and how to actually use it.
Learn to set up and use Hermes Agent, a free open-source AI assistant that runs 24/7 on your own server with no coding required.
Key Takeaways
- Hermes Agent offers full ownership and flexibility unlike commercial AI services.
- No coding skills are needed to set up and use Hermes Agent effectively.
- A VPS is essential for continuous operation and security during learning.
- Hermes can integrate with multiple messaging platforms and tools for seamless workflow.
- Training the agent with personalized prompts and reusable skills maximizes productivity.
What the video covers
- Introduction to Hermes Agent, a free open-source AI assistant that runs on a VPS and integrates with familiar apps like Telegram and Slack.
- Step-by-step beginner-friendly guide with six levels covering setup, remote access, file integration, training, automation of weekly chores, and managing a team of agents.
- Comparison of Hermes with ChatGPT and Claude, highlighting model choice flexibility, ownership, and multi-platform messaging.
- Explanation of why a VPS is recommended for always-on operation and separation from personal devices.
- Use of Hostinger’s one-click Hermes template to simplify installation and management for beginners.
- Details on costs involved: a few dollars per month for the VPS and AI model usage fees.
- Instructions on connecting Hermes to various tools like Google Drive and messaging platforms.
- Training the agent to learn and improve over time by saving reusable skills and handling tasks autonomously.
- Managing tasks via a project board and setting up scheduled jobs with notifications.
- Encouragement to customize prompts based on personal or professional needs for faster learning and better results.
Chapters
- 00:00Introduction and Overview of Hermes Agent
- 03:32Key Differences Between Hermes and Other AI Services
- 06:43Costs and Setting Up Your VPS Server
- 09:41Fixing Onboarding and Initial Setup
- 12:45Using Nous Portal and Model Selection
- 19:16Enabling Messaging Channels and Managing Threads
- 25:14Handling Errors and Restarting Hermes
- 28:31Authentication and Desktop App Setup
- 31:57File Management and Scheduled Jobs
- 35:46Google Drive Integration and API Setup
Full Transcript — Download SRT & Markdown
Speaker A
By the end, you'll have an agent that runs 24/7 on its own machine that you can message from your phone and your desktop, that knows how you like things done, that works inside the tools you already use, and that takes real work off your to-do list.
Speaker A
While you're away, including a small team of agents that work through a task board entirely on their own.
Speaker A
This guide is designed for complete beginners. So you don't need any technical background. And you won't write a single line of code.
Speaker A
I'm Matt, by the way. I use Hermes every day. My own agent runs more than 20 scheduled jobs for me right now.
Speaker A
From bookkeeping checks to trip planning. And hundreds of thousands of people have used my Hermes tutorials on this channel to get their own agent running.
Speaker A
The truth is that the setup is the easy part. The hard part is knowing what to do after.
Speaker A
If you treat Hermes like a chatbot, it feels like just another chatbot. It gets interesting when you treat it like an assistant that learns.
Speaker A
So this video is built as a ladder with six levels. And we climb it from start to finish without skipping anything.
Speaker A
In level one, we hire your agent and give it a home that stays on around the clock for just a few dollars a month.
Speaker A
In level two, we reach it from anywhere with one scan of a QR code and a desktop app that feels like ChatGPT or Claude.
Speaker A
In level three, we put it to work inside your own files and your own accounts.
Speaker A
In level four, we train it. And you'll watch a single correction stick forever. In level five, we hand it a chore you do every week and never do it again.
Speaker A
In level six, we give it a team with a project board where you queue up work and your own agents run it on their own.
Speaker A
Every level is timestamped below. So you can come back to any step later. And if you get stuck, leave a comment.
Speaker A
I read all of them and answer as many as I can. But before we set anything up, you need to know what you're actually hiring because Hermes can use the same AI models as ChatGPT and Claude.
Speaker A
And the difference is everything wrapped around them. Let me show you what that means.
Speaker A
Hermes Agent is a free open source AI agent made by a company called Nous Research.
Speaker A
You pay only for the AI model it thinks with and for the machine it works on, which, for us, is a small rented server that costs just a few dollars a month.
Speaker A
If you've used ChatGPT, Claude, or Gemini, you know the basic idea. You type in plain language.
Speaker A
And it answers you. Hermes can use those exact same AI models underneath too. So the biggest difference is everything built around the models.
Speaker A
If you already pay for ChatGPT or Claude, some of this will sound familiar. Both of those can run a task on a schedule now.
Speaker A
Both of them remember you. Both of them can read your email and send a message for you.
Speaker A
If that's all you wanted, you already have it. But there are three things you can't buy from either of them.
Speaker A
The first is your choice of model. ChatGPT and Claude only let you use the models their own company makes.
Speaker A
Hermes connects to any model from any company. So you can put a cheap one on the boring jobs and a powerful one on the hard jobs, all inside the same agent.
Speaker A
The second is ownership. ChatGPT and Claude are services you rent with weekly limits and features the company can just take away.
Speaker A
Hermes is a free software that runs on a server you own. So no company can throttle it or switch off something you rely on.
Speaker A
The third is where your agent lives. ChatGPT and Claude only talk to you inside their own apps.
Speaker A
Your Hermes Agent gets its own Telegram, WhatsApp, or a Slack account so it can message you in the same app you already use every day.
Speaker A
There's one more difference. In ChatGPT or Claude, you make a project for each thing you work on.
Speaker A
And everything it knows about that work stays inside that project. Hermes turns that around.
Speaker A
There's one agent. It learns your whole world. And you point it out whatever you're working on.
Speaker A
Over time, it writes down what it learns as skills, which are step-by-step notes it can reuse so it gets better at your work specifically instead of starting fresh each time.
Speaker A
And if you've heard of OpenClaw, this is the same category of tool. An always-on personal agent.
Speaker A
If you're coming from it, you don't have to move anything by hand. Once you're set up, just ask your agent to import your OpenClaw configuration.
Speaker A
And it walks you through the whole thing. Now, I have one request before we start.
Speaker A
Throughout this video, I'll use a few different examples, all based around things I'm working on in my life.
Speaker A
Don't copy my examples, word for word. Instead, run each prompt through your own topic.
Speaker A
Whether that's your job, school, or something you do at home. You'll learn much faster, working on something you actually care about.
Speaker A
Your agent needs a machine to live on. And that machine has to be three things.
Speaker A
It has to be awake. Your scheduled jobs, the watching, and the messages it sends you all depend on the machine being on at three in the morning.
Speaker A
And the laptop in a bag isn't on. It has to be separate from your things.
Speaker A
An agent runs commands, installs software, writes files, deletes files. While you're learning, you want all of that happening somewhere with nothing of yours on it.
Speaker A
No personal photos, no passwords, no sensitive work documents, and no browser logged into your bank.
Speaker A
Think of it as a practice workshop. You're going to make a mess in it, which is what it's for.
Speaker A
And if you ever break it badly, you delete the whole thing and start again in 10 minutes.
Speaker A
And it has to be reachable. So your phone, your laptop, and your tablet are all doors into the same agent, which is what we set up in the next level.
Speaker A
That machine is a VPS, a virtual private server. It's a computer in a data centre that stays on all the time and that only you can log into.
Speaker A
And the smallest one costs just a few dollars a month. You can install Hermes on your own computer instead.
Speaker A
But then it only works while that computer is awake. And it's running commands and deleting files right next to everything you own.
Speaker A
So for your first agent, rent the server. Now, you can rent a server from plenty of companies.
Speaker A
And on almost all of them, you'd be installing Hermes yourself from a terminal, which is the text window you may have seen programmers use.
Speaker A
That's the part I want to skip. And it's the reason we're using Hostinger. They have a one-click Hermes template that does the setup a beginner shouldn't have to touch.
Speaker A
It installs Hermes, puts it behind a secure web address, gives you a login page, and adds backups, one-click updates, and a restart button.
Speaker A
It's also one of the cheapest ways to do it. Throughout this tutorial, you'll see how everything looks while I'm recording.
Speaker A
But software like this gets updated all the time. So a button might be somewhere else.
Speaker A
Things might have different labels. And there might even be new features by the time you watch this.
Speaker A
If something seems different, don't panic. Follow the general idea, rather than the exact visuals.
Speaker A
All right, let's get set up. There are just two things that cost money today.
Speaker A
The server, which is just a few dollars a month, and a model account in a few minutes, which costs about the same as a streaming service.
Speaker A
That's the whole bill. So let's start by grabbing our Hostinger VPS. Use the link on screen or the first link in the description below.
Speaker A
It stacks an extra 10% discount on top of any sale Hostinger is already running.
Speaker A
So using that link is the absolute cheapest way to get started. When you click on that link, you'll land on this page here.
Speaker A
And if you click the choose plan button, it'll scroll you down to see the various plans that are available.
Speaker A
Here, you can see that our coupon has already been applied. So all the plan prices that are listed are the prices you'll really pay.
Speaker A
I recommend taking the KVM 1 plan. It's the smallest.
Speaker A
And if you ever outgrow it, you can simply upgrade later without reinstalling anything. So go ahead and click choose plan.
Speaker A
You'll land on the cart. And, here, you'll need to make a few selections. The first one is your registration period.
Speaker A
Longer terms are cheaper per month. But the discount needs at least 12 months. So 12 is the shortest I'd pick.
Speaker A
There's a 30-day money back guarantee whichever you choose. Then there's a box that's already ticked for ready to use AI.
Speaker A
Go ahead and untick it. It signs you up to a different AI service you don't need.
Speaker A
And we're bringing our own in a few minutes. Leave the rest of the add-ons alone.
Speaker A
And then check the discount is showing on this page. And then click continue. On the next page, you'll need to register your account.
Speaker A
Go ahead and register with Google, GitHub, or email address. I'll use an email address.
Speaker A
Create a password. And then enter a phone number if you'd like. But that's optional.
Speaker A
So I'll go ahead and click register. Then on the next page, enter your billing address and your payment details to complete the transaction.
Speaker A
I'll do this offscreen and meet you on the next page. After payment, you'll get this confirmation.
Speaker A
And then Hostinger should route you straight into the Hermes setup flow with a form that gives you an admin username and a password.
Speaker A
If you land on a bare server dashboard instead, that's a glitch that happens sometimes.
Speaker A
And I'll show you how to fix that in a second. Here, you can keep the username as hermes or change the admin username to whatever you want.
Speaker A
Just make sure you remember it. And then go ahead and click the eye to toggle on to see the admin password.
Speaker A
Go ahead and copy this. And then save it somewhere safe. Like a password manager.
Speaker A
Don't save this in text notepad or anywhere insecure that somebody else could find it.
Speaker A
This is how you're going to access your agent. And you want to make sure it's locked down.
Speaker A
Once you've got that saved, go ahead and click deploy. You'll see this loading page for a few moments.
Speaker A
And then you'll be routed into Hostinger's dashboard. You might be presented with this survey.
Speaker A
You can fill it out or click skip. And, here, you can see, at the bottom, at the right, that the application is still being deployed.
Speaker A
If you see this, wait a few minutes until this Hermes Agent application changes to a green check mark status and says running.
Speaker A
Here, you can see I got the application deployed confirmation. So we are all ready to go.
Speaker A
Now, if you were routed into this page without ever going through that onboarding, let me show you how to fix that really quick.
Speaker A
On the left side menu, find the VPS setting. It looks like this. It's a little cloud.
Speaker A
Then, from this menu, see if you have a VPS set up yet. If not, click the get VPS button.
Speaker A
If you do, go ahead and click the manage button on your VPS. From here, you won't likely have your Hermes Agent set up because you need to go through that onboarding in order for this to appear.
Speaker A
So, here, you'll click on the Docker Manager on the left side. Then click catalogue.
Speaker A
In the search bar, search for hermes. Then find Hermes Agent and click select. That'll open the onboarding form that we saw earlier and take you through the whole process.
Speaker A
Once you've got it all set up, you can manage your Docker application that holds Hermes here in the application section of Docker Manager.
Speaker A
We have two things installed here right now. The actual Hermes Agent and Traefik, which secures your connection to the server.
Speaker A
Setting up Hermes used to mean a terminal and a long list of typed commands.
Speaker A
On this route, you never touch a terminal. Your agent is already reachable through a browser.
Speaker A
And you log in just like any other website. On the Docker Manager page, go to your Hermes Agent.
Speaker A
And under the access column, click the open link. This is the login page for your Hermes dashboard.
Speaker A
Go ahead and save this URL in your bookmark so you can access it later.
Speaker A
Once you've done that, then go ahead and sign in with the username and password that you saved earlier.
Speaker A
Welcome to your agent's control room. Let me give you a quick orientation before we change anything.
Speaker A
On the left side, we have the nav menu. Sessions is every conversation your agent has ever had on any device.
Speaker A
And they all share one memory. Models is where you pick the AI that does the thinking.
Speaker A
And that's the first thing we fix. Because, right now, nothing is chosen. Cron is the schedule.
Speaker A
The page that lets it work while you're asleep. Skills is its library of procedures.
Speaker A
The ones it ships with, the ones you instal, and the ones it writes for itself.
Speaker A
And further down, you'll see profiles. These are extra agents you can add later. And if you scroll down further, you'll see, in the plugin section, Kanban.
Speaker A
That's the project board, which is level six. We'll visit most of these rooms today.
Speaker A
In the top of the nav section, if you click chat, it'll load a retro terminal style chat.
Speaker A
And, here, if you try to type something, like, in my case, I'll say, "Hello," you notice that nothing really happens.
Speaker A
In fact, we get this error that says, "No inference provider configured." On the upper right too, you'll see there's an error.
Speaker A
So your agent is installed and running. But it can't think yet. It needs a model provider, which is the account that supplies its intelligence.
Speaker A
And the gateway on the left side is the part that connects it to the outside world.
Speaker A
It carries messages to and from your phone later. Right now, it has nothing to connect to.
Speaker A
We'll fix the model now and the gateway in level two. A model is the AI that does the thinking.
Speaker A
The brain your agent borrows. And it's the same kind of AI that runs ChatGPT and Claude.
Speaker A
Hermes doesn't come with one. So you give it an account that supplies one. We're using Nous Portal, made by the same people who make Hermes.
Speaker A
Nous Portal is built into Hermes. So you simply connect it by logging in with no access codes to copy anywhere.
Speaker A
One account gets you access to more than 200 models from all the major AI companies.
Speaker A
And it supplies the tools your agent works with. Web search, reading, webpages, image generation, browsing, and speech.
Speaker A
$20 a month buys you $22 of credits. And whatever you don't spend rolls over to the next month.
Speaker A
Sign up with a second link in the description below. That link takes $15 off your first paid month.
Speaker A
So month one only costs you $5. And you still get $22 of credits, which covers everything in this video and weeks after it.
Speaker A
That link will take you to this page here, where you can see the invitation banner.
Speaker A
And you'll also see the deal I just mentioned. Go ahead and click the claim your invite link at the bottom.
Speaker A
You'll get this pop-up saying, "Welcome to Nous Portal." Go ahead and create an account with Google, your email, GitHub, Discord, or one of the other options.
Speaker A
I'll use my email again. You may need to enter a six-digit code from your email.
Speaker A
Go ahead and open your inbox and grab that and paste it in. You'll get a success message.
Speaker A
And then you'll be routed to the checkout page. And you can see that the discount has already been applied to your order.
Speaker A
Go ahead and check out. I'll do this offscreen and meet you on the next page.
Speaker A
After checkout, you'll land on this confirmation page. And you'll see, on the left side, that your account has been loaded with $22 of credits.
Speaker A
Now, before we leave Nous Portal, I want to show you where you'll manage your billing in the future.
Speaker A
On the left side, under the account section, click the billing page. Here's where you can go to see your current subscription.
Speaker A
And if you scroll down, you'll see there are different tiers. You can change between subscription tiers at any time.
Speaker A
There is a free tier on this page too. But it runs free models only on standard rate limits.
Speaker A
A single agent job makes dozens of calls. So you'll hit those limits pretty fast, which is why I don't recommend using that free plan.
Speaker A
If you ever run out of credits, simply scroll down this page to the top-up section.
Speaker A
And instead of having to upgrade your entire subscription, you can add whatever dollar amount of credits you want as a top-up at any given time.
Speaker A
All right, let's head back to our Hermes dashboard. I'll close out of my email and the Nous Portal tab.
Speaker A
Back in the Hermes browser dashboard, on the left side nav, scroll down and click into the keys page.
Speaker A
Here, the first option is Nous Portal. Go ahead and click the login button to the right.
Speaker A
Here, this will open a window. And it'll say, "Allow Hermes Agent to use your plus plan." You can click the connect button to activate it.
Speaker A
You'll get a confirmation saying it's connected. And you can close the page and return to the Hermes Agent.
Speaker A
So we'll go ahead and close this tab. And you can see, back on the keys page, that Nous Portal is already connected.
Speaker A
Now, let's pick a model. On the left side, scroll back up and select the models page.
Speaker A
Here, click change next to main model. And, here, we get a whole list of models offered by Nous Portal.
Speaker A
I recommend start with a fast and cheap model because you'll send a lot of messages while you're learning.
Speaker A
DeepSeek Flash is perfect for this. Go ahead and search for DeepSeek or scroll down and find it in the list.
Speaker A
Here, you'll see a few options. There's DeepSeek V4 Pro, DeepSeek V4 Pro 0813, and DeepSeek V4 Flash 0731.
Speaker A
These are the models at the time I'm recording. We want the Flash model. And I recommend just going with the newest version of whatever it is by the time you're watching this.
Speaker A
It could be DeepSeek V4.1 or even 5, 6, 7. Don't pay too much attention to the number.
Speaker A
The one we want is DeepSeek Flash. Now, the reason we're using DeepSeek is because it's good at using tools.
Speaker A
And a message with it costs a fraction of a cent. Later in this video, I'll show you when to move up to a stronger model and how to check what you're spending.
Speaker A
So we'll go ahead and click that. And then click switch at the bottom. It gives you a dialogue asking if it's okay to switch the model.
Speaker A
We'll click reload. Let's go ahead back to our chat page. We'll click that in the upper left.
Speaker A
And then, here, in the upper right, you can see that we now have DeepSeek selected as our model.
Speaker A
Let's go ahead and type a message in the chat here. I'll say, "Hello." And hit return on my keyboard to send it.
Speaker A
And we get a response. It says, "Hello, what can I help you with?" Two minutes ago, this agent couldn't answer anything at all.
Speaker A
Now, we can ask it anything. So let's ask what tools it came with. I'll say, "What tools do you have?" And it comes back with a list.
Speaker A
It has a built-in browser. It can run terminal commands, work with files, do code execution, search the web, delegate to subagents, use memory, schedule things, and a whole bunch of other miscellaneous tools.
Speaker A
Go ahead and read through that list once. Because the rest of this video is us switching those tools on one at a time.
Speaker A
Now, if you already paid for another AI provider, there are two other ways in.
Speaker A
You can buy credit as you go from a service like OpenRouter. Or you can sign in with your ChatGPT subscription, though an agent burns through a plan's limits pretty fast.
Speaker A
There are ways to point Hermes at a Claude subscription. But Anthropic's terms keep those plans to its own apps.
Speaker A
So I'm not walking you through that. Paying per message gets you every Claude model in here anyway on Nous Portal.
Speaker A
So at least for today, it's not necessary. To add a provider later, you can always go back to that keys page or ask your agent and it'll walk you through it.
Speaker A
That's level one done. Your agent has a home and a model. Here's your level up task.
Speaker A
Pause the video. Send your agent three questions about your own topic and see what it can already do with no setup at all.
Speaker A
When you're back, we give it a phone. Level two is reaching your agent from anywhere.
Speaker A
Right now, it lives in a browser tab. But we're going to give it its own Telegram account so it can reach you when you're nowhere near this dashboard.
Speaker A
We're using Telegram because it's the easiest one to set up. It just takes one quick scan of a QR code with no codes to copy anywhere.
Speaker A
And it's a great messaging app to use day-to-day. You get fast replies. And you can send files, voice notes, and photos all in one thread.
Speaker A
Hermes supports more than 20 other platforms too, including Discord, Slack, and WhatsApp. And you can add one of those later from the same page.
Speaker A
Before following along, get your phone out now. Instal Telegram if you don't have it.
Speaker A
And keep it nearby. I put a link in the description below that takes you straight to the Telegram download for your phone or a desktop.
Speaker A
So you don't have to go looking for it. Back in your Hermes web dashboard, on the left side, click channels in the menu.
Speaker A
Then find the Telegram channel. Here, you can see it's disabled. Go ahead and click the toggle at the upper right to turn it on.
Speaker A
You'll need to click the restart button for that change to take effect. So go ahead and click that.
Speaker A
And you'll get a gateway restarting pop-up. You'll get a message saying the gateway is not running.
Speaker A
And then this will take a few minutes for the gateway to turn back on.
Speaker A
If you don't see it turned on after a few minutes, I recommend going ahead and clicking refresh in your browser.
Speaker A
And, here, you can see now that the gateway is running. But there's no bot token configured for Telegram.
Speaker A
And it is turned on. So let's go ahead and click create with QR to connect for Telegram bot.
Speaker A
All right, now grab your phone and get your camera app up. Then go ahead and scan the QR code and open the link.
Speaker A
Here, you'll be taken to a conversation with the NousHoustedHermesBot. Click create bot. And then create.
Speaker A
That'll take you into a conversation with your Hermes Agent. But the bot's not fully connected yet.
Speaker A
So go ahead and hit back on Telegram. And select that NousHoustedHermesBot chat from the list.
Speaker A
Or if it's not there, like it's not in my case, go ahead and scan that QR code again to pull it up.
Speaker A
And that'll take you back to the conversation with the NousHoustedHermesBot. And, here, you can see that there is a message that says, "Connect this bot to your Hermes setup." Go ahead and click the connect this bot button.
Speaker A
And then you'll see, at the same moment that you get a bot connect notification in Telegram, your bot appears in the web dashboard as well.
Speaker A
Here's the bot name. And then here's the allowed user. This is your Telegram ID.
Speaker A
To lock this in, go ahead and click save and restart. And this can take a few minutes, just like that first restart change.
Speaker A
So give it a few minutes to restart. And if, after a while, it doesn't flip back to showing it that everything is running, go ahead and click refresh in your browser.
Speaker A
And then you'll see that Telegram is connected and your gateway is running. Now, in your Telegram app, you can go ahead and tap on the username for your bot.
Speaker A
That's the @hermes_ and series of numbers and letters. That takes you back into the conversation with your bot.
Speaker A
So go ahead and click start. And the start command doesn't actually start the conversation.
Speaker A
You have to send a message. So let's test that out. We'll say, "Hello! Can you hear me?" And you can see that the Hermes Agent is typing a message.
Speaker A
You also get a message saying that there's no home channel set for Telegram. The home channel is where Hermes delivers the cron job results and cross-platform messages.
Speaker A
So go ahead and simply tap the /sethome command. And then that sets Telegram to be your home channel.
Speaker A
Now, anytime a cron job runs, and it needs to send you a notification, Telegram is where it will send it.
Speaker A
And you can see the message that came through here. It says, "Yes, I can hear you loud and clear.
Speaker A
What can I help you with, Matt?" So, now, Telegram is connected. And we can continue chatting with our agent on our phone or even on the Telegram app on the desktop.
Speaker A
Let's cover safety. This flow just put your Telegram account on the allow list. So if anyone else finds your bot and tries to message it, they get nothing back.
Speaker A
And Telegram bot chats aren't end-to-end encrypted. So don't send it passwords or banking details.
Speaker A
For everything we do today though, it's fine. Your conversation on the phone and your conversation in the dashboard are two separate threads.
Speaker A
A lot of people find this confusing at first. So think of it like email.
Speaker A
You can have five email threads going with the same assistant. What you said in one thread doesn't automatically appear in another.
Speaker A
But the assistant has all five threads in their inbox and can look any of them up.
Speaker A
Your agent works the same way. It has one memory, but many separate threads. So if you start something in the app, and you want to continue on your phone, hand the thread over with the handoff command.
Speaker A
If that ever gives you trouble, you can just ask it on Telegram to continue what you started and it'll search its own history.
Speaker A
Let me quickly show you how that handoff works. So go to your chat. And let's start a conversation.
Speaker A
Let's go ahead and ask it something random. Like, "Give me a synopsis of the movie 'The Godfather.'" Now, we've got this answer.
Speaker A
And let's say we wanted to hand off this conversation and continue on Telegram. You can type / and then handoff.
Speaker A
Add a space. Then you'll see a list here. Go ahead and toggle down to Telegram and hit enter on your keyboard.
Speaker A
Then again. And, now, you can see, here on Telegram, it says, "Working here now." Same session.
Speaker A
"What we've got so far: You asked for a synopsis of 'The Godfather.'" And, now, I could continue that conversation here.
Speaker A
And if I flip over to the desktop version of Telegram, you can see that I've got the same thread here too.
Speaker A
Now, there are other slash commands you can send, which are especially useful from Telegram.
Speaker A
If you type /new and then hit enter, that starts a brand new session. And it asks for permission to do this.
Speaker A
So we'll go ahead and click approve once. And, now, we're on a brand new conversation.
Speaker A
By default, if you just keep going, all of your conversations on Telegram will end up in the same thread, which means your model is getting a lot of context every time you send a message.
Speaker A
So a general rule is, if you start a new topic, go ahead and start a new session.
Speaker A
That'll save you from burning through tokens. And we'll cover more about that later. On Telegram, you can also just type /, and scroll through, and see all the different kinds of commands you can send.
Speaker A
You don't have to learn all these now. But just know that these are here and available for you to prompt your agent to do various tasks.
Speaker A
These commands that start with / work on every door. The app, the dashboard, and Telegram.
Speaker A
The other one that's super useful is /model. Typing that allows you to switch between any of the models that are available to you.
Speaker A
For example, I could click into the Nous Portal and see all of the various models that are offered through there.
Speaker A
But I'll go ahead and click cancel because I'm going to stick with DeepSeek for now.
Speaker A
Now, if your bot ever goes quiet, there are three things to check in this order.
Speaker A
First, check your credit and your rate limits with your model provider because that's the usual cause.
Speaker A
And Hermes is good about telling you when it happens. So look at your last message.
Speaker A
Second, the gateway. Here, you can see my gateway status is running. But check this if you're having any issues.
Speaker A
And click restart if you see anything strange. That can help to fix a silent bot almost every time.
Speaker A
And then third, the logs. You can find the logs in the left sidebar near the top of the list.
Speaker A
Here are all the different logs. And you can filter through by type. All debug-related logs.
Speaker A
Info, warning, and errors. Logs is your agent talking about itself. You don't need to read every line.
Speaker A
But just filter for errors when something seems like it's going wrong. And if none of these three steps fix it, restart the whole application from Hostinger Docker Manager.
Speaker A
Click the three dots next to your application. And then restart. Okay, so, now, you can access Hermes through your browser.
Speaker A
You can talk to it on Telegram. But, now, let's get it set up with the official desktop app on your computer.
Speaker A
The browser dashboard is the control room. And it works from anywhere. But it's a tab you have to go find.
Speaker A
For daily use, you want the desktop app. It looks and feels like the ChatGPT or Claude desktop apps.
Speaker A
And it keeps your conversations in a sidebar so you can drag files into it.
Speaker A
And it can listen and talk back. We'll instal it. And instead of letting it run a second agent on this computer, we'll point it at the server we set up so it uses the same model and same memory inside a native app.
Speaker A
I'm on a Mac. But on a Windows, the flow is the same with Control on the keyboard where I press Command.
Speaker A
To download the official desktop app, go to this link here. And I'll also go ahead and put it in the description.
Speaker A
It's hermes-agent.nousresearch.com. And then look for the download button for your operating system on that page.
Speaker A
I'll go ahead and click download for Mac OS. Go ahead and save that installer somewhere.
Speaker A
And then when it's done downloading, go ahead and run that installer. Right now, it's quite simple on Mac.
Speaker A
Simply drag the Hermes icon into the applications folder. Next, open your apps. Then open and run Hermes.
Speaker A
It might ask for permission on Mac. It says, "Hermes is an app downloaded from the internet.
Speaker A
Are you sure you want to open it?" And go ahead and click open. Yes, it's fine.
Speaker A
And then you'll get a popup where you're prompted to instal Hermes. Go ahead and click the instal Hermes button.
Speaker A
And then it'll run through a list of setup steps automatically. This can take a few minutes.
Speaker A
So I'll fast forward to when it's done. Then once the installation is done, go ahead and click to launch Hermes.
Speaker A
You'll eventually see something like this, where it says, "Let's get you set up with Hermes Agent." And it prompts you to connect to a model provider.
Speaker A
Skip this provider screen. Because picking one would create a second agent on this computer.
Speaker A
And we already have one set up on the server. So just click I'll choose a provider later.
Speaker A
Welcome to the Hermes desktop app. Before we go any further, let's connect our server.
Speaker A
Go to the upper right and click the gear icon. Or on Mac, you can hit Command + , on your keyboard.
Speaker A
On Windows, it would be Control + ,. Here, on the left side, find the gateways menu item and select it.
Speaker A
Then select remote gateway in the connection mode. Now, the field we need to enter some information in is the remote URL.
Speaker A
And all we really have to do here is put the link where our server is.
Speaker A
To find this is super easy. Go back to your Hermes Agent dashboard in the browser.
Speaker A
Here, grab the link that's in your address bar. Everything that comes before the /.
Speaker A
Here, I have /chat. So go ahead and just grab everything up to that. Copy that.
Speaker A
Head back to the Hermes desktop app. And then paste in that URL. Here, you can see.
Speaker A
It says, "For authentication, this gateway uses a username and password. Sign in to authorise this desktop app." So go ahead and click the sign in button.
Speaker A
And that'll give you this familiar sign in interface. Go ahead and enter your username and password.
Speaker A
And then sign in. You should see that it says, "You're signed in." Go ahead and click the save and reconnect button.
Speaker A
That'll restart the gateway connection. And, now, we should be able to start a conversation with our agent.
Speaker A
So go ahead and close out the settings page. Head to the chat. And, in the text field, say something like, "What was my first ever message to you on Telegram?" And then send it.
Speaker A
Now, you might see this error that says, "Backend out of date." That pops up if the app is newer than the server.
Speaker A
You can dismiss it or click update Hermes. I'll go ahead and click the X to dismiss it for now.
Speaker A
I'll update later. Here, you can see. The agent is now thinking. And, now, it comes back with a response saying, "Your first message to me on Telegram was, 'Hello, can you hear me?'" So you now have one agent with three doors.
Speaker A
The web browser dashboard, Telegram, and the desktop app. Every door shares the same memory.
Speaker A
Now, one thing that I think is really cool about using the desktop app is the heads-up display mode or HUD.
Speaker A
To access that, hit Command + Shift + H on your keyboard. And that'll open this little floating bar that sits on top of whatever you're working on.
Speaker A
And its position tells Hermes what's under it. So if you say something like this or here, that means the page or app you're looking at.
Speaker A
Now, I don't use it all day. But it saves a window switch when you want a quick answer for something.
Speaker A
And you can drag it around as big as you want or anywhere on your screen.
Speaker A
It's super handy. You can exit this mode by clicking this exit HUD mode button on the right, or use the same command, Command + Shift + H, to put it back to the main desktop app.
Speaker A
Now, here are some other buttons that are useful on the desktop app. For example, here, this microphone button.
Speaker A
You can click it to enable voice dictation. For example, I can click it. And it might ask for permission to your microphone.
Speaker A
You can go ahead and click allow. And, now, you can see that it's dictating.
Speaker A
Hello Hermes. I am currently recording a test dictation message. Can you hear me? And you click the stop dictation button.
Speaker A
It'll quickly transcribe it and turn it into a message that you can send. I'll go ahead and Command + A to select all that and delete it though because I don't need to send that message.
Speaker A
If you want your agent to talk to you, you can turn on this speaker button.
Speaker A
And that will have it read its replies out loud to you. You can also click this wake word button here.
Speaker A
That's where you can just say, "Hey, Hermes" into your microphone at any time to activate Hermes to listen to you.
Speaker A
Finally, this little wave form icon here at the end starts a voice conversation. So you and Hermes can just talk back and forth with voice.
Speaker A
This is quite handy if you just need some quick answers for something or don't want to spend a bunch of time typing in your keyboard.
Speaker A
That's level two. Your agent has a chat on your phone, a desktop, and a voice.
Speaker A
Here's your level up task. Pause here. Message your agent from your phone while you're away from the computer.
Speaker A
And then pick the same thread up in the app. Once you've done that, level three is where it starts doing work for you.
Speaker A
Level three is putting it to work. ChatGPT, Claude, and Gemini can all search the web, read your files, and connect to your accounts.
Speaker A
So what exactly makes Hermes different? Two main things. The first is where the work happens.
Speaker A
Your agent has its own computer. So whatever it makes stays there. Files it can pick up again next week and somewhere for the scheduled jobs in the next level to write to.
Speaker A
The second is that its whole ecosystem is built around doing work autonomously and getting better with each run all on its own.
Speaker A
So we'll run a quick research task to see the tools working and then connect a real account you'll actually use.
Speaker A
In the desktop app, let's initiate a research task. I'm looking for a standing desk right now.
Speaker A
So I'll go ahead and say, "Research the top three standing desks. Give me a one paragraph comparison with links." You watch it load some skills and start to use tools to actually read the web.
Speaker A
Here, you can see, it's running a tool. Saying it's reading New York Times, for example.
Speaker A
It's running code in the background. Reading pages and putting together its recommendations. I'll go ahead and fast forward to when it's done.
Speaker A
It searched, opened the pages, read them, and came back with an answer instead of a list of links, although it did provide me with some citations down here, which is super handy.
Speaker A
Now, let's do something it should stop and ask about first. I'm going to ask it to create some test files and then delete them.
Speaker A
So, here, I'll say this. "Create three test files." And, here, you can see it wrote into a folder on the server.
Speaker A
The opt/data folder. It created three different files. One text file, one markdown file, and one JSON file.
Speaker A
So, now, let's go ahead and have it delete them. I'll say, "Delete those three test files you just made." Now, because I gave it explicit permission to delete those files, it went ahead and did it.
Speaker A
But if it was working on something in the background, and it decided it needed to delete those files as part of the process, it would ask me permission first.
Speaker A
You might see a little something in the desktop app that says, "Run." Or if you're on Telegram, it'll give you a little popup similar to that model selector popup.
Speaker A
And you could improve that task for the session or just once. Simple safe tasks like research don't trigger any sort of permissions prompt.
Speaker A
But deleting things can. And so does running a risky code command. And nothing happens until you say yes.
Speaker A
Now, the way I think about permissions is progressive trust. Don't hand it everything on day one.
Speaker A
Give it a little. Watch how it works. And give it more only when you're comfortable.
Speaker A
In practise, that means a separate account for anything it logs into. And your main email and anything with money in it stays off the list until it has earned them.
Speaker A
Now, let's look at files. In the app, you can drag anything into the chat.
Speaker A
A PDF, a spreadsheet, or a screenshot. And, on Telegram, you can send it a document or photo.
Speaker A
Here, the desktop app, I'll drag in this AI index report PDF. And I'll tell it, "Summarise this and pull out the three numbers I should care about." It gets to work.
Speaker A
And I'll fast forward to when it's done. All right, it came back with an answer.
Speaker A
It gave me three numbers to focus on. These three here in the list. Now, let's tell it something like, "Turn that into a short document I can send to a colleague." And after a minute, it comes back with this AI index brief file as a markdown file.
Speaker A
And it gives me the option to open a preview or download it. So I'll go ahead and download it.
Speaker A
And then put it in my downloads folder. An agent that lives on a server has one catch.
Speaker A
The files it makes live on that server too. And files you send through the chat, like Telegram, have a size limit.
Speaker A
The desktop app allows for greater throughput with the file sizes. But, right now, our agent and us don't really have a shared file system.
Speaker A
So we're going to give you and your agent a shared folder in the cloud.
Speaker A
I'm using Google Drive because many of you already have it. And the same setup gives your agent your Calendar and your Spreadsheets, which we'll use in level five and again in level six.
Speaker A
For this, I use a separate Google account for the agent. Not my main one.
Speaker A
And you should too. Make one now if you need to. It takes two minutes.
Speaker A
You can then share individual folders from your real account into it so it only ever sees what you choose to give it.
Speaker A
In a new session, go ahead and say, "Set up the Google Workspace skills so you can use my Google Drive, Calendar, and Sheets." Now, I'm going to show you something.
Speaker A
This doesn't actually go the way I expected it to here in the recording. So, normally, there's a tool called Clarify that Hermes can call if it has additional questions for you.
Speaker A
Now, I don't really know why it's doing this right now in my particular setup.
Speaker A
And I'm not going to gloss over the fact that, sometimes, with AI, there's just bugs.
Speaker A
That's just what happens. It's not a perfect tool. But it's constantly improving. So what I did here was I took a screenshot of the fact that the form would not load its questions.
Speaker A
And then I sent a message to Hermes with a picture of that and said, "I can't see the questions.
Speaker A
They're just loading forever." And I went ahead and pushed that through the queue manually.
Speaker A
If you ever have issues with your Hermes Agent, you can simply just take a screenshot of what's going on and send that as a message to Hermes.
Speaker A
And it can help you resolve its own issues. So that's pretty cool. And the nice thing here is it came back, asking me the questions outside the special tool.
Speaker A
It just asked them in plain text, which works just fine for us. For you, Clarify may work or it might not.
Speaker A
Either way, now, you know how you can handle it if it doesn't. So, here, it's asking to set the scope for the connection.
Speaker A
It says, "You asked for Drive, Calendar, and Sheets." It'll also include Docs for us.
Speaker A
And, "Say the word if you'd rather also have Gmail and Contacts too or keep it to just three." For me personally, I'll go ahead and add those so I can demo how that works.
Speaker A
But Gmail is a sensitive subject for some people. So only include it if you are comfortable connecting to it.
Speaker A
Here, it also asks me, "Advanced protection. Does your account use hardware security keys?" Mine does not.
Speaker A
So I'll tell it that. So it just asks for me to answer that plainly in a text sentence.
Speaker A
So I'll just say something like, "Regarding the scope, I'd like to use all those tools.
Speaker A
My account does not include any advanced protection such as hardware or security keys." So I'll go ahead and send that and let it think for a second.
Speaker A
So, now, we have some setup steps to do. First, we need to create a Google Cloud Platform project.
Speaker A
It's provided me a link here. So I'll click this. And, here, you can let Hermes browse with a snapshot of your default browser profile.
Speaker A
So, essentially, you can pull in cookies from your existing live browser. So let's go ahead and click use my profile if you want to carry over those credentials.
Speaker A
And this might open for you in a native app browser inside of Hermes. Or if you're interacting with it in something like Telegram, it might open a different browser.
Speaker A
Personally, I'd actually rather work with this in my standard browsers. So I'll go ahead and copy this link out.
Speaker A
And go back to my Brave browser here. Open a new tab and paste in that link.
Speaker A
And, here, it log me right in. I already have another project in Google Cloud.
Speaker A
But you might not. So I'm going to go ahead and click create project, which is what you'll do.
Speaker A
Let's call the project Hermes Agent. And then click create. After a second, it'll load your new project.
Speaker A
You can see that up here in the top bar. It says, "Google Cloud." And next to it, says, "Hermes Agent." Next, we want to enable APIs and services.
Speaker A
And, here you can see. It's already popped out for me. But if you don't see this, go ahead and go to the APIs and services nav item.
Speaker A
And then click enabled APIs and services. I'm going to go ahead and close out these popups.
Speaker A
All right, now, there's a button up top that says, "Enable APIs and services." Click that.
Speaker A
Now, we need to go through and find the tools inside of Google Workspace that we want to connect.
Speaker A
So let's go ahead and start with what we told it. Let's start with Drive.
Speaker A
Search for Drive. And you might get some automatic suggestions here. But we'll just go ahead and search for Drive as it is.
Speaker A
And this one here is the one we want. Google Drive API. We'll select that and then click enable.
Speaker A
After a moment, you'll be taken to a page like this that shows the Google Drive API.
Speaker A
And, up at the top, you'll see something that says, "Disable API." That means it's been enabled.
Speaker A
If you were to click this button, it would stop it. But we don't want that now.
Speaker A
This is just a quick way to check if you are set up. So, now, we'll go ahead and click the back arrow.
Speaker A
And we run through this process again for every Google Workspace tool we want to connect.
Speaker A
And I'll admit. This is the most tedious part of the entire setup. But the nice thing is, once you've set this up, you never have to do it again.
Speaker A
So I'll just go ahead and do this quickly for all the tools that I want to connect.
Speaker A
We'll use Gmail API, enable. Google Calendar API, enable. Google Docs API, enable. Google Sheets API, enable.
Speaker A
And then Google People API, enable. Google People API is the more modern version of Contacts.
Speaker A
All right, now that we've connected the APIs, let's go back and double check what Hermes needs from us next.
Speaker A
I'll go ahead and close out this browser here. You can do that by clicking the little X in the tab.
Speaker A
There we go. Next, after enabling the APIs, it wants us to create some credentials.
Speaker A
You can click the link here. Or I'll show you where to get to it in the Google Cloud Platform.
Speaker A
Here on the left side of the APIs and services page, you'll click credentials. Then click create credentials at the top.
Speaker A
Then OAuth client ID. And then if you see this yellow banner saying, "To create an OAuth client ID, you must configure your consent screen first," go ahead and follow that process too.
Speaker A
So we'll go ahead and click configure consent screen. We'll click get started. And we'll enter some information about our app.
Speaker A
Our app, in this case, is whatever we want to name our current project here.
Speaker A
So we can just call that Hermes. User support email. We'll just use our own email from the dropdown.
Speaker A
Then click next. For audience, internal versus external, go ahead and select external. Then click next.
Speaker A
Then for contact information, just enter your email address again. Then click next. And then click the check box to agree to the terms of service.
Speaker A
And click continue. Then click create. Okay, now, we can go back to the APIs and services.
Speaker A
And then credentials. Now, click create credentials up top. And OAuth client ID. From the application type dropdown, select desktop app.
Speaker A
And then give it a name. Again, we can just call it Hermes. Click create.
Speaker A
And then, here, you have some information that you're not going to want to share with anyone.
Speaker A
The client ID, the client secret. This is all stuff you want to keep private.
Speaker A
Don't share this with anyone. Download the JSON file. Put it somewhere you'll find it.
Speaker A
In my case, I'll put it in downloads. Now, drag that new JSON file it gave you into the Hermes chat, per the instructions here.
Speaker A
Now, there's one more thing we need to do here. We need to add ourselves as a test user if the app is in testing.
Speaker A
Here, it gave me a link. So I'll click that. I'll go ahead and copy this URL again and close out the browser.
Speaker A
Open it in the main browser here. And, here, at the bottom of this audience page, find the test user section.
Speaker A
Click add users. And then add yourself to the test users list. Just enter your email.
Speaker A
Hit enter to lock it in. And then hit save. Okay. Now that we've followed all of Hermes instructions, we can go back to Hermes.
Speaker A
And then just hit enter to send that JSON file. Okay, now, it gives us an authorization URL.
Speaker A
I'm going to go ahead and right click this link it gives me. And then click open in external browser.
Speaker A
Now, I'll log in with the account that I plan to use. It'll say, "Google hasn't verified this app." That's okay because you're technically the developer of this app that we're setting up.
Speaker A
So go ahead and click continue. It'll ask for the scope of things that it's allowed to do inside of Google.
Speaker A
So go ahead and select all. These are the exact things that Hermes has asked permission for.
Speaker A
And you can read through them and change the scope if you want to. But these are all the things that we specifically asked it to do.
Speaker A
So this is just fine. We'll go ahead and click continue. And you'll get what looks like an error.
Speaker A
This is actually the expected behaviour here. So go ahead and copy this URL from your browser address bar.
Speaker A
Return to Hermes. And then paste in that full URL here in the message field.
Speaker A
And send it. This can take a few minutes to process. When it's done though, you should see something that says that the setup is complete.
Speaker A
And it'll confirm that it's got everything connected. Now, if you didn't get to this point successfully, the easiest thing to do to troubleshoot is simply take a picture of whatever it is that you're encountering on screen that's giving you trouble
Speaker A
and send that screenshot to Hermes to help you troubleshoot. Just explain what you're seeing.
Speaker A
Explain what kind of errors are popping up. And this process is pretty straightforward for an agent like Hermes to help walk you through.
Speaker A
And if, for some reason Hermes, itself isn't responding or isn't actually doing a great job at troubleshooting, you could always go to another chatbot somewhere online, like ChatGPT or Claude, to help you out.
Speaker A
Once you're confirmed though, next, let's actually set up our file system in Google Drive.
Speaker A
So let's say, "Create a drive folder called Hermes with Inbox and Outbox inside and give me the link." Here, it just created some folders and gave me a link.
Speaker A
So let's go ahead. And I'll right click this Hermes folder link. And click open in external browser.
Speaker A
Close these old tabs I don't need anymore. And, here, you can see, in my Google Drive folder here, I have the Inbox and Outbox folder that have been created.
Speaker A
So, now, Hermes has access to perform tasks inside of Google Drive for me. This is great.
Speaker A
Let's head back to the Hermes desktop app. And I'll go back to that summary conversation I had where it made a report for me about the AI index.
Speaker A
So now that Hermes has access to my Google Drive, let's tell this summary conversation, "Save the summary you wrote as a document, upload it to Outbox in Google Drive, and send me a link that works from my phone."
Speaker A
After a few minutes of processing, it says it's done. And the summary is now a Google Doc inside my Drive's Outbox folder.
Speaker A
Share it so that anyone with a link can view it. So let's go ahead and open this in an external browser.
Speaker A
And there's my document. Now, I can share this with anyone or even view it on my phone.
Speaker A
Now, we've seen it write content into Google Drive. Let's see it retrieve some content from Google Drive.
Speaker A
I'm going to go back to my Google Drive tab here. Open the Inbox. And then add a new file here.
Speaker A
Click file upload. I'm going to add in this State of AI in 2026 document.
Speaker A
It's uploaded. So I'll head back to Hermes. And let's see how it does, being a little bit ambiguous.
Speaker A
I'll say, "Grab the newest file in my Inbox folder and give me a five bullet summary." So I didn't tell it that there's an Inbox folder in Google Drive.
Speaker A
I just said grab the newest file in the Inbox. And let's see how it does.
Speaker A
It pulled it in. It says, "The newest file in your Inbox folder is the State of AI 2026." And it gave a five bullet summary.
Speaker A
So that's the shared Google Drive folder. Anything you want it to read goes in Inbox.
Speaker A
Anything it makes for you lands in Outbox with a link. And it works from any device.
Speaker A
Now, one thing I want to point out here before we go on to the next subject is that it automatically did a self-improvement review.
Speaker A
It wrote a document delivery skill, wrote a skill credential setup update, and updated its memory.
Speaker A
I didn't have to tell it to do that. It just figured out what the process was for delivering documents.
Speaker A
In this case, sending them to Google Drive. It figured out that there were some updates it needed to make to the skill credential setup, and updated it, and decided to remember things.
Speaker A
This is what makes Hermes really powerful. Now, those two folders are just a starting point.
Speaker A
Once you see what your agent actually produces, you'll want folders that match your own work.
Speaker A
So go ahead and change them. You can also ask your agent to keep them tidy.
Speaker A
Either when you think of it or on a schedule. Filing things where they belong so it doesn't fill up with files you can't find later.
Speaker A
Now, that setup also gave your agent access to Gmail. And this is a question I've seen in many comments.
Speaker A
If it can reach my email, is it going to start just sending things? It does what you ask it to do.
Speaker A
So if you tell it to send an email, it'll send the email. So the control comes from the rule you give it.
Speaker A
Here's the rule I use on everything of mine that runs on a schedule. Write the draught, never send it, and tell me it's ready.
Speaker A
Then I read it, change anything that's off, and press send myself. That's 10 seconds of my time.
Speaker A
And then nothing goes out in my name that I haven't read. After you've edited a draught, ask your agent to look at what you changed and remember it so the next one comes back, needing even less fixing.
Speaker A
Hand things over in this order. Let it read and organise for a few weeks first.
Speaker A
That means sorting, labelling, and pulling out the things you need to see. Then let it draught.
Speaker A
Sending on its own is the last step, if you ever get there. The job we automate in the next level never sends anything at all.
Speaker A
It reads confirmations and files them. And that's it. Google is just one example of a tool you can connect.
Speaker A
Think about the one tool you spend the most time in and connect that one next.
Speaker A
Many tools now speak a standard called MCP, which is a plug that lets an agent read from and act inside that tool.
Speaker A
Hermes has a whole catalogue of ready-made connections. And you can add others by simply pasting a link.
Speaker A
In the Hermes desktop app, go to capabilities on the left side menu. And then, from here, you'll see a couple of different tabs.
Speaker A
There's skills, tools, MCP, and plugins. If you go to MCP, you can see the full existing catalogue of pre-built MCP connections that you can simply toggle on.
Speaker A
All you need to do is authenticate the apps. For example, if you wanted to connect your Canva, you would simply click instal.
Speaker A
Then you can scroll up when it's done installing. And you can see it in the list of servers that are available to you.
Speaker A
Simply click on it. And then click authenticate. And then connect your account. Doing this will enable Hermes to work in your Canva account, just like we set up with the Google Workspace.
Speaker A
I'm actually not going to connect this though personally. And you can toggle this off anytime by simply hitting the toggle.
Speaker A
Or you can delete it entirely by clicking the remove button. The trash can icon.
Speaker A
For you, an MCP connection might be Notion, GitHub, your customer database, or even your accounting software.
Speaker A
Simply connect it, decide what it's allowed to touch, and then give it jobs inside that tool.
Speaker A
Let me show you an example of Hermes doing a job inside the Google Workspace tool we already connected.
Speaker A
Back in the app, let's start a new session. And then I'll say, "Look at my calendar for next week and add a 30-minute block called change studio light bulbs on the least busy afternoon." Now, before I actually run this,
Speaker A
let me show you what the calendar currently looks like so you can see how it changes.
Speaker A
This is my calendar next week. So I imagine it'll find this Friday slot. But I want to see what it comes back with.
Speaker A
Let's go ahead and run that prompt. It'll think for a few minutes. Let's pop over to Google Workspace.
Speaker A
And just like that, you can see change studio light bulbs appear in my calendar.
Speaker A
It went into my calendar and added the block all itself. That's level three. So here's your level up task.
Speaker A
Connect one tool you use every day and give your agent one real job inside it.
Speaker A
Level four is where we stop repeating ourselves. Level four is training it. You've probably had this experience with AI.
Speaker A
You correct it. It fixes that one reply. And then the next session, it's already forgotten.
Speaker A
And you're correcting it again. Watch what happens when you correct Hermes. As an example of something simple that I might correct Hermes on, I'll have it write a short social media post.
Speaker A
I'll say, "Write a short post about studio lighting for complete beginners." And, here, you can see, it comes back with headings, bullets.
Speaker A
And it's quite wordy overall. It's not as short as I'd like. So let's give it a correction.
Speaker A
I'll say, "Too wordy. I prefer short, punchy sentences. Three paragraphs max. No headings or bullets." Save that as a preference.
Speaker A
You can see it saved that to memory. It said it was saved and then came back with a much better structure.
Speaker A
I told it plainly the way you would tell a new hire as you're training them.
Speaker A
And then it wrote that down in its memory. I didn't have to go find a settings page or edit anything.
Speaker A
I just told it. So, now, in the same session, we'll say, "Now, write one about microphones for video production." And this is just another example to show you what happens in the same thread.
Speaker A
And there is the short post. It matches the original style. But, now, watch what happens when we go to a brand new session.
Speaker A
And this is where the real magic happens. Now, you might create new sessions pretty often.
Speaker A
So you can either click new session on the left side, or, at any point, you can start a new session by hitting Command + N on your keyboard, or, on Windows, that would be Control + N.
Speaker A
So I'm going to do that. It's a good keyboard shortcut to note because you'll probably use it often.
Speaker A
Now, I'll say, "Write me a short post about roasting coffee." And there's the exact same structure again.
Speaker A
That was a brand new session with nothing carried over. And it's still in the shape I asked for once in a completely different session.
Speaker A
Any chatbot can hold a preference inside of one conversation because it has the context of that one conversation it reads through when it responds.
Speaker A
This one, on the other hand, holds it across every conversation on every device from now on.
Speaker A
The AI model doesn't remember anything between messages. Every time you send one, it reads a fresh briefing on it.
Speaker A
Who it is, what it knows about you, its notes, and the conversation so far.
Speaker A
Picture a brilliant assistant with total amnesia who keeps excellent notes. Every morning, they walk in with no idea who you are, read their notebook, and then get to work as if they'd never left.
Speaker A
The notebook is the memory. In the Hermes desktop app, you can access the memory by going to settings with the gear icon in the upper right or Command + ,.
Speaker A
On the left side, in the menu, click memory and context. And then, here, you can see all of the memory settings.
Speaker A
There are two main things I want to point out. Persistent memory and user profile.
Speaker A
If you want to see the actual memory files, you can go to the web dashboard.
Speaker A
And then go to the files section in the left side nav. Find the memories folder.
Speaker A
Open it. And then, here, you can see the MEMORY.md, the markdown file for the memory, and then USER.md, which is that other one, the user profile.
Speaker A
And to see them, you can simply click download. I'll save it to my downloads folder here.
Speaker A
I'll open it in finder. Markdown files can be opened in lots of different software options.
Speaker A
If you have a coding tool on your computer, it might default to that. If you have Notion, it might default to that.
Speaker A
But you can always right click and choose exactly what you want to open it with.
Speaker A
In my case, I actually just want to open it with text edit. And, here, I can just see the raw text of the markdown file.
Speaker A
So this is the full memory here. Some things that's remembered so far, for example, are that it has access to a Google Workspace account.
Speaker A
And it lists the email address. And all the tools that it has inside that workspace.
Speaker A
Every time the agent wakes up, it looks at its memory and goes, "Oh, look, I have access to Google.
Speaker A
Cool. I can use that to do my job." Then it looks at the message and says, "What's my job?" Let me explain memory versus user in a bit more detail.
Speaker A
There are two notebooks, like I said. One is about you. It's your name, how you like things written, and what you expect from it.
Speaker A
That's where the correction we just made went. The other one is the agent's own working notes.
Speaker A
What it has learned about your setup, the conventions you use, and the things it got wrong once and then fixed.
Speaker A
You don't write either of them. Your agent does as it goes. And you can read both, like I just showed you.
Speaker A
And if you want to, you could ask it to change things specifically. Both of these are short on purpose.
Speaker A
Just a few hundred words each. Because each word gets sent with every message. So the notebooks hold the things that matter most.
Speaker A
And for everything else, your agent has two more places to look. It can search every conversation you've ever had with it.
Speaker A
And it can read the files and skills it has saved. All of it lives on your server.
Speaker A
I just showed you where you could go to find files. There's a whole lot more files in there.
Speaker A
Now, here's how I run my own personal agent. When the notes fill up, the memory and user, ask it to keep an index.
Speaker A
That's one file that lists where everything else lives. Kind of like a table of contents for all your files.
Speaker A
Then the short memory only has to say start here at the index. And it looks up the rest when a job actually needs it.
Speaker A
Let me show you how you could set up your own index. We'll send a prompt.
Speaker A
Something like this. "From now on, keep an index file of your notes. A table of contents that points to where each topic lives.
Speaker A
And put a one line pointer to it in your memory." And after a minute of processing, it says done.
Speaker A
It created the index file and put it in our file system. It says it's a master index/table of contents, covering the memory, the skills it has, the load instructions.
Speaker A
And it added that detail to the memory. If you wanted to, you could even ask to see the index.
Speaker A
I just said, "Show me the index file." And, here, it came back with a response, showing everything that it included in the file.
Speaker A
So you can go through it and ask it to change things if you want.
Speaker A
It gave us the explanation for where the file lives. So if you wanted to find this in your web dashboard, you'd go to the folder opt/data and then find NOTES_INDEX.md.
Speaker A
So, here, you can see we're already in the file system at /opt/data. So then all we need to do is find that index file.
Speaker A
Scroll down. And here it is, NOTES_INDEX. If we click, we can download it. I'll open it in finder again.
Speaker A
Then open in text edit. And there it is. There's the index file. And it'll change this on its own as time goes on.
Speaker A
Memory holds facts and preferences. Skills hold procedures. A skill is like a recipe. A written list of steps for how to do one job that your agent only reads when that job comes up.
Speaker A
Let's turn what we just taught it into a skill. Back in the Hermes desktop app.
Speaker A
I'm going to go find that post writing chat we had earlier. Go ahead and click that.
Speaker A
And then I'll say, "Save that editing workflow as a skill called concise-writer." And after a minute of thinking, it comes back, saying it saved that as concise-writer and gave the location of the file.
Speaker A
It explains how it encoded the loop it ran on both posts and turned that into the skill.
Speaker A
Now, if we go to the capabilities section in the left side nav, and then look through our skills here, if we scroll down to where it would be alphabetically, we can see here, concise-writer.
Speaker A
Any skills you asked it to learn or any skills it learned on its own, you'll see listed with the Learned tag.
Speaker A
The ones without the Learned tag are the ones that are built in. Here, you can click to open that skill.
Speaker A
I'm going to hide this hub down here. And, here, you can see all the text of the skill written out.
Speaker A
The cool thing is that the skills are just a text file. Like a recipe.
Speaker A
Instructions the agent follows. Now, your agent doesn't just follow those steps blindly. Imagine you ask it something like to check a supplier's price page every week.
Speaker A
One week, the supplier redesigns their website. And the old steps don't find the prices anymore.
Speaker A
Instead of failing every week from then on, Hermes works out the new path and rewrites its own steps so that the next run just works.
Speaker A
It also writes new skills on its own as it solves things. And those show up with the Learn label I showed you.
Speaker A
There's a background process that keeps the library tidy. So skills you stopped using get filed away instead of piling up.
Speaker A
Now, let me show you how you can take any document, webpage, or PDF and turn it into a skill that Hermes can use later.
Speaker A
Here, I found this SEO starter guide. It's a whole PDF on how to optimise copy and web pages for SEO.
Speaker A
This is a great procedural document. So I'm going to take this. Drag it into Hermes.
Speaker A
And then we'll type the command /learn. And then, since the PDF guide is attached, we don't have to give it any additional context other than might want to say, "Turn this PDF into a reusable skill." If you don't have a file you're handing it,
Speaker A
and, instead, you have something like a webpage, you can also just give it the link.
Speaker A
Let's go ahead and send this. And, here, we can see, after a few minutes of thinking, it said, "The skill was created and verified." We have google-seo-starter as the skill that we can call any time.
Speaker A
So, for my project, the skill I wanted it to learn was SEO. For your project, it could be your company's style guide or a course syllabus.
Speaker A
Now, you don't have to make all your own skills. You can also go find preexisting skills.
Speaker A
And there's a couple places to find these. Let's go back to capabilities. And then, here, in the skills page, if you scroll through and go to under where all of the toggled on skills are, there's a bunch of skills that are available to instal.
Speaker A
You can simply click instal on any of these skills that seem interesting to you.
Speaker A
There's also something called Skills Hub. I minimised this earlier. But you can see here at the bottom of the skills page, there's this little text that says, "Skills Hub." And if you click browse the full hub, it opens this little window in the bottom here,
Speaker A
which I'll go ahead and expand now so you can see it more. Skills Hub is a whole library of skills that other people have written.
Speaker A
Now, there's one caution here. Because a skill is just instructions. And you're trusting whoever wrote them.
Speaker A
Hermes scans everything in the hub for things like hidden instructions and data theft. And it blocks dangerous ones outright.
Speaker A
But it won't catch everything. So, as you're browsing, read a skill before you rely on it.
Speaker A
Or even ask your agent to review it for you. And stick to the official trusted sources while you're learning.
Speaker A
Now, let's give your agent a personality. And I think the first one of these things I'm going to show you is really fun.
Speaker A
So if you go into any chat, it could be a brand new one or an old one, you can type /personality.
Speaker A
And then click the personality command. And it opens up a whole list of options, including helpful, concise, creative, teacher.
Speaker A
Even things like catgirl and pirate. So let's go ahead and click pirate. And then submit that command.
Speaker A
After a second, you'll see the personality was set to pirate. And let's ask it something.
Speaker A
I'll say, "Write me a short essay about the sun." And, here, we get a short essay about the sun written in the style of a pirate.
Speaker A
"You'd think a ball of fire 93 million leagues away wouldn't matter much to a mortal.
Speaker A
But there's not a wave that rolls, not a tree that grows, not a breath ye take that isn't the sun's doing." So that's kind of a fun feature.
Speaker A
You can always set it back to a normal personality by sending that command again and choosing none.
Speaker A
Now, we're back to the base agent behaviour. So those presets are temporary. They change how it talks for a session.
Speaker A
And they don't touch anything permanent. The permanent version is a file called SOUL.md, which holds who your agent is and how it behaves.
Speaker A
We already talked about memory, which is what it has learned about you and your work.
Speaker A
Soul is the character it brings to that work. Keep it short. And you can change it by asking.
Speaker A
For example, we can say, "Update your SOUL.md so you always answer in a calm, direct tone.
Speaker A
And never end a reply with a question unless I asked for options." And, here, you can see.
Speaker A
It updated its soul. Its own long-term permanent personality. If I send another message, maybe we'll have it do something online.
Speaker A
We'll ask it something like, "Find the best kebab shop in London." Here, you can see.
Speaker A
It keeps it calm, direct, and avoids the typical AI behaviour of asking a follow-up question at the end.
Speaker A
We've successfully updated its soul. Now, that's level four. It knows how you like things done.
Speaker A
It keeps its own procedures. And it sounds the way you want. So here's your level up task.
Speaker A
And pick whichever one annoys you most. Correct it on something it keeps getting wrong.
Speaker A
Tell it one thing about you it should always remember. Or point the learn command at a document you're tired of explaining to people and let it turn that into a skill.
Speaker A
It takes one minute of teaching. And you never repeat yourself again. Level five is where your agent starts working while you're away.
Speaker A
Hermes ships with 16 ready-made automations. And I went through all of them. Every single one sends you something to read.
Speaker A
A morning brief, a news digest, and a weekly review. That's the problem with a lot of AI automation.
Speaker A
It gives you more to read instead of less to do. So let's build the one that isn't like that.
Speaker A
We're going to give your agent an actual chore. The kind of thing you do every week and hate.
Speaker A
And then you're never going to do it again. Mine is filing receipts. Here's my inbox with a whole bunch of receipts.
Speaker A
Now, every purchase confirmation and invoice that lands on my email has to end up in a spreadsheet at some point.
Speaker A
It takes 20 minutes a week. And it's pure admin work. Now, this will run on the agent's own Google account.
Speaker A
The one we set up in the last level. So it never touches my personal inbox.
Speaker A
Here's the whole thing in one message. This is kind of a long prompt. I'll go ahead and read through what I'm sending here out loud.
Speaker A
And then I'll walk through it before I send it. So I'll say, "Set up a receipt filing clerk for me.
Speaker A
First, set up a Google Sheet called Expenses with a Receipts tab if I don't have one yet.
Speaker A
Then every day, at 7:00, search my Gmail for purchase receipts, order confirmations, and invoices from the last two days that aren't already labelled Files.
Speaker A
For each one, pull out the vendor, the date, the total, and the currency, and append a row to my Google Sheet Expenses on the Receipts tab with columns for date, vendor, amount, currency, category, and the email ID.
Speaker A
Then add the Files label to that email so it never gets counted twice. This is a pre-approved recurring job.
Speaker A
So append the rows and apply the label without asking me each time. Never send, reply to, forward, delete, or archive anything.
Speaker A
Reed and label only. When you're done, message me one line on Telegram with how many you filed and the running total for the month.
Speaker A
Send me a link to the spreadsheet that I can open from any device. Create the job paused so that I can look at it before it goes live." Okay, so there are three parts of the message doing some work.
Speaker A
First, look at the words read and label only. I've told it what it may not do in plain language.
Speaker A
And everything it can do is additive. The worst case is a wrong row in a spreadsheet, which I delete.
Speaker A
Second, look at the label. That's how it knows what it has already filed. So if it runs twice, it doesn't double up.
Speaker A
Your agent starts every scheduled run with no memory of the last one. So a job like this needs something in the real world to check against.
Speaker A
Third, create it paused. It builds the job, but doesn't switch it on so you get to read it first.
Speaker A
That's that progressive trust again. All right, I'll go ahead and fire this message off.
Speaker A
Okay, so instead of just fading from the prompt being sent to the job being done, I'm fast forwarding this so you can see how much thinking and work the agent put into this.
Speaker A
In real time, this took about 10 and a half minutes to complete and set up the job before it came back with this final response that you see here now.
Speaker A
Here, it says it's done, set up, tested, end to end. And if we go and right click on this link and hit open in external browser, we can now see all of the receipts are now identified in the spreadsheet.
Speaker A
If we go into the inbox, we can see that all of the emails have been labelled Files exactly as we asked.
Speaker A
So if we go to the label section on the left side of the inbox and click Files, all the receipts are here.
Speaker A
Now, it ran the job once so it could test everything. But it won't do it going forward until we turn the recurring job on.
Speaker A
So, in Hermes, let's go to scheduled jobs in the left side now. And then, here, you can see, at the top of the list, there is this paused item.
Speaker A
You can tell it's paused because it has the orange indicator in the list. And it says pause right here next to the name.
Speaker A
Here's the receipt filing clerk. It has a prompt that you can read. And what it did was it actually built a deterministic script.
Speaker A
So when it calls this job, it's not actually going to call the LLM, which is great for us because it saves us tokens.
Speaker A
It does the same job the same way every single time. And it's not going to cost us anything on our Nous Portal except the initial trigger that starts the task.
Speaker A
So this will be really cheap in the long run for us. To turn this on, we can simply hit resume.
Speaker A
And then if we wanted to, we could click trigger now. That'll run the job manually so you can see how it works.
Speaker A
Now, here, it failed to trigger the cron job because it said it's already running or claimed by another scheduler.
Speaker A
You might get errors like that sometimes. And that's okay. Honestly, it ran the job anyway.
Speaker A
But you can see how, now, it'll work, going forward. Every day at 7:00 PM, it'll look at the last two days worth of emails and grab all the receipts.
Speaker A
Now, think about your own week. Anything you do on a schedule that follows a rule and that ends in a file, or a folder, or a calendar is a great candidate for this.
Speaker A
You could have a tracker that updates itself from your confirmation emails. 15-minute prep blocks added automatically before tomorrow's meetings.
Speaker A
Or a weekly folder created with the week's notes already inside of it. Just about anything you can think of can be turned into a scheduled job.
Speaker A
That job did a chore. This one watches something and only interrupts you when there's a reason.
Speaker A
I'm going to tell my agent what I care about once. It'll write its own skill for checking, schedule itself, and then, from then on, my phone only buzzes when something's actually new.
Speaker A
In a new session, let's say something like this. "Watch Barcelona to London flight price trends for me.
Speaker A
Build yourself a reusable skill for the check. Keep track of what you've already shown me.
Speaker A
Schedule it every few hours. And message me on Telegram only when there's something new, written with your concise-writer skill." So this job was a cool one to watch.
Speaker A
This one actually took about 30 to 45 minutes. Somewhere in there. I actually stepped away and came back because it was taking a while.
Speaker A
And the reason for that was I didn't give it super, super specific process instructions like I did for the email ones.
Speaker A
So the agent went onto the live internet, figured out the best way to do this job, built the machinery, built a skill, put it on a schedule, and also set up some memory so it can keep track of what it's already shown me.
Speaker A
And then, here, after all that work, is its first response. Basically saying that it'll message on Telegram only when something changes.
Speaker A
If we go to the scheduled jobs, we can see, now, we have this Barcelona-London flight price watch.
Speaker A
Cool thing is too, it's another deterministic script job. So it'll run this script every six hours, fetching the latest prices.
Speaker A
And then you can see, unprompted, after something changed, this is what it looks like when the flight watch cron job finds something new.
Speaker A
I described an outcome once. And, now, it comes to me. My own agent runs more than 20 of these.
Speaker A
One checks my bookkeeping every Monday. One tracks birthdays and starts researching gift ideas while there's time to order.
Speaker A
One works on a trip I'm planning twice a day. Getting quotes and coming back with the best option.
Speaker A
And one follows up on a negotiation so I don't have to remember to do it.
Speaker A
Now that your agent runs jobs on its own, the model choice starts to matter.
Speaker A
Because you're paying for every one of those runs. Models are like workers you can hire by the job.
Speaker A
Some are brilliant and expensive. Some are quick and cheap. And each one can work at different effort levels.
Speaker A
You wouldn't hire the most expensive person you know to sort your mail. Here in Nous Portal, on the left side, if you go to the resources section, and select models, here, you can take a look at the active promos for various models.
Speaker A
How much they cost normally versus how much they cost currently. And, occasionally, there are deals where you can get more power for less money by switching to some of these discounted models.
Speaker A
Now, the more important list is actually the one below. This all models list here.
Speaker A
Here, you can scroll through or even filter and see the list of all the different models offered through Nous Portal and see the portal price.
Speaker A
Here, for example, we're using DeepSeek V4 Flash. And that's seven cents per million tokens in and 14 cents per million tokens out.
Speaker A
Claude Fable 5.1, for comparison, is $10 per million tokens in and $50 per million tokens out.
Speaker A
That is a very expensive model. But it's very powerful. You can also see that there's some free models listed in here.
Speaker A
But keep in mind that not all models perform the same. So you may need to do a little bit of experimentation to see which ones work best for you.
Speaker A
Routine work that follows a skill, like checking a feed or filing receipts, runs just fine on a cheap model, though I'd avoid the very cheapest and the free ones because a model that can't reason will make confident mistakes.
Speaker A
Like filing the wrong thing or throwing away something that you wanted. Work that needs judgement , like planning, analysis, or anything where being wrong costs you, deserves a top model.
Speaker A
And Hermes honestly feels dramatically better on a flagship model, especially one that can hold a lot of information at once.
Speaker A
So if you're underwhelmed by the experience you're getting inside of Hermes, before you blame the tool, try a stronger model or a higher effort setting.
Speaker A
I'll show you where you can switch those now. In the Hermes desktop app, on the right side of the text box where you type your messages, there's a model selector.
Speaker A
If you click that, it'll expand to show you all of the different models that are available to you on your connected provider.
Speaker A
In this case, Nous Portal. So if I had a task that I needed to do that was really, really important, I might switch up to using Fable 5.1.
Speaker A
And then I could send my message here. To the right of that, there's a dropdown that if you click to expand, it shows you the effort level.
Speaker A
Here, right now, Fable 5.1 is set to medium, which, for most tasks, is good.
Speaker A
But you could also scale it up to ultra. Just keep in mind. Ultra on something like Fable 5.1 is going to cost a lot of money.
Speaker A
If you switch back to something like DeepSeek. And, here, I can type to search for it.
Speaker A
DeepSeek V4 Flash. We'll use this one. You could also change the effort level here.
Speaker A
So if you feel like you're not getting quite as much out of DeepSeek as you'd like, you could also increase the effort level to something like ultra or max.
Speaker A
But for everyday chatting, that's kind of overkill. Now, I'll show you how to update that in Telegram as well.
Speaker A
In the Telegram chat, type / and then model. And that'll open up your model configuration.
Speaker A
So, here, you can toggle to see all the different models that are available to you.
Speaker A
So let's say I switched to Fable 5.1 here, for example. Then you can do / and then reasoning.
Speaker A
And this is where you can select your reasoning level. They label it reasoning instead of effort here.
Speaker A
I'll just stick with medium. Now, there are two things I've run into myself that I want to share with you so that you can avoid any surprises in your own setup.
Speaker A
First, you can switch models at any time. And I actually love that about Hermes.
Speaker A
You can change the model and the effort level mid-conversation. But I should point out that switching models in the middle of a long conversation makes the new model reread everything in the conversation from the start, which costs more.
Speaker A
So for a big job, start a fresh thread on the new model. Second, every scheduled job remembers which model it was created with.
Speaker A
If you change your default model later, an unpinned job, meaning it wasn't pinned to a specific model, will refuse to run, rather than quietly switch to a more expensive model.
Speaker A
That's deliberate. And the fix is to pin the job to a model on the cron page.
Speaker A
Now, where you're going to find that is go to the scheduled job section of the desktop app if you're using that.
Speaker A
And then find your cron jobs in the list on the left. Click the three dots next to the name of it.
Speaker A
And click edit cron. Here, you'll see a model selector for jobs that use an LLM.
Speaker A
Now, these are actually both deterministic script jobs in my cron list here. So these don't actually call the LLM.
Speaker A
So there's nothing to switch and select here. So these actually work just fine for me.
Speaker A
But if you have a job that calls a model, you should have a dropdown menu in here to select which model you want to use.
Speaker A
Now, while we're on this subject, let's set up a fallback. Models stop working sometimes.
Speaker A
And plans hit their limits. Usually, while you're asleep and a scheduled job is running.
Speaker A
So give your agent a short list of models it drops to when the first one won't answer.
Speaker A
In my own setup, mine goes from a top model to a second one from a different company.
Speaker A
And then, finally, to a cheap one that's almost always available. It switches mid-conversation without losing anything so that you can keep going.
Speaker A
And your scheduled jobs use the same list. To do this, we'll go back to the chat.
Speaker A
I'll start a new session. And then say something like this. And, here, I've got my sample prompt in brackets.
Speaker A
This is where I would recommend you go through and update your prompt to match your own desired setup.
Speaker A
So, here, I have it saying, "Set up a fallback chain. Try flagship model first.
Speaker A
Then a model from another provider. Then a cheap reliable model. Switch each of these for your own specific desired fallback chain." A good way to figure out which ones you want to use is to use this model list inside your Nous Portal.
Speaker A
So, for example, let's say we wanted to use DeepSeek V4 Pro for our main one.
Speaker A
We can go ahead and copy this name from the list here. Paste it in for our flagship model.
Speaker A
Then a model from another provider. So let's find another brand name. Maybe we'll use something like Claude Sonnet 5, for example.
Speaker A
It's more expensive. But it's reliably available. So we can go ahead and grab Claude Sonnet 5.
Speaker A
And then paste that in for our model from another provider. And then a cheap reliable model.
Speaker A
And for your own setup, this may be where you include something like a free option as a last resort.
Speaker A
Maybe we'll use something here like this Upstage Solar Pro 4. And then place that here in for our final model.
Speaker A
Go ahead and hit send. Here, you can see, after a bit of processing, it comes back and says, "The fallback chain is configured." The primary is whichever model you're working with right now.
Speaker A
And then you have several different fallbacks. Okay, there's one more command for hard questions called mixture of agents.
Speaker A
In your chat, put /moa. And, here, you'll see it in the list. I'll go ahead and select that.
Speaker A
Put that in front of a request. And Hermes will send it to several different models at once.
Speaker A
Then it combines the best of each answer into one reply. You're paying for several answers instead of one.
Speaker A
So save it for decisions that deserve a second and third opinion. For my example here, I might do something like a benchmark test to see if these are actually good models for the fallback chain or use something different.
Speaker A
So I'll say something like, "I want to do a quick benchmark test. Are these the best models to have in my fallback chain if I want to have a flagship model, then reliable model, then cheap or free model for the chain,
Speaker A
all of which are capable of performing my tasks well?" Let's send it. Now, you'll see that the /moa command was recognised.
Speaker A
But after a second, it said that there was a provider error. So no LLM provider was configured for this.
Speaker A
So, to do this, we can actually click the switch provider button that popped up.
Speaker A
And then, in the model settings page that pops up, go ahead and scroll down to the mixture of agents section.
Speaker A
Then you can go in and select your reference models. So we can click in to select something from Nous Portal.
Speaker A
You could pick a specific model. Here, for our reference model one, let's say something like Opus 5 from Anthropic.
Speaker A
For the reference two, we'll grab something else from Nous Portal. We'll say ChatGPT 5.6 Luna as an example.
Speaker A
And maybe one more. We'll grab a DeepSeek model here from Nous Portal. Go ahead and grab DeepSeek V4 Pro.
Speaker A
We've got those locked in. And while we're in here, I'll point out too, if you scroll all the way to the bottom of this page, this is where you manage your fallback models.
Speaker A
Here, you can see that you could actually manage your fallbacks manually if you prefer.
Speaker A
One more thing to check actually too is your aggregator. So what model are you using to actually assemble all the answers that come back from each of those references into one final answer here?
Speaker A
So that'll be your aggregator at the bottom of this reference list. So we want to make sure that this is also selected to a model that's actually available to us.
Speaker A
I'll go ahead and select DeepSeek V4.1 Flash. It's essentially just taking all the best bits of each answer and summarising the final results.
Speaker A
So we'll go ahead and roll with that. Close out of this. And then let's go ahead and copy our prompt.
Speaker A
I'm actually going to open up a fresh session. Type moa again. Paste our prompt.
Speaker A
And go ahead and run it. You'll see. It gets to work. Pulling in the first of the three reference models.
Speaker A
And this can take a while. Now, I want to mention that this might be kind of expensive for credit usage because, not only are we using three different models and an aggregator to compile all the results, but I also selected a couple of more expensive models.
Speaker A
In the next chapter, I'll talk a bit more about costs. But, for now, let's go ahead and fast forward to when this is done.
Speaker A
This can take quite a while. After a bunch of thinking here, which took about 26 minutes, this is what it came back with.
Speaker A
It came back with this chart here, saying this is how it would suggest that I revise my particular fallback chain.
Speaker A
Go with Opus 5 first, DeepSeek V4, then DeepSeek V4 Flash, and then Solar Pro as the last resort.
Speaker A
This is obviously just an example. Models change all the time. And, really, you want to pair the model you're using to the task you're doing.
Speaker A
So this is great for my particular setup. But it's just illustrative. I'd recommend doing something like this that targets your own setup and purpose, rather than copying my example chain here.
Speaker A
Your agent has now done a chore, watched a topic, and used a few different models.
Speaker A
Let's look at what that cost. Here in the Nous Portal dashboard, on the left side, under account, go to usage.
Speaker A
And, here, you can see your total credit spend, your total input tokens, total output tokens, cache reads, cache writes, and total number of requests.
Speaker A
And you can see that over various periods of time here. You can look at it by day, for example.
Speaker A
You can use these filters to change to see it by model, see different chart styles, and even change your display period.
Speaker A
Now, look at the token count. This is the total amount of text going back and forth against the money.
Speaker A
So input tokens are the prompts you send. And output tokens are the outputs from the models.
Speaker A
Both inputs and outputs cost money and deplete your credit balance. Every time you send a message, your agent resends its whole briefing.
Speaker A
Its instructions, its memory, its tool list, and the conversation so far. That rereading is called context.
Speaker A
And it's almost your entire bill. But it stays cheap because the portal recognises the part that didn't change and charges a fraction to read it again.
Speaker A
That's called caching. You don't need to set that up. It just happens by itself.
Speaker A
As far as costs go, the server bill is fixed through Hostinger. So that's just fine.
Speaker A
But the big cost to watch is your model bill. I've been filming this tutorial over a couple of different days.
Speaker A
And, here, you can see my cost per day. Day one was a little bit more than eight cents.
Speaker A
And then, today, as I'm recording this chapter, and the chapter with MOA, mixture of agents, that's today, you can see, that MOA task alone cost me a lot of money.
Speaker A
The calls to Anthropic's Opus 5 cost $1.30. Compare that to all this usage from the previous days where I was just using DeepSeek V4 Flash.
Speaker A
This is way cheaper to run. You can see the cost broken down by model.
Speaker A
So as you're trying to figure out which model to use over time, like I said in the last chapter, make sure you're pairing the model to the task.
Speaker A
If you default to using something like Anthropic's Claude Fable 5.1 for everything, you're going to burn through a lot of money.
Speaker A
That and Astra are super helpful to call when the task demands it. But for most daily work, I recommend trying to find a model that achieves a good balance of cost efficiency and effectiveness.
Speaker A
This chart will help you determine that over time. So that's level five. Your agent does a chore.
Speaker A
It watches what you care about. And, now, you know what you're spending. So here's your level up task.
Speaker A
Pick one thing you do every week that ends in a spreadsheet, a folder, or a calendar.
Speaker A
And give it away. Ask the agent to do it for you. Create the job paused.
Speaker A
Read what it wrote. Then switch it on to run all the time. Level six is giving it a team.
Speaker A
The simplest version needs no setup at all. Ask your agent to split a job.
Speaker A
And it starts helpers that work at the same time and hand you one answer.
Speaker A
For example, I could say something like this. "Plan a three-day trip to Lisbon, Portugal from London, leaving November 5th.
Speaker A
One subagent on things to do. One on best places to eat for authentic local food.
Speaker A
And one on the best flights itinerary. Then give me a single combined plan." Here, you can see that the main agent that we're talking to is delegating the task to three subagents.
Speaker A
And then you can see those three subagents getting to work, each running on DeepSeek V4 Flash.
Speaker A
Now, you could also ask for your main chat to spin up these subagents on specific models.
Speaker A
So they don't all have to run on the exact same one. All these agents are now running in parallel.
Speaker A
So instead of using one model to do all the work, back to back, instead, there are three researchers going out at the same time, and then they'll all come back together, which makes this whole process really quick.
Speaker A
Now, of course, really quick is relative to the task at hand. So I'm still going to go ahead and skip forward to when it's done, just to save time for the video.
Speaker A
And here we go. After a few minutes, I've got this beautiful three-day itinerary for Lisbon.
Speaker A
The agent we're messaging just coordinated three others working at once. Now, those helpers are temporary.
Speaker A
They exist for that one job and then they're gone. For a standing team, you create teammates.
Speaker A
A teammate in Hermes is called a profile. A separate agent on the same server with its own memory, its own skills, its own personality, and its own model.
Speaker A
The desktop app presents profiles as bots with a name, a title, a job description, and a face.
Speaker A
It's the same thing underneath though. And it's a friendlier way to make them. So that's what we'll use.
Speaker A
Now, before I make them, I'm going to have my agent write the job descriptions.
Speaker A
Because it knows what the project board we build next will need better than I do.
Speaker A
In the main chat, let's say this. "I'm about to create three teammate profiles on this server for a project board, a researcher, a writer, and a reviewer.
Speaker A
Write a one-sentence job description for each one, worded so the board can tell from the description alone which of them a task belongs to.
Speaker A
Give me three lines that I can copy and paste." Let's go ahead and run that.
Speaker A
And here are the three one-line job descriptions. I'm going to go ahead and copy these and paste them into a text file real quick, just so I have them somewhere else quickly accessible.
Speaker A
Do a quick little format, just so I can see them. And there we go.
Speaker A
That saves a lot of trial and error. Because the board reads these descriptions to decide who gets what.
Speaker A
Next, we're going to go to the upper left portion of the app here and change from the sessions tab to bots.
Speaker A
Now, if you don't see this here, it might have moved somewhere else on the app.
Speaker A
Search around for it. Or even ask Hermes where that new bots tab is. When you get here, go ahead and click the plus button to add a new bot.
Speaker A
And then click new bot. And then, here, you can give your bot a face.
Speaker A
For example, you could grab from any of these existing bot shapes, click to generate a face by describing your avatar, upload a picture, or even choose from a bunch of pets.
Speaker A
Maybe, for example, I want this penguin. Yeah, that looks fun. Next, let's go ahead and give it a name.
Speaker A
Going back to my text edit here. Our first bot will be called researcher. So I'm going to go ahead and copy this job description.
Speaker A
Get it ready. Copy. And go the name. Researcher. Give it the title. Same thing. Researcher.
Speaker A
And then paste in the job description. You can click advanced, and then scroll down, and give it additional specifications if you'd like to.
Speaker A
If you don't do anything, it'll just clone everything from your original profile. Your default profile.
Speaker A
But, here, I want to set it so that the researcher always uses a higher power model.
Speaker A
So I'll go ahead and change provider to specify Nous Portal. Then model. We'll go ahead and choose Opus 5 for the researcher because it goes deep and can do long-running research tasks.
Speaker A
Everything else, we'll leave the same. We'll go ahead and click create bot. And, now, we have our researcher.
Speaker A
Go ahead and repeat that process for any other bots that you want to make.
Speaker A
I'll do this quickly for the writer and reviewer, choosing models appropriate for each type of task that they'll do.
Speaker A
Okay, I've created those additional two bots. Let's go ahead and say hello to our critic here.
Speaker A
Our reviewer. Just to prove that it works, I'll say, "Hello." It says, "Hey, nothing queued up for me to review right now.
Speaker A
So I'm just here, ready to go when you are." So it already knows its job specifically.
Speaker A
Now, I created the research analyst with a top model because judgement is its whole job.
Speaker A
The other two got cheaper models. And I put them all on different models, just so that we're not experiencing any bias introduced by one single model type.
Speaker A
This helps keep the agent system, as a whole, a bit more honest. Now, let's look at the board.
Speaker A
Hermes has a project board called Kanban that works like Trello. A board of columns that your jobs move through from left to right.
Speaker A
Every job on it is a card. The difference is that these cards don't wait for you.
Speaker A
You put a job on the board. The board splits it into steps, hands each step to the right teammate, and they work through the columns while you watch or while you sleep.
Speaker A
I'm doing this in the desktop app because that's where I work. And it's switched off by default.
Speaker A
So let's turn it on. Head back to the sessions tab. Then go to capabilities.
Speaker A
Then go to the plugins tab. And then, here, you'll find Kanban, like I said, toggled off by default.
Speaker A
Go ahead and toggle it on. Now, it appears in the left side column. Go ahead and click to open it.
Speaker A
And here's the Kanban board. Now, you can't see any columns yet because there aren't any tasks on this board.
Speaker A
So let me just go ahead and quickly create a dummy task. Something with nothing on it.
Speaker A
So you can see how this works. I'll just say, "Test task." And then create task.
Speaker A
Now, here, you can see some columns. Triage is where raw ideas land. And the board turns them into a proper description of the work.
Speaker A
To do is waiting on other cards. Ready means it can start. Running means a teammate has it.
Speaker A
Blocked means someone needs you. Review is a second agent checking the work. And done releases whatever was waiting on it.
Speaker A
There's also a scheduled column, which tasks might get put in if they need to run at a specific time.
Speaker A
Now, three of these columns, you can't drag a card into because the board sets them itself.
Speaker A
Here, you can see. It just picked up the task and put it into running.
Speaker A
Now, if you're using the web dashboard instead, let me show you where to find this.
Speaker A
Here, on the left side, scroll down in the nav bar. And then Kanban is here at the bottom under plugins.
Speaker A
And this will show you all of those different columns right away. Now, before we put a real job on this, here's something I learned from my own mistakes.
Speaker A
It runs quietly in the background, which sounds great until a card gets stuck waiting on you and nothing tells you.
Speaker A
You'd come back an hour later, thinking it had been working the whole time. So there are two fixes.
Speaker A
And I use both. The first is to create your cards from Telegram with the kanban command instead of clicking new task in the app.
Speaker A
Because a card created that way subscribes you to it automatically. And you get a message when it finishes or gets stuck.
Speaker A
The second is to write the notification instructions into the job itself in plain language.
Speaker A
The same way you'd tell a person who to call if they get stuck. Here in Telegram, type the slash command kanban.
Speaker A
And, here, it shows you how to actually use this command. This command takes a bunch of subcommands.
Speaker A
The most important one is create. Now, this looks like a lot of information. So if you're not quite sure how to use this, what you can do is simply ask Hermes to tell you.
Speaker A
I'll say, "Show me an example of how to write a create command for the Kanban board." And, here, it comes back with an example.
Speaker A
So if we scroll up, we can see it from the beginning. It says, "Here's a clean create command." And I actually like the minimal version here.
Speaker A
It says, "Title is the only required argument." So you would type /kanban create. And then give it a title.
Speaker A
So let's go ahead and type a command. We'll say /kanban. And, you'll notice, it doesn't select a command from the list of commands in Telegram.
Speaker A
That's because this is a less commonly used command. That list only contains really common ones.
Speaker A
So don't worry about it if it doesn't look like it selects anything. So then we'll type create.
Speaker A
Then, in quotes, we'll give the task a name. We'll just say, "TEST TASK". Then we'll say, "--".
Speaker A
And, here, there's this issue sometimes if you're typing this on a computer where it merges those dashes.
Speaker A
So here's one quick trick around that. You can type -, and space, -. And then delete that space.
Speaker A
And, now, it works as two dashes. Then we'll say, "Assignee." And then just say, "Default." That way, it just assigns it to the default agent.
Speaker A
All right, we'll go ahead and submit that. And, here, you can see, it created the task.
Speaker A
And it subscribed us to it so that we'll be notified when it's completed. Let's go ahead and check it out in the desktop app to make sure it actually showed up there.
Speaker A
Here in the desktop app, I don't see it. So this is when you can refresh the app.
Speaker A
So, on Mac, you can hit Command + R or, on Windows, it'll be Control + R.
Speaker A
And here it is. The test task is running. So that Telegram command worked for us.
Speaker A
Back in Telegram, when it's done, we'll get a notification. And there we go. The task is complete.
Speaker A
There was nothing really to do. So it came back pretty quickly. Back in the desktop app, we can see the test task is now in the done column.
Speaker A
So that's how you do it from Telegram. Now, let's set one up from the desktop app, which is how I tend to actually do this myself.
Speaker A
Let's give it a real job this time. I'm going to use the board for something I actually use my agent for.
Speaker A
Planning a trip, end to end, including the part I always put off, which is tracking what's actually been booked.
Speaker A
Now, I did a dumbed down version of this earlier, showing you how you could run a command and ask for some subagents to do something.
Speaker A
But, here, I'm not going to tell it exactly how to do the work. I'm just going to give it the task.
Speaker A
And it'll figure out who to assign it to. So let's go ahead and click new task in the upper right.
Speaker A
And I'll say, "Lisbon Trip Plan" for the title. And, for the description, I'll say, "Plan a five-day trip to Lisbon for two people in early spring on a mid-range budget.
Speaker A
Research flights from London, three neighbourhoods to stay in, and three day trips, all with real prices and a direct booking link for anything I'd have to buy.
Speaker A
Build a budget spreadsheet in my Drive Outbox folder with a row for every item.
Speaker A
What it is, the price, the booking link, a booked column, and an actualized column for what I really paid.
Speaker A
Write a day-by-day itinerary next to it. Put placeholder events on my calendar for the travel days and anything time specific.
Speaker A
Check everything for errors and missing prices before you show me. Message me on Telegram when it's done.
Speaker A
And message me there if you get stuck and need an answer from me." All right, that's the full prompt.
Speaker A
It's pretty elaborate. This is where I recommend thinking about a complex big task of your own and how you might want to hand it off to an agent.
Speaker A
Think about the exact deliverables you need. Think about the format you want the work in.
Speaker A
And think about any context it needs. Once you've got all that dumped in there, go ahead and click create task.
Speaker A
Right away, you'll see it goes into the triage column. And it might sit here for a minute, especially with more complex tasks, as the orchestrator agent figures out how to break this task down into smaller pieces and delegate the different subtasks to various agents.
Speaker A
And after a few minutes, you can see it decomposed the task into smaller subtasks.
Speaker A
Now, three of them are running. And then three are in the to-do list below the big Lisbon trip plan.
Speaker A
So, right now, it's currently researching flights, neighbourhoods, and day trips. You'll notice, if I click into any of these items, like, for example, at the placeholder calendar events, it has dependencies.
Speaker A
So it's blocked by this task here. And it blocks these two tasks. It can't start until this is done.
Speaker A
And then these two tasks can't start until this one's done. The main agent is coordinating all the tasks.
Speaker A
Moving them along the pipeline. And so, now, for the most part, this is just going to work in the background.
Speaker A
And we can step away for a while. But before we do, let me just show you one more thing.
Speaker A
If I click into any of these research tasks, you'll see here that this research job was assigned to the researcher.
Speaker A
If I click the drive budget, it kept the default assignee. The review task was assigned to the reviewer.
Speaker A
So the specific agents and models that are best for each task have been selected, based on their individual expertise.
Speaker A
Every card keeps its own history and comments too. So if you scroll down through any of these, like, for example, this currently running task, you'll see the activity logs here.
Speaker A
You'll see all of the work that it has done. And you can even see it's working in real time right now, running commands.
Speaker A
You can add comments to the comments section here if you need to give it some notes about the task.
Speaker A
And the agent running the job, as well as the orchestrator, will add comments in here too as this card moves back and forth between various columns.
Speaker A
This is a full project management ecosystem, managed entirely by an orchestrator AI delegating the tasks to individual experts.
Speaker A
It's almost like a little mini-company. Can a card wait for the hotel to email you back and then carry on when the reply arrives?
Speaker A
Yes. So a teammate that can't finish moves the card to blocked and writes down why.
Speaker A
It'll put that note in the comments. So if you have a job where the task is to manage an email conversation, for example, you need a second job to wake it up.
Speaker A
Because an individual task can't just sit there, watching your inbox. Each task is a job that runs and then ends.
Speaker A
That's why we'll create a second job. It's a scheduled job that checks the board.
Speaker A
Mine runs every half hour. It reads the blocked cards. And when the thing they were waiting for has arrived, it moves them back to ready and they carry on.
Speaker A
If a card genuinely needs me, the scheduled job messages me instead of leaving that card sitting there.
Speaker A
So let's build the simple version of that here. I'll create a new session in the left.
Speaker A
And then we'll say, "Every morning, check my email for booking confirmations for the Lisbon trip.
Speaker A
For each one, fill in the actualized column in the trip spreadsheet with what I actually paid, tick booked, delete the matching placeholder event from my calendar, and create the real one with the confirmation number in the description.
Speaker A
Add [email protected] as a guest on any event they're joining. Label the confirmation email Trip so it isn't counted twice.
Speaker A
Message me one line on Telegram with what changed. Create it paused so I can read it first." Now, here's one warning before you copy that word for word.
Speaker A
Adding guests to a calendar event sends those people a real invitation. So only put names in there when you mean it.
Speaker A
And check the job while it's paused. So that's the shape of every long-running project you'll build with this.
Speaker A
A board does the work. A scheduled job keeps it current and tells you when you're needed.
Speaker A
For your own work, the same board could run a weekly report, a batch of product descriptions, a content calendar, or a research and outreach list.
Speaker A
Two bits of housekeeping before we finish. Because you now own a small system. If anything ever misbehaves, it's the same three checks from level two.
Speaker A
Credit, gateway, and logs. And updates are one click. In the web dashboard, on the left side now, scroll down to system and select that.
Speaker A
Here, you can see a quick overview of everything that's set up and how it's running.
Speaker A
And if you scroll down to the gateway section, you can see that it's currently running.
Speaker A
But you can always click stop, restart it, or start it from this menu here.
Speaker A
Further down, there's even a backup section here. So you can create a backup by clicking this button, which is super helpful to do any time before you make a big change to your configuration.
Speaker A
You can also restore from backups from this section here too. For example, if you download a backup, you can then upload to restore the state from that file here.
Speaker A
Back in the Hostinger Docker Manager, you can also update your software from this menu.
Speaker A
Click on the three dots to the right of the application. Then click update. You'll get a little popup notification, saying your application is being updated.
Speaker A
And this updates the version of Hermes that you have running on your VPS, also known as the backend.
Speaker A
To update the desktop app, open settings. Go to about on the left side. Check if there's a new update available.
Speaker A
And if there is, click update now. It's important to note that the desktop software and the software running on your VPS both update separately.
Speaker A
Now, there's one more page. And this is a fun one. Open your web dashboard.
Speaker A
And, in the plugin section, in the nav, under Kanban, there is an item called achievements.
Speaker A
Go ahead and click on that. Hermes reads back through everything your agent has ever done and gives you badges for it.
Speaker A
There are 60 of them across categories like agent autonomy, debugging chaos, and tool mastery.
Speaker A
They run from copper up to Olympian. And some stay secret until you accidentally trigger them.
Speaker A
Mine here says I've unlocked 7 out of 60. I've discovered 53. And my highest tier is gold.
Speaker A
So I have a little bit of working up to do to get to Olympian.
Speaker A
Most of these achievements are for things I didn't even know I was doing. Once your agent has been running for a week, come back to this page and tell me in the comments which badge surprised you and how many you have unlocked.
Speaker A
I'll read them all. So look at what you've got now. You've got an AI agent on a server that's yours, reachable from your phone and your desktop, connected to your files and your accounts, and trained on how you like things done.
Speaker A
It runs a chore every week that you'll never do again. And it has a board where a team of agents work through a project on their own.
Speaker A
My advice from here is to start small. Ask it questions. Give it one real job.
Speaker A
Let it learn how you work. And once you trust it, hand over the thing you've been dreading all month.
Speaker A
Tell me in the comments which level you're on and which one you're building to this week.
Speaker A
And if you haven't set your agent up yet, go down to the description now and click the first two links.
Speaker A
The first one takes 10% off the server. The second takes $15 off your first month of models.
Speaker A
Thanks for watching.
Topics:Hermes AgentAI assistantopen source AIVPS setupno coding AIHostingerAI automationtask schedulingpersonal AI agentAI integration
