YT YouTube2Text

YouTube2Text → Transcripts → The Ultimate Beginner’s Guide to Hermes Agent

The Ultimate Beginner’s Guide to Hermes Agent — Transcript

by Metics Media · 20,310 words · 2,847 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Welcome to the ultimate beginner's guide
0:01to Hermes Agent. In this video, you're
0:03going to learn how to set up your own AI
0:05agent from scratch and how to actually
0:07use it. By the end, you'll have an agent
0:09that runs 24/7 on its own machine, that
0:12you can message from your phone and your
0:14desktop, that knows how you like things
0:16done, that works inside the tools you
0:18already use, and that takes real work
0:20off your to-do list while you're away,
0:22including a small team of agents that
0:24work through a task board entirely on
0:26their own. This guide is designed for
0:28complete beginners, so you don't need
0:30any technical background, and you won't
0:31write a single line of code. I'm Matt,
0:34by the way. I use Hermes every day. My
0:37own agent runs more than 20 scheduled
0:38jobs for me right now, from bookkeeping
0:40checks to trip planning, and hundreds of
0:43thousands of people have used my Hermes
0:44tutorials on this channel to get their
0:46own agent running. The truth is that the
0:48setup is the easy part. The hard part is
0:50knowing what to do after. If you treat
0:53Hermes like a chatbot, it feels like
0:55just another chatbot. It gets
0:57interesting when you treat it like an
0:58assistant that learns. So, this video is
1:01built as a ladder with six levels and we
1:03climb it from start to finish without
1:04skipping anything. In level one, we hire
1:07your agent and give it a home that stays
1:09on around the clock for just a few
1:10dollars a month. In level two, we reach
1:13it from anywhere with one scan of a QR
1:15code and a desktop app that feels like
1:16ChatGpt or Claude. In level three, we
1:19put it to work inside your own files and
1:21your own accounts. In level four, we
1:23train it and you'll watch a single
1:25correction stick forever. In level five,
1:28we hand it a chore you do every week and
1:29never do it again. In level six, we give
1:32it a team with a project board where you
1:33queue up work and your own agents run it
1:36on their own. Every level is timestamped
1:38below so you can come back to any step
1:40later. And if you get stuck, leave a
1:43comment. I read all of them and answer
1:44as many as I can. But before we set
1:47anything up, you need to know what
1:49you're actually hiring because Hermes
1:51can use the same AI models as ChatgPT
1:54and Claude. And the difference is
1:56everything wrapped around them. Let me
1:58show you what that means.
2:01Hermes Agent is a free open- source AI
2:03agent made by a company called Nouse
2:05Research. You pay only for the AI model
2:08it thinks with and for the machine it
2:10works on, which for us is a small rented
2:12server that costs just a few dollars a
2:14month. If you've used chat GPT, Claude
2:16or Gemini, you know the basic idea. You
2:19type in plain language and it answers
2:20you. Hermes can use those exact same AI
2:23models underneath too. So the biggest
2:24difference is everything built around
2:26the models. If you already pay for Chat
2:29GPT or Claude, some of this will sound
2:31familiar. Both of those can run a task
2:33on a schedule now. Both of them remember
2:35you. Both of them can read your email
2:37and send a message for you. If that's
2:39all you wanted, you already have it. But
2:43there are three things you can't buy
2:44from either of them. The first is your
2:47choice of model. ChatBT and Claude only
2:50let you use the models their own company
2:52makes. Hermes connects to any model from
2:55any company. So you can put a cheap one
2:57on the boring jobs and a powerful one on
2:59the hard jobs all inside the same agent.
3:03The second is ownership. Chat GPT and
3:05Claude are services you rent with weekly
3:07limits and features the company can just
3:09take away. Hermes is a free software
3:11that runs on a server you own. So no
3:14company can throttle it or switch off
3:16something you rely on. The third is
3:18where your agent lives. ChatBT and
3:21Claude only talk to you inside their own
3:22apps. Your Hermes agent gets its own
3:25Telegram, WhatsApp, or Slack account, so
3:27it can message you in the same app you
3:29already use every day. There's one more
3:32difference. In chat GPT or Claude, you
3:34make a project for each thing you work
3:36on, and everything it knows about that
3:38work stays inside that project. Hermes
3:41turns that around. There's one agent. It
3:43learns your whole world, and you point
3:45it out whatever you're working on. Over
3:47time, it writes down what it learns as
3:49skills, which are step-by-step notes it
3:52can reuse, so it gets better at your
3:53work specifically instead of starting
3:55fresh each time. And if you've heard of
3:57OpenClaw, this is the same category of
3:59tool, an always personal agent. If
4:02you're coming from it, you don't have to
4:04move anything by hand. Once you're set
4:06up, just ask your agent to import your
4:08open clock configuration, and it walks
4:09you through the whole thing. Now, I have
4:12one request before we start. Throughout
4:14this video, I'll use a few different
4:15examples, all based around things I'm
4:17working on in my life. Don't copy my
4:20examples word for word. Instead, run
4:22each prompt through your own topic,
4:24whether that's your job, school, or
4:26something you do at home. You'll learn
4:28much faster working on something you
4:30actually care about.
4:32Your agent needs a machine to live on,
4:34and that machine has to be three things.
4:37It has to be awake. Your scheduled jobs,
4:40the watching, and the messages it sends
4:42you all depend on the machine being on
4:44at 3:00 in the morning. And a laptop in
4:46a bag isn't on. It has to be separate
4:50from your things. An agent runs
4:52commands, installs software, writes
4:54files, deletes files. While you're
4:56learning, you want all of that happening
4:58somewhere with nothing of yours on it.
5:00No personal photos, no passwords, no
5:02sensitive work documents, and no browser
5:05logged into your bank. Think of it as a
5:07practice workshop. You're going to make
5:09a mess in it, which is what it's for.
5:11And if you ever break it badly, you
5:12delete the whole thing and start again
5:14in 10 minutes. And it has to be
5:17reachable. So your phone, your laptop,
5:18and your tablet are all doors into the
5:21same agent, which is what we set up in
5:22the next level. That machine is a VPS, a
5:26virtual private server. It's a computer
5:28in a data center that stays on all the
5:30time and that only you can log into. And
5:33the smallest one costs just a few
5:34dollars a month. You can install Hermes
5:36on your own computer instead, but then
5:38it only works while that computer is
5:39awake and it's running commands and
5:41deleting files right next to everything
5:43[clears throat] you own. So for your
5:46first agent, rent the server. Now, you
5:49can rent a server from plenty of
5:50companies, and on almost all of them,
5:52you'd be installing Hermes yourself from
5:54a terminal, which is the text window you
5:56may have seen programmers use. That's
5:58[snorts] the part I want to skip, and
6:00it's the reason we're using Hostinger.
6:02They have a one-click Hermes template
6:04that does the setup a beginner shouldn't
6:06have to touch. It installs Hermes, puts
6:09it behind a secure web address, gives
6:11you a login page, and adds backups,
6:13one-click updates, and a restart button.
6:16It's also one of the cheapest ways to do
6:18it. Throughout this tutorial, you'll see
6:20how everything looks while I'm
6:21recording, but software like this gets
6:23updated all the time, so a button might
6:26be somewhere else. Things might have
6:28different labels, and there might even
6:29be new features by the time you watch
6:31this. If something seems different,
6:33don't panic. Follow the general idea
6:36rather than the exact visuals.
6:39All right, let's get set up. There are
6:41just two things that cost money today.
6:43the server, which is just a few dollars
6:45a month, and a model account in a few
6:47minutes, which costs about the same as a
6:48streaming service. That's the whole
6:50bill. So, let's start by grabbing our
6:52Hostinger VPS. Use the link on screen or
6:55the first link in the description below.
6:57It stacks an extra 10% discount on top
6:59of any sale Hostinger is already
7:00running. So, using that link is the
7:02absolute cheapest way to get started.
7:04When you click on that link, you'll land
7:05on this page here. And if you click the
7:07choose plan button, it'll scroll you
7:09down to see the various plans that are
7:11available. Here you can see that our
7:12coupon has already been applied. So all
7:14the plan prices that are listed are the
7:16prices you'll really pay. I recommend
7:19taking the KVM1 plan. It's the smallest
7:22plan and it runs one agent comfortably.
7:24And if you ever outgrow it, you can
7:26simply upgrade later without
7:27reinstalling anything. So go ahead and
7:29click choose plan. You'll land on the
7:32cart. And here you'll need to make a few
7:33selections. The first one is your
7:35registration period. Longer terms are
7:38cheaper per month, but the discount
7:40needs at least 12 months. So 12 is the
7:43shortest I'd pick. There's a 30-day
7:45money back guarantee, whichever you
7:47choose. Then there's a box that's
7:49already ticked for ready to use AI. Go
7:52ahead and untick it. It signs you up to
7:55a different AI service you don't need,
7:56and we're bringing our own in a few
7:58minutes. Leave the rest of the add-ons
8:00alone, and then check the discount is
8:02showing on this page, and then click
8:04continue.
8:06On the next page, you'll need to
8:07register your account. Go ahead and
8:09register with Google, GitHub, or email
8:12address. I'll use an email address.
8:14Create a password and then enter a phone
8:16number if you'd like, but that's
8:17optional, so I'll go ahead and click
8:19register. Then on the next page, enter
8:21your billing address and your payment
8:23details to complete the transaction.
8:25I'll do this off screen and meet you on
8:27the next page. After payment, you'll get
8:28this confirmation and then Hostinger
8:31should route you straight into the
8:32Hermes setup flow with a form that gives
8:34you an admin username and a password. If
8:37you land on a bare server dashboard
8:39instead, that's a glitch that happens
8:41sometimes, and I'll show you how to fix
8:43that in a second. Here, you can keep the
8:45username as Hermes or change the admin
8:48username to whatever you want. Just make
8:49sure you remember it. And then [snorts]
8:51go ahead and click the I to toggle on to
8:53see the admin password. Go ahead and
8:56copy this and then save it somewhere
8:58safe like a password manager. Don't save
9:01this in a text notepad or anywhere
9:03insecure that somebody else could find
9:04it. This is how you're going to access
9:06your agent and you want to make sure
9:07it's locked down. Once you've got that
9:09saved, go ahead and click deploy. You'll
9:12see this loading page for a few moments
9:14and then you'll be routed into
9:16Hostinger's dashboard. You might be
9:18presented with this survey. You can fill
9:20it out or click skip. And here you can
9:22see at the bottom at the right that the
9:24application is still being deployed. If
9:27you see this, wait a few minutes until
9:28this Hermes agent application changes to
9:32a green check mark status and says
9:34running here. You can see I got the
9:37application deployed confirmation. So we
9:39are all ready to go. Now if you were
9:41routed into this page without ever going
9:44through that onboarding, let me show you
9:46how to fix that really quick. On the
9:48left side menu, find the VPS setting. It
9:51looks like this. It's a little cloud.
9:53Then from this menu, see if you have a
9:55VPS set up yet. If not, click the get
9:59VPS button. If you do, go ahead and
10:02click the manage button on your VPS.
10:04From here, you won't likely have your
10:06Hermes agent set up because you need to
10:07go through that onboarding in order for
10:09this to appear. So, here you'll click on
10:11the Docker Manager on the left side,
10:13then click catalog.
10:15In the search bar, search for Hermes.
10:18Then find Hermes agent and click select.
10:22That'll open the onboarding form that we
10:23saw earlier and take you through the
10:25whole process.
10:27Once you've got it all set up, you can
10:29manage your Docker application that
10:31holds Hermes here in the application
10:33section of Docker Manager. We have two
10:35things installed here right now. The
10:36actual Hermes agent and traffic, which
10:39secures your connection to the server.
10:41Setting up Hermes used to mean a
10:43terminal and a long list of typed
10:45commands. On this route, you never touch
10:48a terminal. Your agent is already
10:49reachable through a browser and you log
10:51in just like any other website. On the
10:53Docker Manager page, go to your Hermes
10:55agent and under the access column, click
10:57the open link. This is the login page
11:00for your Hermes dashboard. Go ahead and
11:03save this URL in your bookmark so you
11:05can access it later. Once you've done
11:07that, then go ahead and sign in with the
11:09username and password that you saved
11:11earlier. Welcome to your agent's control
11:13room. Let me give you a quick
11:14orientation before we change anything.
11:17On the left side, we have the nav menu.
11:19Sessions is every conversation your
11:21agent has ever had on any device, and
11:23they all share one memory. Models is
11:26where you pick the AI that does the
11:27thinking, and that's the first thing we
11:29fix because right now, nothing is
11:30chosen. Craw is the schedule, the page
11:34that lets it work while you're asleep.
11:36Skills is its library of procedures. The
11:38ones it ships with, the ones you
11:40install, and the ones it writes for
11:42itself. And further down, you'll see
11:44profiles. These are extra agents you can
11:46add later. And if you scroll down
11:48further, you'll see in the plug-in
11:49section, conbon. That's the project
11:52board, which is level six. We'll visit
11:54most of these rooms today. In the top of
11:56the nav section, if you click chat,
11:58it'll load a retro terminal style chat.
12:01And here if you try to type something
12:02like in my case I'll say hello you
12:06notice that nothing really happens. In
12:09fact we get this error that says no
12:11inference provider configured. On the
12:13upper right too you'll see there's an
12:15error. So your agent is installed and
12:16running but it can't think yet. It needs
12:19a model provider which is the account
12:20that supplies its intelligence. And the
12:23gateway on the left side is the part
12:25that connects it to the outside world.
12:27It carries messages to and from your
12:29phone later. Right now, it has nothing
12:31to connect to. We'll fix the model now
12:34and the gateway in level two. A model is
12:37the AI that does the thinking, the brain
12:39your agent borrows, and it's the same
12:41kind of AI that runs chat GBT and
12:43Claude. Hermes doesn't come with one, so
12:46you give it an account that supplies
12:47one. We're using now portal made by the
12:50same people who make Hermes. Now, Portal
12:53is built into Hermes, so you simply
12:55connect it by logging in with no access
12:56codes to copy anywhere. One account gets
12:59you access to more than 200 models from
13:02all the major AI companies, and it
13:03supplies the tools your agent works
13:05with: web search, reading, web pages,
13:08image generation, browsing, and speech.
13:10$20 a month buys you $22 of credits, and
13:14whatever you don't spend rolls over to
13:15the next month. Sign up with a second
13:17link in the description below. That link
13:19takes $15 off your first paid month. So,
13:22month one only costs you $5 and you
13:25still get $22 of credits which covers
13:28everything in this video and weeks after
13:30it. That link will take you to this page
13:32here where you can see the invitation
13:34banner and you'll also see the deal I
13:37just mentioned. Go ahead and click the
13:39claim your invite link at the bottom.
13:42You'll get this pop-up saying welcome to
13:43Nouse portal. Go ahead and create an
13:45account with Google, your email, GitHub,
13:48Discord, or one of the other options.
13:50I'll use my email again.
13:54You may need to enter a six-digit code
13:56from your email. Go ahead and open your
13:58inbox and grab that and paste it in.
14:00You'll get a success message and then
14:02you'll be routed to the checkout page
14:04and you can see that the discount has
14:05already been applied to your order. Go
14:07ahead and check out. I'll do this off
14:09screen and meet you on the next page.
14:12After checkout, you'll land on this
14:13confirmation page and you'll see on the
14:14left side that your account has been
14:16loaded with $22 of credits. Now, before
14:19we leave now portal, I want to show you
14:20where you'll manage your billing in the
14:22future. On the left side, under the
14:24account section, click the billing page.
14:26Here's where you can go to see your
14:27current subscription. And if you scroll
14:28down, you'll see there are different
14:30tiers. You can change between
14:32subscription tiers at any time. There is
14:34a free tier on this page, too. But it
14:36runs free models only on standard rate
14:39limits. A single agent job makes dozens
14:42of calls, so you'll hit those limits
14:44pretty fast, which is why I don't
14:46recommend using that free plan. If you
14:48ever run out of credits, simply scroll
14:49down this page to the topup section. And
14:52instead of having to upgrade your entire
14:54subscription, you can add whatever
14:56dollar amount of credits you want as a
14:58topup at any given time. All right,
15:00let's head back to our Hermes dashboard.
15:02I'll close out of my email and the Nouse
15:04portal tab. Back in the Hermes browser
15:06dashboard, on the left side nav, scroll
15:08down and click into the keys page. Here,
15:11the first option is NUS portal. Go ahead
15:14and click the login button to the right
15:16here. here. This will open a window and
15:18it'll say allow Hermes agent to use your
15:20plus plan. You can click the connect
15:22button to activate it. You'll get a
15:24confirmation saying it's connected and
15:25you can close the page and return to the
15:27Hermes agent. So, we'll go ahead and
15:28close this tab and you can see back on
15:31the keys page that Nouse portal is
15:32already connected. Now, let's pick a
15:34model on the left side. Scroll back up
15:37and select the models page. Here,
15:39[snorts] click change next to main
15:41model. And here we get a whole list of
15:43models offered by NUS portal. I
15:45recommend start with a fast and cheap
15:47model because you'll send a lot of
15:48messages while you're learning. Deepseek
15:51flash is perfect for this. Go ahead and
15:53search for DeepSeek or scroll down and
15:55find it in the list. Here you'll see a
15:58few options. There's Deepseek V4 Pro,
16:01Deepseek V4 Pro0813, and Deepseek V4
16:04Flash0731.
16:07These are the models at the time I'm
16:08recording. We want the flash model. And
16:11I recommend just going with the newest
16:13version of whatever it is by the time
16:14you're watching this. It could be
16:16Deepseek V4.1 or even 5 6 7. Don't pay
16:20too much attention to the number. The
16:22one we want is DeepSeek Flash. Now, the
16:24reason we're using DeepSeek is because
16:26it's good at using tools and a message
16:29with it costs a fraction of a scent.
16:31Later in this video, I'll show you when
16:32to move up to a stronger model and how
16:35to check what you're spending. So, we'll
16:36go ahead and click that and then click
16:38switch at the bottom. It gives
16:40[clears throat] you a dialogue asking if
16:41it's okay to switch the model. We'll
16:42click reload. Let's go ahead back to our
16:45chat page. We'll click that in the upper
16:47left. And then here in the upper right,
16:49you can see that we now have DeepSk
16:51selected as our model. Let's go ahead
16:54and type a message in the chat here.
16:55I'll say hello and hit return on my
16:59keyboard to send it. And we get a
17:01response. It says hello. What can I help
17:03you with? Two minutes ago, this agent
17:05couldn't answer anything at all. Now, we
17:08can ask it anything. So, let's ask what
17:10tools it came with. I'll say, "What
17:12tools do you have,
17:16and it comes back with a list?" It has a
17:18built-in browser. It can run terminal
17:20commands, work with files, do code
17:22execution, search the web, delegate to
17:25sub agents, use memory, schedule things,
17:28and a whole bunch of other miscellaneous
17:30tools. Go ahead and read through that
17:32list once because the rest of this video
17:34is us switching those tools on one at a
17:36time. Now, if you already pay for
17:38another AI provider, there are two other
17:40ways in. You can buy credit as you go
17:42from a service like Open Router, or you
17:45can sign in with your chat GPT
17:47subscription, though an agent burns
17:50through a plan's limits pretty fast.
17:51There are ways to point Hermes at a
17:53claude subscription, but Anthropic's
17:55terms keep those plans to its own apps,
17:57so I'm not walking you through that.
17:58Paying per message gets you every clawed
18:01model in here anyway on Nouse portal.
18:03So, at least for today, it's not
18:04necessary. To add a provider later, you
18:06can always go back to that keys page or
18:08ask your agent and it'll walk you
18:09through it. That's level one done. Your
18:12agent has a home and a model. Here's
18:15your level up task. Pause the video,
18:18send your agent three questions about
18:19your own topic, and see what it can
18:21already do with no setup at all. When
18:24you're back, we give it a phone. Level
18:26two is reaching your agent from
18:28anywhere. Right now, it lives in a
18:30browser tab, but we're going to give it
18:31its own Telegram account so it can reach
18:33you when you're nowhere near this
18:35dashboard. We're using Telegram because
18:38it's the easiest one to set up. It just
18:40takes one quick scan of a QR code with
18:42no codes to copy anywhere. And it's a
18:44great messaging app to use day-to-day.
18:46You get fast replies, and you can send
18:48files, voice notes, and photos all in
18:50one thread. Hermes supports more than 20
18:53other platforms, too, including Discord,
18:55Slack, and WhatsApp. And you can add one
18:57of those later from the same page.
18:59Before following along, get your phone
19:01out now. Install Telegram if you don't
19:03have it, and keep it nearby. I put a
19:05link in the description below that takes
19:06you straight to the Telegram download
19:08for your phone or desktop so you don't
19:09have to go looking for it. Back in your
19:11Hermes web dashboard, on the left side,
19:14click channels in the menu. Then find
19:16the Telegram channel. Here you can see
19:18it's disabled. Go ahead and click the
19:20toggle at the upper right to turn it on.
19:22You'll need to click the restart button
19:24for that change to take effect. So, go
19:26ahead and click that. And you'll get a
19:27gateway restarting popup. You'll get a
19:29message saying the gateway is not
19:31running. And then this will take a few
19:32minutes for the gateway to turn back on.
19:34If you don't see it turned on after a
19:36few minutes, I recommend going ahead and
19:37clicking refresh in your browser. And
19:39here you can see now that the gateway is
19:42running, but there's no bot token
19:44configured for Telegram and it is turned
19:46on. So, let's go ahead and click create
19:48with QR to connect for Telegram bot. All
19:51right. Now, grab your phone and get your
19:52camera app up. Then, go ahead and scan
19:54the QR code and open the link. Here
19:57you'll be taken to a conversation with
19:59the now hosted Hermes bot. Click create
20:01bot and then create.
20:05That'll take you into a conversation
20:07with your Hermes agent. But the bot's
20:10not fully connected yet. Let's go ahead
20:11and hit back on Telegram and select that
20:13now hosted HermesBot chat from the list.
20:16Or if it's not there, like it's not in
20:18my case, go ahead and scan that QR code
20:20again to pull it up.
20:23And that'll take you back to the
20:24conversation with the now hosted Hermes
20:26bot. And here you can see that there is
20:28a message that says connect this bot to
20:31your Hermes setup. Go ahead and click
20:33the connect this bot button.
20:36And then you'll see at the same moment
20:37that you get a bot connect notification
20:39in Telegram, your bot appears in the web
20:42dashboard as well. Here's the bot name.
20:44And then here's the allowed user. This
20:46is your Telegram ID. To lock this in, go
20:48ahead and click save and restart. And
20:50this can take a few minutes just like
20:51that first restart change. So give it a
20:54few minutes to restart. And if after a
20:55while it doesn't flip back to showing it
20:57that everything is running, go ahead and
20:59click refresh in your browser. And then
21:00you'll see the Telegram is connected and
21:02your gateway is running. Now, in your
21:04Telegram app, you can go ahead and tap
21:06on the username for your bot. That's the
21:08Hermes and series of numbers and
21:10letters. That takes you back into the
21:12conversation with your bot. So, go ahead
21:14and click start. And the start command
21:16doesn't actually start the conversation.
21:18You have to send a message. So, let's
21:19test that out. We'll say, "Hello,
21:23can you hear me?"
21:28And you can see that the Hermes agent is
21:29typing a message. You also get a message
21:32saying that there's no home channel set
21:33for Telegram. The home channel is where
21:35Hermes delivers the cron job result and
21:38crossplatform messages. So go ahead and
21:40simply tap the / set home command and
21:43then that sets telegram to be your home
21:45channel. Now anytime a cron job runs and
21:47it needs to send you a notification,
21:49telegram is where it will send it. And
21:51you can see the message that came
21:52through here. It says, "Yes, I can hear
21:53you loud and clear. What can I help you
21:54with, Matt?" So now Telegram is
21:56connected and we can continue chatting
21:58with our agent on our phone or even on
22:00the Telegram app on the desktop. Let's
22:03cover safety. This flow just put your
22:05Telegram account on the allow list. So
22:07if anyone else finds your bot and tries
22:09to message it, they get nothing back.
22:11And Telegram bot chats aren't endtoend
22:13encrypted. So don't send it passwords or
22:15banking details. For everything we do
22:17today though, it's fine. Your
22:19conversation on the phone and your
22:20conversation in the dashboard are two
22:23separate threads. A lot of people find
22:24this confusing at first. So, think of it
22:27like email. You can have five email
22:29threads going with the same assistant.
22:31What you said in one thread doesn't
22:33automatically appear in another. But the
22:35assistant has all five threads in their
22:37inbox and can look any of them up. Your
22:39agent works the same way. It has one
22:42memory but many separate threads. So, if
22:44you start something in the app and you
22:46want to continue on your phone, hand the
22:48thread over with the handoff command. If
22:50that ever gives you trouble, you can
22:51just ask it on Telegram to continue what
22:53you started and it'll search its own
22:55history. Let me quickly show you how
22:57that handoff works. So, go to your chat
23:00and let's start a conversation. Let's go
23:02ahead and ask it something random like,
23:04"Give me a synopsis of the movie The
23:06Godfather."
23:08Now, we've got this answer and let's say
23:09we wanted to hand off this conversation
23:11and continue on Telegram. You can type
23:13forward slash and then handoff. Add a
23:16space. Then, you'll see a list here. Go
23:18ahead and toggle down to Telegram and
23:20hit enter on your keyboard. Then again,
23:23and now you can see here on Telegram, it
23:25says working here now. Same session.
23:27What we've got so far, you asked for a
23:29synopsis of the Godfather. And now I
23:31could continue that conversation here.
23:33And if I flip over to the desktop
23:34version of Telegram, you can see that
23:36I've got the same thread here, too. Now,
23:39there are other slash commands you can
23:40send which are especially useful from
23:42Telegram. If you type slash new and then
23:46hit enter, that starts a brand new
23:47session and it asks for permission to do
23:50this. So, we'll go ahead and click
23:51approve once.
23:54And now we're on a brand new
23:55conversation. By default, if you just
23:57keep going, all of your conversations on
23:59Telegram will end up in the same thread,
24:02which means your model is getting a lot
24:04of context every time you send a
24:06message. So, a general rule is if you
24:09start a new topic, go ahead and start a
24:11new session. That'll save you from
24:13burning through tokens, and we'll cover
24:14more about that later. On Telegram, you
24:16can also just type slash and scroll
24:18through and see all the different kinds
24:19of commands you can send. You don't have
24:21to learn all these now, but just know
24:22that these are here and available for
24:24you to prompt your agent to do various
24:26tasks. These commands that start with
24:28slash work on every door, the app, the
24:31dashboard, and Telegram. The other one
24:34that's super useful is slashmodel.
24:36Typing that allows you to switch between
24:38any of the models that are available to
24:40you. For example, I could click into the
24:42NUS portal and see all of the various
24:44models that are offered through there,
24:46but I'll go ahead and click cancel
24:47because I'm going to stick with Deep
24:48Seek for now. Now, if your bot ever goes
24:50quiet, there are three things to check
24:53in this order. First, check your credit
24:56and your rate limits with your model
24:58provider because that's the usual cause.
25:00And Hermes is good about telling you
25:02when it happens. So, look at your last
25:03message. Second, the gateway. Here you
25:07can see my gateway status is running,
25:09but check this if you're having any
25:10issues and click restart if you see
25:13anything strange. That can help to fix a
25:14silent bot almost every time. And then
25:17third, the logs. You can find the logs
25:19in the left sidebar near the top of the
25:21list. Here are all the different logs
25:23and you can filter through by type. All
25:26debug related logs, info, warning, and
25:29errors. Logs is your agent talking about
25:32itself. You don't need to read every
25:33line, but just filter for errors when
25:35something seems like it's going wrong.
25:37And if none of these three steps fix it,
25:40restart the whole application from
25:41Hostinger's Docker Manager. Click the
25:44three dots next to your application and
25:46then restart.
25:48Okay, so now you can access Hermes
25:50through your browser. You can talk to it
25:52on Telegram, but now let's get it set up
25:54with the official desktop app on your
25:56computer. The browser dashboard is the
25:59control room and it works from anywhere,
26:01but it's a tab you have to go find. For
26:03daily use, you want the desktop app. It
26:06looks and feels like the chat GPT or
26:08cloud desktop apps and it keeps your
26:09conversations in a sidebar so you can
26:11drag files into it and it can listen and
26:13talk back. We'll install it and instead
26:15of letting it run a second agent on this
26:17computer, we'll point it at the server
26:19we set up so it uses the same model and
26:21same memory inside a native app. I'm on
26:23a Mac, but on a Windows, the flow is the
26:25same with control on the keyboard where
26:27I press command. To download the
26:29official desktop app, go to this link
26:31here, and I'll also go ahead and put it
26:33in the description. It's
26:34her-agent.nouse.com
26:38and then look for the download button
26:39for your operating system on that page.
26:41I'll go ahead and click download for Mac
26:43OS. Go ahead and save that installer
26:46somewhere. And then when it's done
26:48downloading, go ahead and run that
26:49installer.
26:51Right now, it's quite simple. On Mac,
26:52simply drag the Hermes icon into the
26:54applications folder.
26:56Next, open your apps, then open and run
27:00Hermes. It might ask for permission. On
27:02Mac, it says Hermes is an app downloaded
27:03from the internet. Are you sure you want
27:04to open it? And go ahead and click open.
27:06Yes, it's fine. And then you'll get a
27:08popup where you're prompted to install
27:09Hermes. Go ahead and click the install
27:11Hermes button. And then it'll run
27:14through a list of setup steps
27:15automatically. This can take a few
27:17minutes, so I'll fast forward to when
27:18it's done. Then once the installation is
27:20done, go ahead and click to launch
27:22Hermes. You'll eventually see something
27:24like this where it says, "Let's get you
27:26set up with Hermes agent." And it
27:27prompts you to connect to a model
27:30provider. Skip this provider screen
27:32because picking one would create a
27:34second agent on this computer and we
27:36already have one set up on the server.
27:37So just click I'll choose a provider
27:39later. Welcome to the Hermes desktop
27:42app. Before we go any further, let's
27:44connect our server. Go to the upper
27:46right and click the gear icon. or on Mac
27:48you can hit command comma on your
27:51keyboard. On Windows it would be control
27:53comma. Here on the left side find the
27:56gateways menu item and select it. Then
27:59select remote gateway in the connection
28:01mode. Now the field we need to enter
28:03some information in is the remote URL.
28:05And all we really have to do here is put
28:08the link where our server is. To find
28:10this is super easy. Go back to your
28:12Hermes agent dashboard in the browser
28:14here. Grab the link that's in your
28:16address bar. everything that comes
28:18before the forward slash. Here I have
28:20forward/ chat. So go ahead and just grab
28:23everything up to that. Copy that. Head
28:26back to the Hermes desktop app and then
28:28paste in that URL. Here you can see it
28:31says for authentication, this gateway
28:32uses a username and password. Sign in to
28:35authorize this desktop app. So go ahead
28:37and click the signin button and that'll
28:39give you this familiar signin interface.
28:41Go ahead and enter your username and
28:43password and then sign in. You should
28:46see that it says you're signed in. Go
28:48ahead and click the save and reconnect
28:50button. That'll restart the gateway
28:52connection. And now we should be able to
28:54start a conversation with our agent. So
28:56go ahead and close out the settings
28:57page. Head to the chat and in the text
29:00field say something like, "What was my
29:03first ever message to you on Telegram?"
29:05and then send it. Now you might see this
29:07error that says backend out of date.
29:09That pops up if the app is newer than
29:11the server. You can dismiss it or click
29:14update Hermes. I'll go ahead and click
29:16the X to dismiss it for now. I'll update
29:18later. Here you can see the agent is now
29:21thinking. And now it comes back with a
29:23response saying your first message to me
29:25on Telegram was hello, can you hear me?
29:27So you now have one agent with three
29:29doors. The web browser dashboard,
29:32Telegram, and the desktop app. Every
29:35door shares the same memory. Now, one
29:38thing that I think is really cool about
29:40using the desktop app is the heads up
29:42display mode or HUD. To access that, hit
29:45commandshifthh on your keyboard. And
29:48that'll open this little floating bar
29:50that sits on top of whatever you're
29:52working on. And its position tells
29:54Hermes what's under it. So, if you say
29:56something like this or here, that means
29:59the page or app you're looking at. Now,
30:01I don't use it all day, but it saves a
30:03window switch when you want a quick
30:05answer for something. And you can drag
30:06it around as big as you want or anywhere
30:08on your screen. It's super handy. You
30:10can exit this mode by clicking this exit
30:13HUD mode button on the right or use the
30:15same command command shift H to put it
30:17back to the main desktop app. Now, here
30:19are some other buttons that are useful
30:21on the desktop app. For example, here
30:23this microphone button. You can click it
30:25to enable voice dictation. For example,
30:27I can click it and it might ask for
30:29permission to your microphone. You can
30:31go ahead and click allow. And now you
30:32can see that it's dictating. Um, hello
30:35Hermes. I am currently recording a test
30:37dictation message. Can you hear me?
30:42Then you click the stop dictation
30:44button. It'll quickly transcribe it and
30:46turn it into a message that you can
30:47send. I'll go ahead and command A to
30:50select all that and delete it though cuz
30:51I don't need to send that message. If
30:53you want your agent to talk to you, you
30:55can turn on this speaker button and that
30:58will have it read its replies out loud
31:00to you. You can also click this wakeword
31:03button here. That's where you can just
31:05say, "Hey, Hermes," into your microphone
31:07at any time to activate Hermes to listen
31:10to you. Finally, this little waveform
31:12icon here at the end starts a voice
31:14conversation, so you and Hermes can just
31:17talk back and forth with voice. This is
31:19quite handy if you just need some quick
31:21answers for something or don't want to
31:22spend a bunch of time typing into your
31:24keyboard. That's level two. Your agent
31:26has a chat on your phone, a desktop, and
31:29a voice. Here's your level up task.
31:32Pause here. Message your agent from your
31:34phone while you're away from the
31:35computer and then pick the same thread
31:37up in the app. Once you've done that,
31:39level three is where it starts doing
31:41work for you.
31:43Level three is putting it to work.
31:45Chatbt, Claude, and Gemini can all
31:47search the web, read your files, and
31:49connect to your accounts. So, what
31:51exactly makes Hermes different? Two main
31:53things. The first is where the work
31:56happens. Your agent has its own
31:58computer, so whatever it makes stays
32:00there. files you can pick up again next
32:02week and somewhere for the scheduled
32:03jobs in the next level to write to. The
32:06second is that its whole ecosystem is
32:09built around doing work autonomously and
32:11getting better with each run all on its
32:13own. So, we'll run a quick research task
32:15to see the tools working and then
32:17connect a real account you'll actually
32:19use. In the desktop app, let's initiate
32:22a research task. I'm looking for a
32:25standing desk right now. So, I'll go
32:27ahead and say, "Research the top three
32:29standing desks. Give me a one paragraph
32:32comparison with links." You watch it
32:34load some skills and start to use tools
32:37to actually read the web. Here you can
32:38see it's running a tool saying it's
32:40reading New York Times, for example.
32:42It's running code in the background,
32:44reading pages, and putting together its
32:46recommendations. I'll go ahead and fast
32:48forward to when it's done. It searched,
32:50opened the pages, read them, and came
32:52back with an answer instead of a list of
32:54links. Although it did provide me with
32:56some citations down here, which is super
32:58handy. Now, let's do something it should
33:00stop and ask about first. I'm going to
33:02ask it to create some test files and
33:04then delete them. So, here I'll say this
33:07create three test files. And here you
33:10can see it wrote into a folder on the
33:12server, the opta
33:14folder. It created three different
33:17files. One text file, one markdown file,
33:19and one JSON file. So, now let's go
33:22ahead and have it delete them. I'll say
33:24delete those three test files you just
33:25made. Now, because I gave it explicit
33:29permission to delete those files, it
33:31went ahead and did it. But if it was
33:33working on something in the background
33:35and it decided it needed to delete those
33:37files as part of the process, it would
33:39ask me permission first. You might see a
33:42little something in the desktop app that
33:44says run or if you're on Telegram, it'll
33:47give you a little popup similar to that
33:48model selector popup and you can improve
33:51that task for the session or just once.
33:54Simple safe tasks like research don't
33:57trigger any sort of permissions prompt.
33:59But deleting things can and so does
34:01running a risky code command. And
34:04nothing happens until you say yes. Now,
34:07the way I'd think about permissions is
34:09progressive trust. Don't hand it
34:10everything on day one. Give it a little
34:13watch how it works and give it more only
34:15when you're comfortable. In practice,
34:17that means a separate account for
34:19anything it logs into. And your main
34:21email and anything with money in it
34:23stays off the list until it has earned
34:25them. Now, let's look at files. In the
34:28app, you can drag anything into the
34:30chat, a PDF, a spreadsheet, or a
34:32screenshot. And on Telegram, you can
34:34send it a document or photo. Here in the
34:36desktop app, I'll drag in this AI index
34:38report PDF and I'll tell it summarize
34:42this and pull out the three numbers I
34:44should care about. It gets to work and
34:46I'll fast forward to when it's done. All
34:48right, it came back with an answer. It
34:50gave me three numbers to focus on these
34:52three here in the list. Now, let's tell
34:55it something like, turn that into a
34:57short document I can send to a
34:58colleague. And after a minute, it comes
35:01back with this AI index brief file as a
35:04markdown file. And it gives me the
35:07option to open a preview or download it.
35:09So I'll go ahead and download it and
35:10then put it in my downloads folder. An
35:12agent that lives on a server has one
35:14catch. The files it makes live on that
35:17server, too. And files you send through
35:19the chat, like Telegram, have a size
35:21limit. The desktop app allows for
35:23greater throughput with the file sizes.
35:26But right now, our agent and us don't
35:28really have a shared file system. So,
35:32we're going to give you and your agent a
35:34shared folder in the cloud. I'm using
35:36Google Drive because many of you already
35:38have it. And the same setup gives your
35:39agent your calendar and your
35:41spreadsheets, which we'll use in level
35:43five and again in level six. For this, I
35:46use a separate Google account for the
35:48agent, not my main one. And you should
35:50too. Make one now if you need to. It
35:52takes 2 minutes. You can then share
35:54individual folders from your real
35:55account into it so it only ever sees
35:57what you choose to give it. In a new
36:00session, go ahead and say, "Set up the
36:02Google Workspace skills so you can use
36:04my Google Drive, Calendar, and Sheets."
36:07Now, I'm going to show you something.
36:08This doesn't actually go the way I
36:10expected it to here in the recording.
36:12So, normally there's a tool called
36:14Clarify that Hermes can call if it has
36:16additional questions for you. Now, I
36:18don't really know why it's doing this
36:19right now in my particular setup, and
36:21I'm not going to gloss over the fact
36:22that sometimes with AI, there's just
36:24bugs. That's just what happens. It's not
36:26a perfect tool, but it's constantly
36:28improving. So, what I did here was I
36:30took a screenshot of the fact that the
36:33form would not load its questions. And
36:36then I sent a message to Hermes with a
36:38picture of that and said, I can't see
36:41the questions. They're just loading
36:42forever. And I went ahead and pushed
36:45that through the queue manually. If you
36:48ever have issues with your Hermes agent,
36:49you can simply just take a screenshot of
36:51what's going on and send that as a
36:52message to Hermes and it can help you
36:54resolve its own issues. So, that's
36:55pretty cool. And the nice thing here is
36:57it came back asking me the questions
36:59outside the special tool. It just asked
37:00them in plain text, which works just
37:02fine for us. For you, clarify may work
37:05or it might not. Either way, now you
37:07know how you can handle it if it
37:08doesn't. So, here it's asking to set the
37:10scope for the connection. It says you
37:13asked for drive, calendar, and sheets.
37:14It'll also include Docs for us. and say
37:17the word if you'd rather also have Gmail
37:19and contacts too or keep it to just
37:20three. For me personally, I'll go ahead
37:23and add those so I can demo how that
37:25works. But Gmail is a sensitive subject
37:27for some people. So only include it if
37:29you are comfortable connecting it. Here
37:31it also asks me advanced protection.
37:33Does your account use hardware security
37:34keys? Mine does not. So I'll tell it
37:36that. So it just asks for me to answer
37:38that plainly in a text sentence. So I'll
37:41just say something like regarding the
37:43scope, I'd like to use all those tools.
37:45My account does not include any advanced
37:47protection such as hardware or security
37:49keys. So I'll go ahead and send that and
37:51let it think for a second. So now we
37:54have some setup steps to do. First, we
37:56need to create a Google Cloud Platform
37:59project. It's provided me a link here.
38:01So I'll click this. And here you can let
38:03Hermes browse with a snapshot of your
38:06default browser profile. So essentially
38:08you can pull in cookies from your
38:10existing live browser. So, let's go
38:12ahead and click use my profile if you
38:14want to carry over those credentials.
38:15And this might open for you in a native
38:18app browser inside of Hermes. Or if
38:20you're interacting with it in something
38:22like Telegram, it might open a different
38:24browser. Personally, I'd actually rather
38:26work with this in my standard browser.
38:28So, I'll go ahead and copy this link out
38:30and go back to my Brave browser here,
38:33open a new tab, and paste in that link.
38:35And here, it logged me right in. I
38:37already have another project in Google
38:38Cloud, but you might not. So, I'm going
38:40to go ahead and click create project,
38:41which is what you'll do. Let's call the
38:44project Hermes Agent and then click
38:46create. After a second, it'll load your
38:49new project. You can see that up here in
38:51the top bar, it says Google Cloud and
38:52next to it says Hermes Agent. Next, we
38:55want to enable APIs and services. And
38:57here, you can see it's already popped
38:58out for me. But if you don't see this,
39:00go ahead and go to the APIs and services
39:03nav item and then click enabled APIs and
39:05services. I'm going to go ahead and
39:07close out these pop-ups. All right. Now,
39:09there's a button up top that says enable
39:11APIs and services. Click that. Now, we
39:14need to go through and find the tools
39:16inside of Google Workspace that we want
39:17to connect. So, let's go ahead and start
39:19with what we told it. We'll start with
39:21Drive. Search for Drive. And you might
39:23get some automatic suggestions here, but
39:25we'll just go ahead and search for Drive
39:27as it is.
39:29And this one here is the one we want,
39:30Google Drive API. We'll select that and
39:32then click enable. After a moment,
39:35you'll be taken to a page like this that
39:37shows the Google Drive API. And up at
39:39the top, you'll see something that says
39:41disable API. That means it's been
39:42enabled. If you were to click this
39:44button, it would stop it. But we don't
39:45want that now. This is just a quick way
39:47to check if you are set up. So now we'll
39:50go ahead and click the back arrow. And
39:53we run through this process again for
39:55every Google Workspace tool we want to
39:57connect. And I'll admit this is the most
39:59tedious part of the entire setup. But
40:00the nice thing is once you've set this
40:02up, you never have to do it again. So
40:04I'll just go ahead and do this quickly
40:05for all the tools that I want to
40:07connect.
40:08We'll use Gmail API enable.
40:13Google Calendar API enable
40:19Google Docs API.
40:21Enable
40:25Google Sheets API.
40:27Enable.
40:32And then Google People API. Enable.
40:35Google People API is the more modern
40:37version of contacts.
40:40All right. Now that we've connected the
40:41APIs, let's go back and double check
40:43what Hermes needs from us next. I'll go
40:46ahead and close out this browser here.
40:48You can do that by clicking the little
40:49X, the tab. There we go.
40:52Next, after enabling the APIs, it wants
40:55us to create some credentials. You can
40:57click the link here or I'll show you
40:58where to get to it in the Google Cloud
41:00Platform. Here on the left side of the
41:01APIs and services page, you'll click
41:03credentials, then click create
41:06credentials at the top, then OOTH client
41:08ID. And then if you see this yellow
41:10banner saying to create an OOTH client
41:12ID, you must configure your consent
41:14screen first. Go ahead and follow that
41:16process, too. So, we'll go ahead and
41:17click configure consent screen. We'll
41:19click get started and we'll enter some
41:21information about our app. Our app in
41:23this case is whatever we want to name
41:24our current project here. So we can just
41:26call that Hermes.
41:29User support email. We'll just use our
41:31own email from the drop down. Then click
41:33next. For audience internal versus
41:36external, go ahead and select external.
41:39Then click next.
41:41Then for contact information just enter
41:43your email address again. Then click
41:44next. and then click the check box to
41:47agree to the terms of service and click
41:49continue. Then click create. Okay, now
41:52we can go back to the APIs and services
41:56and then credentials.
41:58Now click create credentials up top and
42:00OOTH client ID. From the application
42:03type dropdown, select desktop app and
42:06then give it a name. Again, we can just
42:08call it Hermes. Click create. And then
42:12here you have some information that
42:14you're not going to want to share with
42:15anyone. the client ID, the client
42:16secret. This is all stuff you want to
42:18keep private. Don't share this with
42:19anyone. Download the JSON file.
42:22Put it somewhere you'll find it. In my
42:23case, I'll put it in downloads. Now,
42:26drag that new JSON file it gave you into
42:28the Hermes chat
42:31per the instructions here. Now, there's
42:32one more thing we need to do here. We
42:34need to add ourselves as a test user. If
42:36the app is in testing here, it gave me a
42:38link. So, I'll click that. I'll go ahead
42:41and copy this URL again and close out
42:43the browser. Open it in the main browser
42:44here. And here at the bottom of this
42:46audience page, find the test user
42:49section, click add users, and then add
42:52yourself to the test users list. Just
42:53enter your email, hit enter to lock it
42:56in, and then hit save.
42:59Okay, now that we've followed all of
43:00Hermes instructions, we can go back to
43:02Hermes and then just hit enter to send
43:04that JSON file. Okay, now it gives us an
43:07authorization URL. I'm going to go ahead
43:09and rightclick this link it gives me and
43:11then click open in external browser. Now
43:15I'll log in with the account that I plan
43:16to use.
43:18It'll say Google hasn't verified this
43:20app. That's okay because you're
43:22technically the developer of this app
43:23that we're setting up. So go ahead and
43:25click continue.
43:28It'll ask for the scope of things that
43:30it's allowed to do inside of Google. So
43:32go ahead and select all. These are the
43:34exact things that Hermes has asked
43:37permission for, and you can read through
43:38them and change the scope if you want
43:40to, but these are all the things that we
43:41specifically asked it to do. So, this is
43:43just fine. We'll go ahead and click
43:45continue. And you'll get what looks like
43:47an error. This is actually the expected
43:49behavior here. So, go ahead and copy
43:52this URL from your browser address bar,
43:55return to Hermes, and then paste in that
43:58full URL here in the message field and
44:00send it.
44:02This can take a few minutes to process.
44:04When it's done though, you should see
44:06something that says that the setup is
44:07complete and it'll confirm that it's got
44:10everything connected. Now, if you didn't
44:12get to this point successfully, the
44:13easiest thing to do to troubleshoot is
44:15simply take a picture of whatever it is
44:17that you're encountering on screen
44:19that's giving you trouble and send that
44:21screenshot to Hermes to help you
44:23troubleshoot. Just explain what you're
44:25seeing. Explain what kind of errors are
44:27popping up. And this process is pretty
44:30straightforward for an agent like Hermes
44:32to help walk you through. And if for
44:34some reason Hermes itself isn't
44:35responding or isn't actually doing a
44:37great job at troubleshooting, you could
44:39always go to another chatbot somewhere
44:41online like chat GPT or Claude to help
44:44you out. Once you're confirmed though,
44:46next let's actually set up our file
44:48system in Google Drive. So let's say
44:51create a drive folder called Hermes with
44:54inbox and outbox inside and give me the
44:57link
44:58here. It just created some folders and
45:00gave me a link. So let's go ahead and
45:02I'll rightclick this Hermes folder link
45:04and click open in external browser.
45:06Close these old tabs I don't need
45:08anymore. And here you can see in my
45:11Google Drive folder here I have the
45:12inbox and outbox folder that have been
45:14created. So now Hermes has access to
45:16perform tasks inside of Google Drive for
45:19me. This is great. Let's head back to
45:20the Hermes desktop app and I'll go back
45:22to that summary conversation I had where
45:24it made a report for me about the AI
45:26index. So now that Hermes has access to
45:28my Google Drive, let's tell this summary
45:30conversation, save the summary you wrote
45:32as a document, upload it to Outbox in
45:35Google Drive, and send me a link that
45:37works for my phone.
45:39After a few minutes of processing, it
45:41says it's done and the summary is now a
45:43Google doc inside my drives outbox
45:45folder shared so that anyone with a link
45:47can view it. So, let's go ahead and open
45:49this in an external browser. And there's
45:52my document. Now, I could share this
45:54with anyone or even view it on my phone.
45:56Now, we've seen it write content into
45:57Google Drive. Let's see it retrieve some
45:59content from Google Drive. I'm going to
46:01go back to my Google Drive tab here,
46:03open the inbox, and then add a new file
46:06here. We'll click file upload. I'm going
46:09to add in this state of AI 2026
46:12document.
46:14It's uploaded. So I'll head back to
46:15Hermes and let's see how it does being a
46:17little bit ambiguous. I'll say grab the
46:19newest file in my inbox folder and give
46:21me a five bullet summary. So I didn't
46:23tell it that there was an inbox folder
46:25in Google Drive. I just said grab the
46:27newest file in the inbox. Let's see how
46:30it does.
46:34It pulled it in. and it says the newest
46:35file in your inbox folder is the state
46:37of AI 2026 and it gave a five bullet
46:40summary. So that's the shared Google
46:42Drive folder. Anything you want it to
46:44read goes in inbox. Anything it makes
46:46for you lands in Outbox with a link and
46:48it works from any device. Now one thing
46:50I want to point out here before we go on
46:51to the next subject is that it
46:53automatically did a self-improvement
46:55review. It wrote a document delivery
46:58skill, wrote a skill credential setup
47:01update and updated its memory. I didn't
47:04have to tell it to do that. It just
47:05figured out what the process was for
47:07delivering documents, in this case,
47:08sending them to Google Drive. It figured
47:11out that there were some updates it
47:12needed to make to the skill credential
47:14setup and updated it and decided to
47:16remember things. This is what makes
47:18Hermes really powerful. Now, those two
47:21folders are just a starting point. Once
47:24you see what your agent actually
47:25produces, you'll want folders that match
47:27your own work. So, go ahead and change
47:29them. You can also ask your agent to
47:31keep them tidy, either when you think of
47:33it or on a schedule, filing things where
47:35they belong so it doesn't fill up with
47:37files you can't find later. [snorts]
47:39Now, that setup also gave your agent
47:41access to Gmail. And this is a question
47:43I've seen in many comments. If it can
47:46reach my email, is it going to start
47:47just sending things? It does what you
47:50ask it to do. So, if you tell it to send
47:52an email, it'll send the email. So, the
47:54control comes from the rule you give it.
47:57Here's the rule I use on everything of
47:59mine that runs on a schedule. Write the
48:01draft, never send it, and tell me it's
48:03ready. Then I read it, change anything
48:06that's off, and press send myself.
48:08That's 10 seconds of my time, and then
48:09nothing goes out in my name that I
48:11haven't read. After you've edited a
48:13draft, ask your agent to look at what
48:14you changed and remember it so the next
48:16one comes back needing even less fixing.
48:20Hand things over in this order. Let it
48:22read and organize for a few weeks first.
48:24That means sorting, labeling, and
48:26pulling out the things you need to see.
48:28Then let it draft. Sending on its own is
48:31the last step if you ever get there. The
48:34job we automate in the next level never
48:36sends anything at all. It reads
48:37confirmations and files them, and that's
48:39it. Google is just one example of a tool
48:42you can connect. Think about the one
48:44tool you spend the most time in and
48:46connect that one next. Many tools now
48:48speak a standard called MCP, which is a
48:51plug that lets an agent read from and
48:53act inside that tool. Hermes has a whole
48:55catalog of readym made connections, and
48:57you can add others by simply pasting a
48:59link. In the Hermes desktop app, go to
49:02capabilities on the left side menu.
49:05And then from here, you'll see a couple
49:07of different tabs. There's skills,
49:09tools, MCP, and plugins. If you go to
49:11MCP, you can see the full existing
49:14catalog of pre-built MCP connections
49:17that you can simply toggle on. All you
49:18need to do is authenticate the apps. For
49:21example, if you wanted to connect your
49:23Canva, you would simply click install.
49:26Then you can scroll up when it's done
49:28installing and you can see it in the
49:29list of servers that are available to
49:30you. Simply click on it and then click
49:33authenticate and then connect your
49:34account. Doing this will enable Hermes
49:37to work in your Canva account just like
49:40we set up with the Google Workspace. I'm
49:42actually not going to connect this
49:43though personally. And you can toggle
49:44this off any time by simply hitting the
49:47toggle. Or you can delete it entirely by
49:48clicking the remove button, the trash
49:50can icon. For you, an MCP connection
49:52might be notion, GitHub, your customer
49:55database, or even your accounting
49:56software. Simply connect it, decide what
49:59it's allowed to touch, and then give it
50:00jobs inside that tool. Let me show you
50:03an example of Hermes doing a job inside
50:05the Google Workspace tool we already
50:06connected. Back in the app, let's start
50:08a new session. And then I'll say, look
50:11at my calendar for next week and add a
50:1330-minute block called change studio
50:15light bulbs on the leaked busy
50:17afternoon. Now, before I actually run
50:19this, let me show you what the calendar
50:20currently looks like so you can see how
50:21it changes. This is my calendar next
50:24week. So, I imagine it'll find this
50:27Friday slot, but I want to see what it
50:28comes back with. Let's go ahead and run
50:30that prompt.
50:31It'll think for a few minutes. Let's pop
50:33over to Google Workspace.
50:36And just like that, you can see change
50:38studio light bulbs appear to my
50:39calendar. It went into my calendar and
50:41added the block all itself. That's level
50:44three. So, here's your levelup task.
50:46Connect one tool you use every day and
50:49give your agent one real job inside it.
50:53Level four is where we stop repeating
50:55ourselves.
50:56Level four is training it. You've
50:58probably had this experience with AI.
51:00You correct it, it fixes that one reply,
51:03and then the next session, it's already
51:04forgotten and you're correcting it
51:06again. Watch what happens when you
51:07correct Hermes. As an example of
51:10something simple that I might correct
51:12Hermes on, I'll have it write a short
51:14social media post. I'll say, "Write a
51:16short post about studio lighting for
51:18complete beginners."
51:21And here you can see it comes back with
51:22headings, bullets, and it's quite wordy
51:25overall. It's not as short as I'd like.
51:27So, let's give it a correction. I'll say
51:29too wordy. I prefer short, punchy
51:32sentences, three paragraphs max, no
51:34headings or bullets. Save that as a
51:36preference.
51:38You can see it saved that to memory. It
51:40said it was saved and then came back
51:42with a much better structure. I told it
51:45plainly the way you would tell a new
51:47hireer as you're training them. And then
51:48it wrote that down in its memory. I
51:50didn't have to go find a settings page
51:52or edit anything. I just told it. So now
51:55in the same session we'll say now write
51:57one about microphones for video
51:59production. And this is just another
52:00example to show you what happens in the
52:02same thread. And there is the short
52:04post. It matches the original style. But
52:07now watch what happens when we go to a
52:10brand new session. And this is where the
52:12real magic happens. Now you might create
52:14new sessions pretty often. So you can
52:15either click new session on the left
52:17side or at any point you can start a new
52:20session by hitting commandn on your
52:22keyboard or on Windows that would be
52:24controln. So I'm going to do that. It's
52:26a good keyboard shortcut to know cuz
52:27you'll probably use it often. Now I'll
52:29say write me a short post about roasting
52:32coffee.
52:33And there's the exact same structure
52:36again. That was a brand new session with
52:38nothing carried over and it's still in
52:40the shape I asked for once in a
52:42completely different session. Any
52:44chatbot can hold a preference inside of
52:46one conversation because it has the
52:48context of that one conversation it
52:50reads through when it responds. This
52:52one, on the other hand, holds it across
52:54every conversation on every device from
52:56now on. The AI model doesn't remember
53:00anything between messages. Every time
53:02you send one, it reads a fresh briefing
53:04on who it is, what it knows about you,
53:06its notes, and the conversation so far.
53:08Picture a brilliant assistant with total
53:11amnesia who keeps excellent notes. Every
53:14morning they walk in with no idea who
53:15you are, read their notebook, and then
53:18get to work as if they'd never left. The
53:20notebook is the memory. In the Hermes
53:22desktop app, you can access the memory
53:25by going to settings with the gear icon
53:27in the upper right or command, comma. On
53:30the left side in the menu, click memory
53:32and context. And then here you can see
53:35all of the memory settings. There are
53:37two main things I want to point out.
53:38Persistent memory and user profile. If
53:42you want to see the actual memory files,
53:43you can go to the web dashboard and then
53:47go to the files section in the left side
53:50nav. Find the memories folder. Open it
53:54and then here you can see the memory.
53:56MD, the markdown file for the memory and
53:59then user.md, which is that other one,
54:02the user profile. And to see them, you
54:04can simply click download. I'll save it
54:06to my downloads folder here. I'll open
54:09it in Finder.
54:11Markdown files can be opened in lots of
54:13different software options. If you have
54:15a coding tool on your computer, it might
54:17default [clears throat] to that. If you
54:19have notion, it might default to that.
54:21But you can always rightclick and choose
54:22exactly what you want to open it with.
54:23In my case, I actually just want to open
54:25it with text edit. And here I can just
54:28see the raw text of the markdown file.
54:30So this is the full memory here. Some
54:32things that's remembered so far. for
54:34example, are that it has access to a
54:37Google Workspace account and it lists
54:39the email address and all the tools that
54:41it has inside that workspace. Every time
54:43the agent wakes up, it looks at its
54:44memory and goes, "Oh, look, I have
54:46access to Google. Cool. I can use that
54:48to do my job." Then it looks at the
54:49message and says, "What's my job?" Let
54:51me explain memory versus user in a bit
54:53more detail. There are two notebooks,
54:56like I said. One is about you. It's your
54:59name, how you like things written, and
55:00what you expect from it. That's where
55:02the correction we just made went. The
55:05other one is the agent's own working
55:07notes. What it has learned about your
55:09setup, the conventions you use, and the
55:11things it got wrong once and then fixed.
55:14You don't write either of them. Your
55:16agent does as it goes. And you can read
55:19both like I just showed you. And if you
55:20want to, you could ask it to change
55:23things specifically. Both of these are
55:25short on purpose, just a few hundred
55:27words each because each word gets sent
55:29with every message. So, the notebooks
55:32hold the things that matter most, and
55:34for everything else, your agent has two
55:36more places to look. It can search every
55:38conversation you've ever had with it,
55:40and it can read the files and skills it
55:42has saved. All of it lives on your
55:44server. I just showed you where you
55:45could go to find files. There's a whole
55:47lot more files in there. Now, here's how
55:49I run my own personal agent. When the
55:51notes fill up, the memory and user, ask
55:54it to keep an index. That's one file
55:57that lists where everything else lives.
56:00kind of like a table of contents for all
56:02your files. Then the short memory only
56:05has to say start here at the index and
56:08it looks up the rest when a job actually
56:11needs it. Let me show you how you could
56:12set up your own index. We'll send a
56:14prompt something like this. From now on,
56:16keep an index file of your notes, a
56:19table of contents that points to where
56:20each topic lives, and put a oneline
56:23pointer to it in your memory.
56:25And after a minute of processing, it
56:27says done. It created the index file and
56:31put it in our file system. It says it's
56:33a master index/t of contents covering
56:36the memory, the skills it has, the load
56:39instructions, and it added that detail
56:41to the memory. If you wanted to, you
56:44could even ask to see the index. I just
56:46said, "Show me the index file." And here
56:48it came back with a response showing
56:50everything that it included in the file.
56:52So, you can go through it and ask it to
56:54change things if you want. It gave us
56:55the explanation for where the file
56:56lives. So if you wanted to find this in
56:58your web dashboard, you'd go to the
57:00folder opt/data and then find
57:02notes_index.mmd.
57:04So here you can see we're already in the
57:06file system at /opt/data.
57:09So then all we need to do is find that
57:11index file. Scroll down and here it is
57:14notes_index.
57:16If we click, we can download it. I'll
57:18open it in finder again. Then open in
57:21text edit. And there it is. There's the
57:24index file. and it'll change this on its
57:26own as time goes on. Memory holds facts
57:29and preferences. Skills hold procedures.
57:33A skill is like a recipe, a written list
57:35of steps for how to do one job that your
57:37agent only reads when that job comes up.
57:40Let's turn what we just taught it into a
57:42skill. Back in the Hermes desktop app,
57:44I'm going to go find that postwriting
57:47chat we had earlier. Go ahead and click
57:48that. And then I'll say save that
57:51editing workflow as a skill called
57:53concise writer.
57:56And after a minute of thinking, it comes
57:58back saying it saved that as concise
58:00writer and gave the location of the
58:02file. It explains how it encoded the
58:04loop it ran on both posts and turned
58:06that into the skill. Now, if we go to
58:09the capabilities section in the left
58:10side nav and then look through our
58:13skills here, if we scroll down to where
58:16it would be alphabetically, we can see
58:18here concise writer. Any skills you
58:21asked it to learn or any skills it
58:23learned on its own, you'll see listed
58:25with the learned tag. The ones without
58:27the learned tag are the ones that are
58:28built in here. You can click to open
58:30that skill. I'm going to hide this hub
58:33down here. And here you can see all the
58:36text of the skill written out. The cool
58:38thing is that the skills are just a text
58:40file like a recipe instructions the
58:42agent follows. Now your agent doesn't
58:44just follow those steps blindly. Imagine
58:47you ask it something like to check a
58:50supplier's price page every week. One
58:52week the supplier redesigns their
58:54website and the old steps don't find the
58:57prices anymore. Instead of failing every
59:00week from then on, Hermes works out the
59:02new path and rewrites its own steps so
59:06that the next run just works. It also
59:09writes new skills on its own as it
59:11solves things and those show up with the
59:13learned label I showed you. There's a
59:15background process that keeps the
59:16library tidy so skills you stopped using
59:19get filed away instead of piling up.
59:22Now, let me show you how you can take
59:23any document, web page, or PDF and turn
59:27it into a skill that Hermes can use
59:29later. Here, I found this SEO starter
59:31guide. It's a whole PDF on how to
59:33optimize copy and web pages for SEO.
59:37This is a great procedural document. So,
59:40I'm going to take this, drag it into
59:43Hermes, and then we'll type the command
59:46forward slashlearn.
59:49And then since the PDF guide is
59:51attached, we don't have to give it any
59:52additional context other than might want
59:54to say turn this PDF into a reusable
59:56skill. If you don't have a file you're
59:58handing it and instead you have
59:59something like a web page, you can also
1:00:00just give it the link. Let's go ahead
1:00:02and send this. And here we can see after
1:00:04a few minutes of thinking, it said the
1:00:06skill was created and verified. We have
1:00:08Google SEO starter as the skill that we
1:00:10can call anytime. So for my project, the
1:00:13skill I wanted it to learn was SEO. for
1:00:16your project. It could be your company's
1:00:18style guide or a course syllabus. Now,
1:00:21you don't have to make all your own
1:00:23skills. You can also go find
1:00:26pre-existing skills. And there's a
1:00:27couple of places to find these. Let's go
1:00:29back to capabilities. And then here in
1:00:32the skills page, if you scroll through
1:00:35and go to under where all of the toggled
1:00:38on skills are, there's a bunch of skills
1:00:40that are available to install. You can
1:00:43simply click install on any of these
1:00:45skills that seem interesting to you.
1:00:47There's also something called skills
1:00:49hub. I minimized this earlier, but you
1:00:51can see here at the bottom of this
1:00:52skills page, there's this little text
1:00:54that says skills hub. And if you click
1:00:56browse the full hub, it opens this
1:00:59little window in the bottom here, which
1:01:01I'll go ahead and expand now so you can
1:01:03see it more. Skillsh is a whole library
1:01:07of skills that other people have
1:01:09written. Now, there's one caution here
1:01:11because a skill is just instructions,
1:01:13and you're trusting whoever wrote them.
1:01:16Hermes scans everything in the hub for
1:01:18things like hidden instructions and data
1:01:20theft, and it blocks dangerous ones
1:01:22outright, but it won't catch everything.
1:01:24So, as you're browsing, read a skill
1:01:27before you rely on it, or even ask your
1:01:29agent to review it for you, and stick to
1:01:30the official trusted sources while
1:01:32you're learning. Now, let's give your
1:01:35agent a personality. And I think the
1:01:37first one of these things I'm going to
1:01:38show you is really fun. So, if you go
1:01:40into any chat, it could be a brand new
1:01:42one or an old one, you can type
1:01:43/personality,
1:01:45and then click the personality command,
1:01:48and it opens up a whole list of options
1:01:50including helpful, concise, creative,
1:01:53teacher, even things like catgirl and
1:01:56pirate. So, let's go ahead and click
1:01:58pirate and then submit that command.
1:02:01After a second, you'll see the
1:02:02personality was set to pirate. And let's
1:02:04ask it something. I'll say, "Write me a
1:02:08short essay about the sun."
1:02:11And here we get a short essay about the
1:02:13sun written in the style of a pirate.
1:02:16You'd think a ball of fire 93 million
1:02:18leagues away wouldn't matter much to a
1:02:20mortal, but there's not a wave that
1:02:22rolls, not a tree that grows, not a
1:02:23breath ye take that isn't the sun's
1:02:25doing. So, that's kind of a fun feature.
1:02:27You can always set it back to a normal
1:02:28personality by sending that command
1:02:31again and choosing none.
1:02:34Now we're back to the base agent
1:02:36behavior. So those presets are
1:02:38temporary. They change how it talks for
1:02:40a session and they don't touch anything
1:02:42permanent. The permanent version is a
1:02:44file called soul.md
1:02:46which holds who your agent is and how it
1:02:49behaves. We already talked about memory
1:02:51which is what it has learned about you
1:02:53and your work. Soul is the character it
1:02:56brings to that work. Keep it short and
1:02:58you can change it by asking. For
1:03:00example, we can say, "Updateyou soulm so
1:03:03you always answer in a calm, direct tone
1:03:06and never end a reply with a question
1:03:07unless I asked for options."
1:03:10And here you can see it updated its
1:03:13soul, its own long-term permanent
1:03:16personality. If I send another message,
1:03:18maybe we'll have it do something online.
1:03:19We'll ask it something like, "Find the
1:03:21best kebab shop in London." Here you can
1:03:25see it keeps it calm, direct, and avoids
1:03:27the typical AI behavior of asking a
1:03:30follow-up question at the end. We've
1:03:32successfully updated its soul. Now,
1:03:34that's level four. It knows how you like
1:03:37things done. It keeps its own
1:03:38procedures, and it sounds the way you
1:03:40want. So, here's your level up task, and
1:03:43pick whichever one annoys you most.
1:03:46Correct it on something it keeps getting
1:03:48wrong. Tell it one thing about you it
1:03:51should always remember. Or point the
1:03:54learn command at a document you're tired
1:03:56of explaining to people and let it turn
1:03:58that into a skill. It takes 1 minute of
1:04:01teaching and you never repeat yourself
1:04:04again. [snorts] Level five is where your
1:04:06agent starts working while you're away.
1:04:09Hermes ships with 16 readymade
1:04:11automations and I went through all of
1:04:13them. Every single one sends you
1:04:15something to read. a morning brief, a
1:04:17news digest, and a weekly review. That's
1:04:20the problem with a lot of AI automation.
1:04:22It gives you more to read instead of
1:04:23less to do. So, let's build the one that
1:04:26isn't like that. We're going to give
1:04:28your agent an actual chore, the kind of
1:04:30thing you do every week and hate, and
1:04:32then you're never going to do it again.
1:04:35Mine is filing receipts. Here's my inbox
1:04:38with a whole bunch of receipts. Now,
1:04:41every purchase, confirmation, and
1:04:42invoice that lands in my email has to
1:04:44end up in a spreadsheet at some point.
1:04:46It takes 20 minutes a week and it's pure
1:04:48admin work. Now, this will run on the
1:04:51agent's own Google account, the one we
1:04:53set up in the last level, so it never
1:04:55touches my personal inbox. Here's the
1:04:58whole thing in one message. This is kind
1:05:00of a long prompt. I'll go ahead and read
1:05:02through what I'm sending here out loud
1:05:04and then I'll walk through it before I
1:05:05send it. So, I'll say, "Set up a receipt
1:05:09filing clerk for me. First, set up a
1:05:12Google sheet called expenses with a
1:05:14receipts tab if I don't have one yet.
1:05:16Then, every day at 7 p.m., search my
1:05:19Gmail for purchase receipts, order
1:05:21confirmations, and invoices from the
1:05:23last 2 days that aren't already labeled
1:05:26files. For each one, pull out the
1:05:29vendor, the date, the total, and the
1:05:31currency, and append a row to my Google
1:05:33sheet, expenses on the receipts tab with
1:05:36columns for date, vendor, amount,
1:05:39currency, category, and the email ID.
1:05:42Then add the files label to that email
1:05:45so it never gets counted twice. This is
1:05:47a pre-approved, recurring job, so append
1:05:50the rows and apply the label without
1:05:52asking me each time. Never send, reply
1:05:56to, forward, delete, or archive
1:05:58anything. Read and label only. When
1:06:01you're done, message me one line on
1:06:03Telegram with how many you filed and the
1:06:05running total for the month. Send me a
1:06:07link to the spreadsheet that I can open
1:06:09from any device. Create the job paused
1:06:11so that I can look at it before it goes
1:06:13live. Okay, so there are three parts of
1:06:15the message doing some work. First, look
1:06:17at the words read and label only. I've
1:06:20[snorts] told it what it may not do in
1:06:23plain language and everything it can do
1:06:25is additive. The worst case is a wrong
1:06:28row in a spreadsheet which I delete.
1:06:31Second, look at the label. That's how it
1:06:34knows what it has already filed. So if
1:06:36it runs twice, it doesn't double up.
1:06:38Your agent starts every scheduled run
1:06:40with no memory of the last one. So a job
1:06:44like this needs something in the real
1:06:46world to check against. Third, create it
1:06:49paused. It builds the job but doesn't
1:06:52switch it on, so you get to read it
1:06:54first. That's that progressive trust
1:06:55again. All right, I'll go ahead and fire
1:06:57this message off.
1:07:00Okay, so instead of just fading from the
1:07:03prompt being sent to the job being done,
1:07:05I'm fast forwarding this so you can see
1:07:08how much thinking and work the agent put
1:07:10into this.
1:07:12in real time. This took about 10 and a
1:07:15half minutes to complete and set up the
1:07:17job before it came back with this final
1:07:19response that you see here. Now, here it
1:07:22says it's done setup, tested end to end.
1:07:25And if we go and rightclick on this link
1:07:29and hit open in external browser, we can
1:07:31now see all of the receipts are now
1:07:35identified in the spreadsheet. If we go
1:07:38into the inbox, we can see that all of
1:07:40the emails have been labeled files
1:07:43exactly as we asked. So, if we go to the
1:07:45label section on the left side of the
1:07:47inbox and click files, all the receipts
1:07:49are here. Now, it ran the job once so we
1:07:52could test everything, but it won't do
1:07:54it going forward until we turn the
1:07:56recurring job on. So, in Hermes, let's
1:07:59go to scheduled jobs in the left side
1:08:01nav. And then here, you can see at the
1:08:03top of the list there is this paused
1:08:05item. You can tell it's paused because
1:08:06it has the orange indicator in the list
1:08:09and it says pause right here next to the
1:08:10name. Here's the receipt filing clerk.
1:08:13It has a prompt that you can read. And
1:08:15what it did was it actually built a
1:08:17deterministic script. So when it calls
1:08:20this job, it's not actually going to
1:08:21call the LLM, which is great for us
1:08:23because it saves us tokens. It does the
1:08:26same job the same way every single time.
1:08:29And it's [clears throat] not going to
1:08:30cost us anything on our NUS portal
1:08:32except the initial trigger that starts
1:08:34the task. So this will be really cheap
1:08:36in the long run for us. To turn this on,
1:08:39we can simply hit resume. And then if we
1:08:41wanted to, we could click trigger now.
1:08:43That'll run the job manually. So you can
1:08:45see how it works. Now here it failed to
1:08:47trigger the crown job because it said
1:08:48it's already running or claimed by
1:08:49another scheduler. You might get errors
1:08:51like that sometimes, and that's okay.
1:08:52Honestly, it ran the job anyway, but you
1:08:54can see how now it'll work going
1:08:57forward. Every day at 7:00 p.m., it'll
1:08:59look at the last 2 days worth of emails
1:09:01and grab all the receipts. Now, think
1:09:04about your own week. Anything you do on
1:09:06a schedule that follows a rule and that
1:09:08ends in a file or a folder or a calendar
1:09:10is a great candidate for this. You could
1:09:12have a tracker that updates itself from
1:09:14your confirmation emails. 15-minute prep
1:09:17blocks added automatically before
1:09:18tomorrow's meetings, or a weekly folder
1:09:21created with the week's notes already
1:09:22inside of it. Just about anything you
1:09:24can think of can be turned into a
1:09:26scheduled job. That job did a chore.
1:09:30This one watches something and only
1:09:32interrupts you when there's a reason.
1:09:33I'm going to tell my agent what I care
1:09:35about once. It'll write its own skill
1:09:37for checking, schedule itself, and then
1:09:39from then on, my phone only buzzes when
1:09:41something's actually new. In a new
1:09:44session, let's say something like this.
1:09:47Watch Barcelona to London flight price
1:09:49trends for me. Build yourself a reusable
1:09:52skill for the check. Keep track of what
1:09:53you've already shown me. Schedule it
1:09:55every few hours. and message me on
1:09:57Telegram only when there's something new
1:09:59written with your concise writer skill.
1:10:02So, this job was a cool one to watch.
1:10:03This one actually took about 30 to 45
1:10:07minutes somewhere in there. I actually
1:10:08stepped away and came back because it
1:10:10was taking a while. And the reason for
1:10:12that was I didn't give it super super
1:10:14specific process instructions like I did
1:10:16for the email one. So the agent went on
1:10:19to the live internet, figured out the
1:10:21best way to do this job, built the
1:10:23machinery, built a skill, put on a
1:10:26schedule, and also set up some memory so
1:10:27it can keep track of what it's already
1:10:29shown me. And then here after all that
1:10:32work is its first response. Basically
1:10:34saying that it'll message on Telegram
1:10:36only when something changes. If we go to
1:10:39the schedule jobs, we can see now we
1:10:42have this Barcelona London flight price
1:10:44watch. Cool thing is too, it's another
1:10:46deterministic script job. So it'll run
1:10:49this script every 6 hours fetching the
1:10:52latest prices. And there you can see
1:10:55unprompted after something changed. This
1:10:58is what it looks like when the flight
1:10:59watch job finds something new. I
1:11:02described an outcome once and now it
1:11:04comes to me. My own agent runs more than
1:11:0720 of these. One checks my bookkeeping
1:11:09every Monday. One tracks birthdays and
1:11:11starts researching gift ideas while
1:11:13there's time to order. One works on a
1:11:15trip. I'm planning twice a day, getting
1:11:17quotes, and coming back with the best
1:11:19option. And one follows up on a
1:11:21negotiation, so I don't have to remember
1:11:23to do it. Now that your agent runs jobs
1:11:25on its own, the model choice starts to
1:11:28matter because you're paying for every
1:11:30one of those runs. Models are like
1:11:32workers you can hire by the job. Some
1:11:34are brilliant and expensive, some are
1:11:36quick and cheap, and each one can work
1:11:39at different effort levels. You wouldn't
1:11:41hire the most expensive person, you
1:11:42know, to sort your mail. Here in Nouse
1:11:45portal on the left side, if you go to
1:11:47the resources section and select models.
1:11:51Here you can take a look at the active
1:11:53promos for various models, how much they
1:11:55cost normally versus how much they cost
1:11:57currently. And occasionally there are
1:11:59deals where you can get more power for
1:12:01less money by switching to some of these
1:12:04discounted models. Now, the more
1:12:05important list is actually the one below
1:12:07this all models list here. Here you can
1:12:10scroll through or even filter and see
1:12:12the list of all the different models
1:12:14offered through NAS portal and see the
1:12:18portal price. Here for example, we're
1:12:19using DeepSeek V4 Flash and that's 7
1:12:22cents per million tokens in and 14 cents
1:12:26per million tokens out. Claude Fable 5.1
1:12:29for comparison is $10 per million tokens
1:12:33in and $50 per million tokens out. That
1:12:36is a very expensive model, but it's very
1:12:38powerful. You can also see that there's
1:12:40some free models listed in here, but
1:12:42keep in mind that not all models perform
1:12:44the same. So, you may need to do a
1:12:46little bit of experimentation to see
1:12:47which ones work best for you. Routine
1:12:49work that follows a skill, like checking
1:12:51a feed or filing receipts runs just fine
1:12:54on a cheap model. Though I'd avoid the
1:12:56very cheapest and the free ones because
1:12:57a model that can't reason will make
1:13:00confident mistakes like filing the wrong
1:13:02thing or throwing away something that
1:13:04you wanted. Work that needs judgment
1:13:07like planning analysis or anything where
1:13:09being wrong costs you deserves a top
1:13:12model. And Hermes honestly feels
1:13:14dramatically better on a flagship model
1:13:16especially one that can hold a lot of
1:13:18information at once. So, if you're
1:13:20underwhelmed by the experience you're
1:13:22getting inside of Hermes, before you
1:13:24blame the tool, try a stronger model or
1:13:27a higher effort setting. I'll show you
1:13:29where you can switch those. Now, in the
1:13:31Hermes desktop app, on the right side of
1:13:33the text box where you type your
1:13:34messages, there's a model selector. If
1:13:36you click that, it'll expand to show you
1:13:39all of the different models that are
1:13:40available to you on your connected
1:13:42provider. In this case, N portal. So, if
1:13:45I had a task that I needed to do that
1:13:47was really, really important, I might
1:13:50switch up to using Fable 5.1. And then I
1:13:53could send my message here. To the right
1:13:54of that, there's a drop down that if you
1:13:56click to expand, it shows you the effort
1:13:58level here. Right now, Fable 5.1 is set
1:14:01to medium, which for most tasks is good,
1:14:03but you could also scale it up to ultra.
1:14:06Just keep in mind ultra on something
1:14:07like Fable 5.1 is going to cost a lot of
1:14:10money. If you switch back to something
1:14:12like deepsek and here I can type to
1:14:14search for it deep seek v4 flash. We'll
1:14:18use this one. You could also change the
1:14:21effort level here. So if you feel like
1:14:22you're not getting quite as much out of
1:14:24deepseeek as you'd like, you can also
1:14:26increase the effort level to something
1:14:27like ultra or max. But for everyday
1:14:29chatting, that's kind of overkill. Now
1:14:30I'll show you how to update that in
1:14:32Telegram as well. In the Telegram chat,
1:14:34type slash and then model. And that'll
1:14:38open up your model configuration. So
1:14:39here you can toggle to see all the
1:14:41different models that are available to
1:14:43you. So let's say I switched to Fable
1:14:445.1 here for example. Then you can do
1:14:47slash and then reasoning.
1:14:50And this is where you can select your
1:14:52reasoning level. They label it reasoning
1:14:54instead of effort here. I'll just stick
1:14:55with medium. Now there are two things
1:14:58I've run into myself that I want to
1:14:59share with you so that you can avoid any
1:15:01surprises in your own setup. First, you
1:15:03can switch models at any time. And I
1:15:06actually love that about Hermes. You can
1:15:07change the model and the effort level
1:15:09mid conversation, but I should point out
1:15:12that switching models in the middle of a
1:15:14long conversation makes the new model
1:15:16reread everything in the conversation
1:15:18from the start, which costs more. So for
1:15:22a big job, start a fresh thread on the
1:15:24new model. Second, every scheduled job
1:15:26remembers which model it was created
1:15:28with. If you change your default model
1:15:30later, an unpinned job, meaning it
1:15:33wasn't pinned to a specific model will
1:15:36refuse to run rather than quietly switch
1:15:38to a more expensive model. That's
1:15:40deliberate. And the fix is to pin the
1:15:42job to a model on the cron page. Now,
1:15:44where you go to find that is go to the
1:15:46schedule job section of the desktop app
1:15:48if you're using that and then find your
1:15:51cron jobs in the list on the left. Click
1:15:53the three dots next to the name of it
1:15:55and click edit cron. Here you'll see a
1:15:58model selector for jobs that use an LLM.
1:16:01Now, these are actually both
1:16:03deterministic script jobs in Micron list
1:16:06here. So, these don't actually call the
1:16:07LLM. So, there's nothing to switch and
1:16:10select here. So, these actually will
1:16:11work just fine for me. But, if you have
1:16:14a job that calls a model, you should
1:16:15have a drop-own menu in here to select
1:16:18which model you want to use. Now, while
1:16:20we're on this subject, let's set up a
1:16:22fallback. Models stop working sometimes
1:16:25and plans hit their limits, usually
1:16:27while you're asleep and a scheduled job
1:16:29is running. So, give your agent a short
1:16:31list of models it drops to when the
1:16:34first one won't answer. In my own setup,
1:16:36mine goes from a top model to a second
1:16:38one from a different company and then
1:16:40finally to a cheap one that's almost
1:16:42always available. It switches mid-con
1:16:43conversation without losing anything so
1:16:45that you can keep going and your
1:16:47scheduled jobs use the same list. To do
1:16:49this, we'll go back to the chat.
1:16:52I'll start a new session and then say
1:16:54something like this. And here I've got
1:16:57my sample prompt in brackets. This is
1:16:59where I would recommend you go through
1:17:01and update your prompt to match your own
1:17:04desired setup. So here I have it saying
1:17:06set up a fallback chain. Try flagship
1:17:09model first, then a model from another
1:17:12provider, then a cheap reliable model.
1:17:15Switch each of these for your own
1:17:18specific desired fallback chain. A good
1:17:20way to figure out which ones you want to
1:17:22use is to use this model list inside
1:17:24your NUS portal. So for example, let's
1:17:27say we wanted to use DeepSeek V4 Pro for
1:17:31our main one. We can go ahead and copy
1:17:34this name from the list here, paste it
1:17:35in for our flagship model,
1:17:39then a model from another provider. So
1:17:41let's find another brand name. Maybe
1:17:42we'll use something like Claude Sonnet
1:17:445, for example. It's more expensive, but
1:17:46it's reliably available. So we can go
1:17:49ahead and grab claw sonnet 5 and then
1:17:52paste that in for our model from another
1:17:54provider and then a cheap reliable
1:17:57model. And for your own setup, this may
1:17:59be where you include something like a
1:18:01free option as a last resort. Maybe
1:18:04we'll use something here like this
1:18:06upstage solar pro 4 and then place that
1:18:10here in for our final model. Go ahead
1:18:14and hit send here. You can see after a
1:18:16bit of processing, it comes back and
1:18:18says the fallback chain is configured.
1:18:21The primary is whichever model you're
1:18:23working with right now. And then you
1:18:24have several different fallbacks. Okay.
1:18:27There's one more command for hard
1:18:29questions called mixture of agents. In
1:18:33your chat, put slash M O A. And here
1:18:37you'll see it in the list. I'll go ahead
1:18:38and select that. put that in front of a
1:18:40request and Hermes will send it to
1:18:43several different models at once. Then
1:18:46it combines the best of each answer into
1:18:49one reply. You're paying for several
1:18:51answers instead of one, so save it for
1:18:54decisions that deserve a second and
1:18:56third opinion. For my example here, I
1:18:59might do something like a benchmark test
1:19:01to see if these are actually good models
1:19:02for the fallback chain or use something
1:19:04different. So, I'll say something like,
1:19:06I want to do a quick benchmark test. Are
1:19:08these the best model to have in my
1:19:10fallback chain? If I want to have a
1:19:12flagship model, then reliable model,
1:19:14then cheap or free model for the chain,
1:19:17all of which are capable of performing
1:19:19my tasks. Well, let's send it. Now,
1:19:22you'll see that the / MOA command was
1:19:25recognized, but after a second, it said
1:19:27that there was a provider error. So, no
1:19:29LLM provider was configured for this.
1:19:31So, to do this, we can actually click
1:19:32the switch provider button that popped
1:19:34up. And then in the model settings page
1:19:36that pops up, go ahead and scroll down
1:19:38to the mixture of agents section. Then
1:19:40you can go in and select your reference
1:19:42models. So we can click in to select
1:19:45something from now portal. You could
1:19:47pick a specific model here. For our
1:19:49reference model one, let's say something
1:19:51like
1:19:53opus 5 from anthropic. For the reference
1:19:562, we'll grab something else from now
1:19:58portal. We'll say
1:20:01chatgpt 5.6 6 Luna as an example. And
1:20:05maybe one more. We'll grab a Deepseek
1:20:08model here from Nouse Portal. Go ahead
1:20:10and grab Deep Seek V4 Pro. We've got
1:20:14those locked in. And while we're in
1:20:15here, I'll point out too, if you scroll
1:20:17all the way to the bottom of this page,
1:20:18this is where you manage your fallback
1:20:19models. Here you can see that you could
1:20:21actually manage your fallbacks manually
1:20:23if you prefer. One more thing to check
1:20:25actually too is your aggregator. So what
1:20:27model are you using to actually assemble
1:20:30all the answers that come back from each
1:20:32of those references into one final
1:20:34answer here? So that'll be your
1:20:35aggregator at the bottom of this
1:20:37reference list. So we want to make sure
1:20:39that this is also selected to a model
1:20:41that's actually available to us. I'll go
1:20:42ahead and select DeepS v4.1 flash. It's
1:20:46essentially just taking all the best
1:20:47bits of each answer and summarizing the
1:20:50final results. So we'll go ahead and
1:20:51roll with that. Close out of this. And
1:20:54then let's go ahead and copy our prompt.
1:20:56I'm actually going to open up a fresh
1:20:57session,
1:20:59type MOA again, paste our prompt, and go
1:21:04ahead and run it. You'll see it gets to
1:21:05work pulling in the first of the three
1:21:07reference models. And this can take a
1:21:09while. Now, I want to mention that this
1:21:11might be kind of expensive for credit
1:21:12usage because not only are we using
1:21:15three different models and an aggregator
1:21:17to compile all the results, but I also
1:21:19selected a couple of more expensive
1:21:20models. In the next chapter, I'll talk a
1:21:22bit more about costs, but for now, let's
1:21:24go ahead and fast forward to when this
1:21:26is done. This could take quite a while.
1:21:29After a bunch of thinking here, which
1:21:31took about 26 minutes, this is what it
1:21:34came back with. It came back with this
1:21:36chart here saying this is how it would
1:21:39suggest that I revise my particular
1:21:41fallback chain. Go with Opus 5 first,
1:21:44Deepseek V4, then Deepseek V4 Flash, and
1:21:47then Solar Pro as the last resort. This
1:21:50is obviously just an example. Models
1:21:52change all the time and really you want
1:21:54to pair the model you're using to the
1:21:56task you're doing. So this is great for
1:21:58my particular setup, but it's just
1:21:59illustrative. I'd recommend doing
1:22:01something like this that targets your
1:22:02own setup and purpose rather than
1:22:04copying my example chain here. Your
1:22:07agent has now done a chore, watched a
1:22:09topic, and used a few different models.
1:22:12Let's look at what that cost. Here in
1:22:14the Nouse portal dashboard on the left
1:22:17side under account go to usage and here
1:22:21you can see your total credit spend your
1:22:24total input tokens total output tokens
1:22:26cash reads cash rights and total number
1:22:29of requests and you can see that over
1:22:31various periods of time here you can
1:22:33look at it by day for example you can
1:22:35use these filters to change to see it by
1:22:37model see different chart styles and
1:22:40even change your display period. Now
1:22:42look at the token count. This is the
1:22:44total amount of text going back and
1:22:46forth against the money. So input tokens
1:22:49are the prompts you send and output
1:22:52tokens are the outputs from the models.
1:22:55Both inputs and outputs cost money and
1:22:58deplete your credit balance. Every time
1:23:01you send a message, your agent resends
1:23:03its whole briefing, its instructions,
1:23:06its memory, its tool list, and the
1:23:08conversation so far. That rereading is
1:23:11called context and it's almost your
1:23:14entire bill, but it stays cheap because
1:23:17the portal recognizes the part that
1:23:19didn't change and charges a fraction to
1:23:22read it again. That's called caching.
1:23:25You don't need to set that up. It just
1:23:26happens by itself. As far as costs go,
1:23:29the server bill is fixed through
1:23:31Hostinger, so that's just fine. But the
1:23:34big cost to watch is your model bill.
1:23:36I've been filming this tutorial over a
1:23:38couple of different days. And here you
1:23:40can see my cost per day. Day one was a
1:23:43little bit more than 8 cents. And then
1:23:45today, as I'm recording this chapter and
1:23:47the chapter with MOA, a mixture of
1:23:49agents, that's today. You can see that
1:23:51MOA task alone cost me a lot of money.
1:23:54The calls to Anthropics Opus 5 cost
1:23:58$130. Compare that to all this usage
1:24:01from the previous days where I was just
1:24:02using Deepseek V4 Flash. This is way
1:24:05cheaper to run. You can see the cost
1:24:06broken down by model. So, as you're
1:24:09trying to figure out which model to use
1:24:10over time, like I said in the last
1:24:12chapter, make sure you're pairing the
1:24:14model to the task. If you default to
1:24:17using something like Anthropics Claude
1:24:19Fable 5.1 for everything, you're going
1:24:21to burn through a lot of money. That and
1:24:24Astra are super helpful to call when the
1:24:26task demands it. But for most daily
1:24:29work, I recommend trying to find a model
1:24:31that achieves a good balance of cost
1:24:34efficiency and effectiveness. This chart
1:24:36will help you determine that over time.
1:24:38So, that's level five. Your agent does a
1:24:41chore. It watches what you care about,
1:24:43and now you know what you're spending.
1:24:45So, here's your level up task. Pick one
1:24:48thing you do every week that ends in a
1:24:50spreadsheet, a folder, or a calendar,
1:24:52and give it away. Ask the agent to do it
1:24:55for you. Create the job paused. read
1:24:57what it wrote, then switch it on to run
1:25:00all the time. Level six is giving it a
1:25:03team. The simplest version needs no
1:25:05setup at all. Ask your agent to split a
1:25:08job and it starts helpers that work at
1:25:10the same time and hand you one answer.
1:25:12For example, I could say something like
1:25:14this. Plan a 3-day trip to Lisbon,
1:25:16Portugal from London, leaving November
1:25:185th. one sub agent on things to do, one
1:25:22on best places to eat for authentic
1:25:24local food, and one on the best flights
1:25:26itinerary, then give me a single
1:25:29combined plan. Here you can see that the
1:25:32main agent that we're talking to is
1:25:34delegating the task to three sub aents.
1:25:36And then you can see those three sub
1:25:38aents getting to work, each running on
1:25:40DeepSeek V4 Flash. Now, you could also
1:25:43ask for your main chat to spin up these
1:25:45sub agents on specific models. So, they
1:25:48don't all have to run on the exact same
1:25:50one. All these agents are now running in
1:25:52parallel. So, instead of using one model
1:25:54to do all the work back to back, instead
1:25:57there are three researchers going out at
1:25:59the same time and then they'll all come
1:26:01back together, which makes this whole
1:26:02process really quick. Now, of course,
1:26:04really quick is relative to the task at
1:26:07hand. So, I'm still going to go ahead
1:26:08and skip forward to when it's done just
1:26:10to save time for the video. And here we
1:26:12go. After a few minutes, I've got this
1:26:14beautiful 3-day itinerary for Lisbon.
1:26:17The agent we're messaging just
1:26:19coordinated three others working at
1:26:21once. Now, those helpers are temporary.
1:26:25They exist for that one job and then
1:26:27they're gone. For a standing team, you
1:26:30create teammates. A teammate in Hermes
1:26:33is called a profile. a separate agent on
1:26:35the same server with its own memory, its
1:26:37own skills, its own personality, and its
1:26:40own model. The desktop app presents
1:26:42profiles as bots with a name, a title, a
1:26:44job description, and a face. It's the
1:26:46same thing underneath, though, and it's
1:26:48a friendlier way to make them. So,
1:26:49that's what we'll use. Now, before I
1:26:51make them, I'm going to have my agent
1:26:53write the job descriptions because it
1:26:54knows what the project board we build
1:26:56next will need better than I do. In the
1:26:59main chat, let's say this. I'm about to
1:27:02create three teammate profiles on this
1:27:04server for a project board, a
1:27:06researcher, a writer, and a reviewer.
1:27:09Write a one-s sentence job description
1:27:10for each one worded so the board can
1:27:12tell from the description alone which of
1:27:14them a task belongs to. Give me three
1:27:17lines that I can copy and paste. Let's
1:27:20go ahead and run that. And here are the
1:27:23three oneline job descriptions. I'm
1:27:25going to go ahead and copy these and
1:27:26paste them into a text file real quick
1:27:30just so I have them somewhere else
1:27:31quickly accessible.
1:27:35Do a quick little format just so I can
1:27:36see them. And there we go. That saves a
1:27:39lot of trial and error because the board
1:27:41reads these descriptions to decide who
1:27:43gets what. Next, we're going to go to
1:27:45the upper left portion of the app here
1:27:48and change from the sessions tab to
1:27:50bots. Now, if you don't see this here,
1:27:53it might have moved somewhere else on
1:27:54the app. Search around for it, or even
1:27:56ask Hermes where that new bots tab is.
1:27:59When you get here, go ahead and click
1:28:01the plus button to add a new bot, and
1:28:04then click new bot. And then here, you
1:28:06can give your bot a face. For example,
1:28:09you could grab from any of these
1:28:10existing bot shapes. Click to generate a
1:28:13face by describing your avatar, upload a
1:28:16picture, or even choose from a bunch of
1:28:19pets. Maybe for example, I want this
1:28:22penguin. Yeah, that looks fun. Next,
1:28:25let's go ahead and give it a name. Going
1:28:27back to my text edit here, our first bot
1:28:29will be called researcher. So, I'm going
1:28:30to go ahead and copy this job
1:28:31description. Get it ready. Copy and give
1:28:35the name researcher.
1:28:38Give it the title. Same thing,
1:28:40researcher. And then paste in the job
1:28:42description. You can click advanced and
1:28:44then scroll down and give it additional
1:28:46specifications if you'd like to. If you
1:28:48don't do anything, it'll just clone
1:28:49everything from your original profile,
1:28:52your default profile. But here, I want
1:28:54to set it so that the researcher always
1:28:56uses a higher power model. So, I'll go
1:28:59ahead and change provider to specify NAS
1:29:02portal. Then model, we'll go ahead and
1:29:05choose opus 5 for the researcher cuz it
1:29:08goes deep and can do long running
1:29:10research tasks. Everything else we'll
1:29:12leave the same. We'll go ahead and click
1:29:13create bot. And now we have our
1:29:15researcher. Go ahead and repeat that
1:29:17process for any other bots that you want
1:29:19to make. I'll do this quickly for the
1:29:20writer and reviewer, choosing models
1:29:22appropriate for each type of task that
1:29:24they'll do. Okay, I've created those
1:29:26additional two bots. Let's go ahead and
1:29:28say hello to our critic here, our
1:29:30reviewer. Just to prove that it works,
1:29:32I'll say hello.
1:29:36It says, hey, nothing queued up for me
1:29:38to review right now. So, I'm just here
1:29:39ready to go. And you are. So, it already
1:29:41knows its job specifically. Now, I
1:29:43created the research analyst with a top
1:29:46model because judgment is its whole job.
1:29:49The other two got cheaper models, and I
1:29:51put them all on different models just so
1:29:53that we're not experiencing any bias
1:29:55introduced by one single model type.
1:29:58This helps keep the agent system as a
1:30:01whole a bit more honest. Now, let's look
1:30:03at the board. Hermes has a project board
1:30:05called Conbon that works like Trello, a
1:30:08board of columns that your jobs move
1:30:09through from left to right. Every job on
1:30:12it is a card. The difference is that
1:30:14these cards don't wait for you. You put
1:30:17a job on the board. The board splits it
1:30:19into steps, hands each step to the right
1:30:21teammate, and they work through the
1:30:22columns while you watch or while you
1:30:24sleep. I'm doing this in the desktop app
1:30:26because that's where I work, and it's
1:30:28switched off by default. So, let's turn
1:30:30it on. Head back to the sessions tab,
1:30:33then go to capabilities, then go to the
1:30:36plugins tab, and then here you'll find
1:30:38conbon. Like I said, toggled off by
1:30:41default. Go ahead and toggle it on. Now
1:30:44it appears in the left side column. Go
1:30:46ahead and click to open it. And here's
1:30:48the conbon board. Now, you can't see any
1:30:50columns yet because there aren't any
1:30:52tasks on this board. So, let me just go
1:30:53ahead and quickly create a dummy task.
1:30:55Something with nothing on it. So, you
1:30:56can see how this works. I'll just say
1:31:01test task and then create task. Now,
1:31:04here you can see some columns. Triage is
1:31:07where raw ideas land. and the board
1:31:09turns them into a proper description of
1:31:11the work. To do is waiting on other
1:31:15cards. Ready means it can start. Running
1:31:18means a teammate has it. Blocked means
1:31:22someone needs you. Review is a second
1:31:26agent checking the work. And done
1:31:29releases whatever was waiting on it.
1:31:31There's also a scheduled column which
1:31:32tasks might get put in if they need to
1:31:34run at a specific time. Now, three of
1:31:36these columns you can't drag a card into
1:31:38because the board sets them itself. Here
1:31:40you can see it just picked up the task
1:31:42and put it into running. Now, if you're
1:31:44using the web dashboard instead, let me
1:31:46show you where to find this. Here on the
1:31:48left side, scroll down in the navbar and
1:31:50then conbon is here at the bottom under
1:31:52plugins. And this will show you all of
1:31:54those different columns right away. Now,
1:31:56before we put a real job on this, here's
1:31:58something I learned from my own
1:31:59mistakes. It runs quietly in the
1:32:02background, which sounds great until a
1:32:04card gets stuck waiting on you and
1:32:06nothing tells you. You'd come back an
1:32:08hour later thinking it had been working
1:32:09the whole time. So, there are two fixes
1:32:11and I use both. The first is to create
1:32:14your cards from Telegram with the conbon
1:32:17command instead of clicking new task in
1:32:19the app because a card created that way
1:32:22subscribes you to it automatically and
1:32:24you get a message when it finishes or
1:32:26gets stuck. The second is to write the
1:32:29notification instructions into the job
1:32:31itself in plain language, the same way
1:32:33you'd tell a person who to call if they
1:32:35get stuck. Here in Telegram, type the
1:32:38slash command conbon.
1:32:42And here it shows you how to actually
1:32:44use this command.
1:32:46This command takes a bunch of
1:32:48subcomands.
1:32:50The most important one is create. Now,
1:32:53this looks like a lot of information.
1:32:54So, if you're not quite sure how to use
1:32:57this, what you can do is simply ask
1:32:59Hermes to tell you. I'll say, "Show me
1:33:02an example of how to write a create
1:33:04command for the conbon board."
1:33:07And here it comes back with an example.
1:33:09So, if we scroll up, we can see it from
1:33:10the beginning. It says, "Here's a clean
1:33:12create command." And I actually like the
1:33:14minimal version here. It says title is
1:33:16the only required argument. So you would
1:33:18type slashconbon
1:33:21create and then give it a title. So
1:33:23let's go ahead and type a command. We'll
1:33:26say slashconbon.
1:33:28And you'll notice it doesn't select a
1:33:30command from the list of commands in
1:33:31Telegram. That's because this is a less
1:33:34commonly used command. That list only
1:33:36contains really common ones. So don't
1:33:38worry about it if it doesn't look like
1:33:39it selects anything. So then we'll type
1:33:42create. Then in quotes we'll give the
1:33:44task a name. We'll just say test task.
1:33:47Then we'll say dash dash. And here
1:33:50there's this issue sometimes if you're
1:33:51typing this on a computer where it
1:33:52merges those dashes. So here's one quick
1:33:54trick around that. You can type dash and
1:33:57space dash and then delete that space
1:34:00and now it works as two dashes. Then
1:34:02we'll say assigney and then just say
1:34:05default. That way it just assigns it to
1:34:07the default agent. All right, we'll go
1:34:10ahead and submit that. And here you can
1:34:12see it created the task and it
1:34:14subscribed us to it so that we'll be
1:34:16notified when it's completed. Let's go
1:34:18ahead and check it out in the desktop
1:34:20app to make sure it actually showed up
1:34:22there. Here in the desktop app, I don't
1:34:23see it. So this is when you can refresh
1:34:25the app. So on Mac, you can hit command
1:34:27R or on Windows control R.
1:34:31And here it is. The test task is
1:34:33running. So that Telegram command worked
1:34:35for us. Back in Telegram, when it's
1:34:37done, we'll get a notification. And
1:34:40there we go.
1:34:41The task is complete. There was nothing
1:34:43really to do, so it came back pretty
1:34:44quickly. Back in the desktop app, we can
1:34:46see the test task is now in the done
1:34:49column. So, that's how you do it from
1:34:50Telegram. Now, let's set one up from the
1:34:53desktop app, which is how I tend to
1:34:55actually do this myself. Let's give it a
1:34:57real job. This time, I'm going to use
1:34:59the board for something I actually use
1:35:01my agent for, planning a trip end to
1:35:04end, including the part I always put
1:35:06off, which is tracking what's actually
1:35:08been booked.
1:35:10Now, I did a dumped down version of this
1:35:11earlier showing you how you could run a
1:35:13command and ask for some sub agents to
1:35:16do something, but here I'm not going to
1:35:18tell it exactly how to do the work. I'm
1:35:20just going to give it the task and it'll
1:35:22figure out who to assign it to. So,
1:35:24let's go ahead and click new task in the
1:35:26upper right. And I'll say Lisbon trip
1:35:30plan for the title. And for the
1:35:32description, I'll say plan a 5-day trip
1:35:34to Lisbon for two people in early spring
1:35:36on a mid-range budget. research flights
1:35:38from London, three neighborhoods to stay
1:35:40in, and three day trips, all with real
1:35:42prices, and a direct booking link for
1:35:44anything I'd have to buy. Build a budget
1:35:47spreadsheet in my drive outbox folder
1:35:49with a row for every item. What it is,
1:35:51the price, the booking link, a booked
1:35:54column, and an actualized column for
1:35:56what I really paid. Write a dayby-day
1:35:58itinerary next to it. Put placeholder
1:36:01events on my calendar for the travel
1:36:02days, and anything time-specific. Check
1:36:05everything for errors and missing prices
1:36:07before you show me. Message me on
1:36:09Telegram when it's done and message me
1:36:11there if you get stuck and need an
1:36:13answer from me. All right, that's the
1:36:15full prompt. It's pretty elaborate. This
1:36:17is where I recommend thinking about a
1:36:19complex [clears throat] big task of your
1:36:21own and how you might want to hand it
1:36:23off to an agent. Think about the exact
1:36:26deliverables you need. Think about the
1:36:28format you want the work in. And think
1:36:29about any context it needs. Once you've
1:36:31got all that dumped in there, go ahead
1:36:33and click create task. Right away,
1:36:35you'll see it goes into the triage
1:36:37column. And it might sit here for a
1:36:39minute, especially with more complex
1:36:40tasks as the orchestrator agent figures
1:36:43out how to break this task down into
1:36:46smaller pieces and delegate the
1:36:48different subtasks to various agents.
1:36:51And after a few minutes, you can see it
1:36:53decomposed the task into smaller
1:36:55subtasks. Now, three of them are running
1:36:58and then three are in the to-do list
1:37:00below the big Lisbon trip plan. So,
1:37:04right now, it's currently researching
1:37:05flights, neighborhoods, and day trips.
1:37:08You'll notice if I click into any of
1:37:10these items, like for example, add the
1:37:12placeholder calendar events, it has
1:37:16dependencies. So, it's blocked by this
1:37:19task here, and it blocks these two
1:37:22tasks. It can't start until this is
1:37:24done. And then these two tasks can't
1:37:27start until this one's done. The main
1:37:29agent is coordinating all the tasks,
1:37:31moving them along the pipeline. And so
1:37:33now, for the most part, this is just
1:37:35going to work in the background, and we
1:37:36can step away for a while. But before we
1:37:38do, let me just show you one more thing.
1:37:40If I click into any of these research
1:37:41tasks, you'll see here that this
1:37:43research job was assigned to the
1:37:46researcher. If I click the drive budget,
1:37:49it kept the default assigne. The review
1:37:53task was assigned to the reviewer. So
1:37:55the specific agents and models that are
1:37:58best for each task have been selected
1:38:01based on their individual expertise.
1:38:04Every card keeps its own history and
1:38:06comments too. So if you scroll down
1:38:08through any of these, like for example,
1:38:10this currently running task, you'll see
1:38:13the activity logs here. You'll see all
1:38:16of the work that it has done. And you
1:38:18can even see it's working in real time
1:38:19right now running commands. You can add
1:38:21comments to the comments section here if
1:38:23you need to give it some notes about the
1:38:25task. And the agent running the job as
1:38:27well as the orchestrator will add
1:38:28comments in here too as this card moves
1:38:31back and forth between various columns.
1:38:33This is a full project management
1:38:36ecosystem managed entirely by an
1:38:39orchestrator AI delegating the tasks to
1:38:42individual experts. It's almost like a
1:38:45little mini company. Can a card wait for
1:38:47the hotel to email you back and then
1:38:49carry on when the reply arrives? Yes.
1:38:51So, a teammate that can't finish moves
1:38:54the card to blocked and writes down why.
1:38:57It'll put that note in the comments. So,
1:39:00if you have a job where the task is to
1:39:02manage an email conversation, for
1:39:04example, you need a second job to wake
1:39:07it up because an individual task can't
1:39:10just sit there watching your inbox. Each
1:39:13task is a job that runs and then ends.
1:39:16That's why we'll create a second job.
1:39:17It's a scheduled job that checks the
1:39:19board. Mine runs every half hour. It
1:39:22reads the blocked cards and when the
1:39:23thing they were waiting for is arrived,
1:39:25it moves them back to ready and they
1:39:27carry on. If a card genuinely needs me,
1:39:30the scheduled job messages me instead of
1:39:32leaving that card sitting there. So,
1:39:33let's build the simple version of that
1:39:35here. I'll create a new session in the
1:39:37left
1:39:39and then we'll say every morning check
1:39:41my email for booking confirmations for
1:39:43the Lisbon trip. For each one fill in
1:39:45the actualized column in the trip
1:39:47spreadsheet with what I actually paid.
1:39:49Tick booked. Delete the matching
1:39:51placeholder event from my calendar and
1:39:52create the real one with the
1:39:54confirmation number in the description.
1:39:55Add guest at medexmat.com as a guest on
1:39:58any event they're joining. Label the
1:40:00confirmation email trip so it isn't
1:40:03counted twice. Message me one line on
1:40:05Telegram with what changed. Create it
1:40:07paused so I can read it first. Now,
1:40:09here's one warning before you copy that
1:40:11word for word. Adding guests to a
1:40:13calendar event sends those people a real
1:40:14invitation. So, only put names in there
1:40:16when you mean it. And check the job
1:40:18while it's paused. So, that's the shape
1:40:21of every longunning project you'll build
1:40:23with this. A board does the work. A
1:40:25scheduled job keeps it current and tells
1:40:27you when you're needed. For your own
1:40:29work, the same board could run a weekly
1:40:31report, a batch of product descriptions,
1:40:33a content calendar, or a research and
1:40:36outreach list. Two bits of housekeeping
1:40:38before we finish. Because you now own a
1:40:40small system, if anything ever
1:40:42misbehaves, it's the same three checks
1:40:44from level two, credit, gateway, and
1:40:48logs. And updates are one click. In the
1:40:51web dashboard on the left side nav,
1:40:53scroll down to system and select that.
1:40:56Here you can see a quick overview of
1:40:57everything that's set up and how it's
1:40:59running. And if you scroll down to the
1:41:00gateway section, you can see that it's
1:41:02currently running. But you can always
1:41:04click stop, restart it, or start it from
1:41:06this menu here. Further down, there's
1:41:09even a backup section here. So you can
1:41:11create a backup by clicking this button,
1:41:13which is super helpful to do any time
1:41:15before you make a big change to your
1:41:17configuration. You can also restore from
1:41:19backups from this section here, too. For
1:41:20example, if you download a backup, you
1:41:22can then upload to restore the state
1:41:24from that file here. Back in the hosting
1:41:27or docker manager, you can also update
1:41:29your software from this menu. Click on
1:41:31the three dots to the right of the
1:41:33application. Then click update. You'll
1:41:36get a little pop-up notification saying
1:41:38your application is being updated. And
1:41:40this updates the version of Hermes that
1:41:42you have running on your VPS, also known
1:41:44as the backend. To update the desktop
1:41:47app, open settings, go to about on the
1:41:49left side.
1:41:51Check if there's a new update available,
1:41:53and if there is, click update. Now,
1:41:57it's important to note that the desktop
1:41:59software and the software running on
1:42:01your VPS both update separately. Now,
1:42:04there's one more page, and this is a fun
1:42:07one. Open your web dashboard, and in the
1:42:09plug-in section in the nav under conbon,
1:42:12there is an item called achievements. Go
1:42:14ahead and click on that. Hermes reads
1:42:15back through everything your agent has
1:42:18ever done and gives you badges for it.
1:42:20There are 60 of them across categories
1:42:22like agent autonomy, debugging chaos,
1:42:25and tool mastery. They run from copper
1:42:28up to Olympian, and some stay secret
1:42:30until you accidentally trigger them.
1:42:32Mine here says I've unlocked seven out
1:42:34of 60. I've discovered 53, and my
1:42:37highest tier is gold, so I have a little
1:42:39bit of working up to do to get to
1:42:41Olympian. Most of these achievements are
1:42:42for things I didn't even know I was
1:42:44doing. Once your agent has been running
1:42:46for a week, come back to this page and
1:42:47tell me in the comments which badge
1:42:49surprised you and how many you've
1:42:50unlocked. I'll read them all. So, look
1:42:52at what you've got now. You've got an AI
1:42:54agent on a server that's yours,
1:42:56reachable from your phone and your
1:42:57desktop, connected to your files and
1:42:59your accounts, and trained on how you
1:43:01like things done. It runs a chore every
1:43:03week that you'll never do again, and it
1:43:06has a board where a team of agents work
1:43:08through a project on their own. My
1:43:10advice from here is to start small. Ask
1:43:12it questions. Give it one real job. Let
1:43:15it learn how you work. And once you
1:43:17trust it, hand over the thing you've
1:43:19been dreading all month. Tell me in the
1:43:21comments which level you're on and which
1:43:22one you're building to this week. And if
1:43:24you haven't set your agent up yet, go
1:43:26down to the description now and click
1:43:28the first two links. The first one takes
1:43:3010% off the server. The second takes $15
1:43:33off your first month of models. Thanks
1:43:36for watching.

About this transcript

This page contains the full transcript of The Ultimate Beginner’s Guide to Hermes Agent by Metics Media, generated from the public captions YouTube serves with the video. The transcript has 20,310 words across 2,847 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.