YT YouTube2Text

YouTube2Text → Transcripts → ایجنت هرمس رو توی کامپیوتر خودش نصب کنیم

ایجنت هرمس رو توی کامپیوتر خودش نصب کنیم — Transcript

by Jadi Mirmirani · 5,940 words · 845 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:01Well, hello. I’m Jadi, I hope
0:03you’re doing well. We’re talking a
0:06bit about AI agents. In the first
0:09lesson, a while ago, I talked about how
0:11to run an LLM on our own computer so we
0:13could connect our agent to it. In the
0:17previous section, I talked about agents
0:19and what an agent is in general. We
0:24realized an agent is basically a
0:26harness or a toolset that wraps around
0:28an LLM—which is just the brain—so I
0:30can work with it; it has features, like
0:32a loop that can bring it closer to a
0:34goal. When you ask it what the
0:39temperature is today, it understands
0:41that it needs to Google it; so it
0:43searches, finds weather sites, searches
0:45within one to find the answer, and if
0:47that site doesn’t work, it searches
0:48the next one. It understands that
0:54it’s in a loop, constantly measuring
0:56its actions against the goal to see if
0:58it has succeeded; when it reaches the
1:00goal or fails for some reason after a
1:02while—for instance, after an hour,
1:04ten minutes, or too many tokens—it
1:06responds to you, but returns with an
1:08error. Since it’s precisely in a loop
1:12, it sees that this site gave an error,
1:14so it tests the next site. I couldn't
1:16connect this way, so I don't have these
1:17programs, so I will install the
1:19programs. And we understood that it is
1:20something like this. Besides this loop,
1:28there are things like the ability to
1:30call tools, make API calls, connect to
1:32the command line or a browser via its
1:34tools, or in agents we see now, perform
1:36a task every day, and along with these,
1:38have a memory; meaning it remembers the
1:40events that have happened. When you
1:48tell it, "I want to remember to call my
1:50mom tomorrow," it says, "Okay, it needs
1:52to be remembered; for this, I have to
1:54write it down somewhere. What is the
1:56date tomorrow?" So it runs a command on
1:58the computer to see what today’s date
2:00is, then writes in its memory that on
2:02such-and-such date, it needs to call
2:04so-and-so; if I ask tomorrow what I
2:05have to do, it checks that and says, "
2:07Oh, you said you should call your mom."
2:09Or I might tell it to check everything
2:11every day at 8 AM and tell me what I
2:13need to do; consequently, it checks
2:15this at 8 AM and does it. This is the
2:18concept of our agent. Now, today we
2:21want to install a common agent. There
2:24are many agents, and as you have
2:26realized, these are essentially
2:27programs built around an LLM. Basically
2:31, I need to install a "harness," that
2:32thing you wear around your body; I
2:34don't know the exact Persian term for
2:36it, the one you strap on when you climb
2:38buildings or mountains or whatever, and
2:43connect it to an LLM. That’s my job.
2:46Which agent should I install? There’s
2:48an agent that’s very trendy these
2:49days; of course, these trends change
2:51every week, don't think that just
2:52because something cooler comes out you
2:54have to switch every time it’s
2:55necessary. Its name is "Hermes," as we
2:59call it, or "Hermis" as they say abroad
3:01, and its slogan is an agent that grows
3:03with you. We still joke with our
3:06friends asking each other, "Hasn't your
3:07agent grown yet? Hasn't it done
3:09anything new?" And things like that,
3:11but there are others too; there are all
3:13sorts of agents. One of the ones that
3:16popularized this, the idea of an agent
3:18fundamentally, was "OpenInterpreter,"
3:20those that run on a computer, have full
3:22access, and do whatever they want.
3:25I’ll go with Hermes for now because
3:27it’s very common these days, has
3:29various capabilities, and I use it
3:31personally too; it’s fun. But you can
3:35install whatever you want; I said
3:36something here, I said it installs on a
3:38computer, and I don’t have any extra
3:40computers around here because I
3:42generally don’t have many; but as we
3:44said, agents can run things like a
3:45command line, they can connect to a
3:47browser, they can connect to a set of
3:49tools; all this happens on your
3:51computer. Do I really want an AI that
3:55might make mistakes—at the end of the
3:57day, it's just a very advanced
3:58auto-correct—to be allowed to run
4:00commands on my computer? Probably not.
4:05If you are very, very, very excited, an
4:11amateur, thrill-seeking, risky; you can
4:13be all of these. I mean, if you did
4:16that, you’re free to, it’s a choice
4:18, but it doesn’t make sense to me. As
4:21a result, many were encouraged to buy a
4:23"Mac Mini" and run agents on that.
4:26Second-hand ones abroad are, say, $ 200
4:29, $ 100; but it’s still expensive.
4:32There’s no need, I run these in a
4:33virtual environment. Back in the day,
4:37we had "VMs," then "Docker" became
4:39popular, and you can run your agent in
4:41Docker; but something I use these days
4:43for this kind of work is called "Incus"
4:46(or "In-cus," depends on your
4:47pronunciation, but "Incus" is probably
4:49the more correct one). It’s much,
4:54much, much more compact than Docker.
4:56You know, you remember the idea of a VM
4:58, right? VMs create a complete virtual
5:02machine on my computer; they have their
5:04own kernel, their own memory, their own
5:06CPU allocation, and their own file
5:08system. You're running a full Linux
5:13system just to run one program, which
5:14is a bit of overkill. Docker came along
5:17to solve this problem. Docker said it
5:23would inherit things like the kernel,
5:25network stack, and many other parts
5:26from the host machine, creating only a
5:28container for the application you're
5:30running. It was an application-level
5:37container, but everything else was
5:38inherited by default from my system.
5:42The idea. LXC was an intermediate idea.
5:49LXC said it would use the kernel from
5:51the host machine. And for everything
5:54else, it would put a full Linux file
5:56system inside. It's a cool idea. It's
6:00something between a VM and Incus, and
6:02now Incus makes managing it much easier
6:04. That's why I'm using Incus for this
6:07right now. If you don't have it, it's
6:12installed on Linux, and you can
6:14logically install it with a simple
6:16command like ` sudo apt install incus`.
6:21I already have it, but it says it wants
6:23to upgrade. Go ahead; an upgrade
6:27doesn't hurt, so it upgrades. Once you
6:30have Incus, you can bring up a machine.
6:32We wanted to install Hermes; Hermes,
6:34Hermis, Hermez, or Hermes; we just call
6:36it Hermes in Persian. One of its
6:39strange features is that its website,
6:41on the very first page with that big
6:42image and slogan, tells you everything
6:44you need for installation. It says you
6:48can install it in the terminal with one
6:50command; just copy this. It takes it
6:52and runs it in bash; clearly, you
6:54shouldn't run something on your
6:55computer if you don't know what it is.
6:57Everyone trusts Hermes. We are also
7:00running it here in a virtual machine.
7:02When I say everyone trusts it, I mean
7:03many people are running it. It's not
7:05too bad. It doesn't mean it's secure,
7:07though. So first, I want an LXC virtual
7:11machine that I run with Incus. Incus is
7:14very straightforward. Incus doesn't
7:20even show the command. You can always
7:22type help. It says these are the things
7:25I can do in Incus: Images, so-and-so,
7:31rename, resume. So why isn't the one we
7:34want in there? Oh, here it is; all the
7:39things it can do. Look, "launch" is
7:42what I want; create, start, instance
7:44from images. See how simple and
7:46straightforward these are. I can say "
7:51incus launch" from images, I want "
7:53debian 13", and name it "my-hermes". It
7:59creates a machine and downloads
8:00whatever is needed from Debian. It's
8:04here because I've used it, but normally
8:05it would need to download like a
8:07gigabyte, and a tool called My Hermis
8:09creates it. It says you have a "My
8:14Hermis" here, for example, a Matrix
8:16backend too, and a "Hermis" that is
8:18stopped; since I can record this, nice!
8:21Okay, so let's connect to this. I'll
8:24say "My"...sorry, execute "exec" on
8:27this, run a command in it. How much
8:29more straightforward it's become than
8:31Docker! My virtual machine is also much
8:33faster and less troublesome; you can.
8:35You can start configuring it to have
8:37this much memory, this much storage—I
8:39don't know another word for memory,
8:41storage I guess—this much CPU, be
8:43like this, mount this file; you can do
8:45a bit of work, but its basic version
8:47started up very straightforwardly. I
8:49say execute in My "Hermis" the "bash"
8:53command. Now I'm connected to it, see,
8:56its name is "My Hermis," my user is "
8:58root," and basically I'm saying run "
9:00bash" in that machine you built. I'm in
9:03that machine now, so my work has become
9:05much more straightforward. It had a
9:08one-line command here that runs this;
9:10let's run this inside it. Be careful
9:13not to get things mixed up here, I'm in
9:15that machine, in "My Hermis," it gets
9:18this file and executes it. It's very
9:21cool. Almost this single line does
9:23everything. The reason Hermis became so
9:26famous was that it had everything
9:27inside it. It installed directly and
9:29effortlessly to Telegram. It installs
9:31everything it needs by itself and is
9:33very hassle-free. Open Cloud and such
9:35are just as simple. But ultimately If
9:39you wanted, you can obviously scroll
9:41down and see what it says. Go check its
9:44documentation. It says it has a web
9:46interface that looks like this. When
9:48you connect to it—I don't know why
9:50it's not here anymore, it's just
9:52loading; what happened? You can go into
9:54its documentation. It's installing here
9:57too; it's saying for itself that I'm
9:59downloading Node. It works with Node;
10:01it said Git wasn't found, it tries to
10:03install Git itself. Node wasn't
10:05installed, it tries to install Node;
10:07let's hope it can install successfully,
10:09sometimes it hits a snag. Yeah, see it
10:11couldn't install Node, it says I want
10:13Node.js but its download gave an error
10:15because, well, the document has
10:16something that doesn't work here. I
10:19installed a "Debian" on my Hermis;
10:21remember I can say apt install nodejs,
10:24and install Node.js myself so that Node
10:26exists. It's a lot like, "AI is here,
10:31machine learning is here," poor machine
10:33learning, no! Agents are here;
10:35programmers don't do anything anymore,
10:37they can't. The most I do with the
10:38terminal these days is keeping these
10:40things alive. Once they advance a bit
10:42more, it'll get better; but for now, we
10:44really need to understand this stuff. I
10:46installed Node.js manually, let me
10:48start the Hermes installation one more
10:51time. It says my Node is too old, it
10:57wants its own version; Yes, it wanted
11:02its own version. Can't I tell it to
11:07install 22? Apt search. Node.js. Apt
11:18update. Okay. What version is my
11:31Node.js? Node.js is 20, we wanted
11:53Node.js 22, what do we do? One way is
12:04to search how to easily install Node.js
12:0722. Install. Debian systems are another
12:20story, you can use something newer,
12:22instead of installing whatever, go run
12:24these commands, and yeah, it actually
12:26goes and sets up 20 and installs it.
12:30Let's run it. I was thinking about
12:33something quite pointless. Not that you
12:35think this way or that it matters to me
12:37, but many might say, "test these
12:39things once before recording a video."
12:41Actually, it's the opposite. These are
12:43the problems you encounter in real life
12:45and must be able to solve if you want
12:46to be a professional. Now it's going to
12:50get Node.js. It installs Node.js 22.
12:55Node version is still on 20. What we
13:00did here is paste the command to inject
13:03the LTS repository, now I want to
13:04install or upgrade it. It upgrades
13:18Node.js, let's see how it goes; these
13:20are things you constantly run into in
13:22real life and must be able to solve. If
13:27you want to do this kind of work, my
13:29idea now is to install a full Hermes on
13:31my compact system using Emacs, which is
13:33actually lightweight. The good thing is
13:38that it sees that environment like a
13:40real Linux, only its kernel is shared
13:41with my computer. As a result, I can
13:44give it plenty of access. Whatever
13:46command it wants to run in the command
13:48line, at worst it only ruins that space
13:50; there's no danger. It's quite funny
13:52when they say, "I don't know, an agent
13:54deleted our production system." The
13:57real answer is, sure it's funny, but if
13:59that really happened and you actually
14:01did that and are truly upset, you’re
14:02an idiot for giving an agent access to
14:04a production system. Let's go once more
14:08to......run the Hermes curl and install
14:13it to see what happens; this time we're
14:15ahead, see, this is how we move forward
14:17. Here, let's also take the opportunity
14:32to learn this stuff. It's a very, very
14:36useful and cool thing. Install it, use
14:40it, try to build a machine, connect its
14:42network. It has very, very, very direct
14:45commands, and they are also very easy.
14:49It has a config index; for example, you
14:51say config, add device to "MyHermes",
14:54name it something like "MyNetwork",
14:56listen in proxy mode to this, and
14:58connect to that. These two links
15:02connect to each other; anyone listening
15:04to that port will listen to this;
15:06it’s very straightforward. Or you can
15:09share your disks; for instance, say
15:11share this partition with that for such
15:13and such reason. Here it says how to
15:17install it and various other things. It
15:19has getting started and quick start
15:21guides; it supports platforms like
15:22Android and even runs on Termux, which
15:24is really cool. The "Using" section
15:27explains everything, how you can
15:29configure it, and the rest of the story
15:31. It actually does everything itself;
15:33you can proceed without reading any
15:35documentation as long as you don't get
15:37stuck. This is a harness, as I said,
15:41the LLM is behind it; its own programs
15:43are running here. You can talk to it
15:49here like ChatGPT or anything else,
15:50write things here and see, but I never
15:52want to sit behind this computer and
15:54start going into this text interface to
15:56talk to it. Therefore, it is very
16:00common in agents to connect it to your
16:01messengers. For example, one of the
16:04good things about Hermes is that it
16:06connects very easily to Telegram. The
16:08good thing is that not all of them
16:09connect to others. Or, for example, it
16:12connects to IRC or Discord, and
16:14eventually, you can interact with it
16:16entirely in Telegram; it sends the
16:18answers there for you, and you work
16:20with it there. There is no need to come
16:22and sit behind this computer. In fact,
16:24you leave this computer on the shelf,
16:32let it move forward a little. My
16:33recommendation is that you don't
16:35connect it to IRC. The first time I
16:38connected it to IRC, my account or
16:40connection was banned on that IRC
16:41server because it was against the rules
16:43to install bots. Fortunately, it opened
16:47up after 24 hours, but if you want to
16:49connect it somewhere to test it. For us
16:53, the most logical place is probably
16:55Telegram in the end. Well, while this
17:06is moving forward, let me say one other
17:08thing. That is, as I have said many
17:11times, you need an LLM in the
17:12background; you can use things like
17:18small models that you can run on your
17:20own computer, run them. Just to see how
17:22it works, learn, and move forward;
17:24don't have the dependency that "I don't
17:26have an account now, I haven't paid
17:28anywhere, I don't have any LLM, I can't
17:30work." You can easily work with this. I
17:32have said this before; and The output
17:37might not be that great, your agent
17:39might do nonsensical things, not know
17:41what to do, be very slow, and so on,
17:43but ultimately you can move forward and
17:45keep learning. But usually, people have
17:47accounts behind them that are connected
17:49to an LLM. And one of the "famous" ones
17:53is OpenRouter. What OpenRouter does is
17:56that it has gone to a lot of places and
18:01gotten APIs to use their LLMs. I mean,
18:04OpenRouter...what happened? OpenRouter
18:07...what happened to Open? Router itself
18:10says: "I am here, OpenRouter." And it
18:15has gotten accounts from many different
18:17places, Anthropic, GPT, DeepSeek, I
18:19don't know, Moonshot, Kimi, and
18:21everywhere else. You create an account
18:25and a token here for yourself and you
18:30can connect it to any of these. The
18:34good thing is that you have one thing
18:36here; you can set it to use whatever
18:37model you're using, put limits on it,
18:39and other things. Another good thing is
18:42that it obscures things a bit; meaning
18:44the request it sends doesn't exactly
18:46say "Jadi" sent this. If you connect
18:48directly, it always sees you; here, it
18:50gets a bit mixed up; it still sees the
18:52content, but it's not directly linked
18:54to the individual. Come here to
18:57OpenRouter, create an account here too;
18:59soon it will ask us for this. I have an
19:04OpenRouter account; you can come into
19:05your personal workspace. Here, first of
19:08all, you can see all the different
19:10models to see what models it has. You
19:16can see what the models are and where
19:17they work. Search based on price and
19:20these kinds of things. In general, they
19:23usually talk based on tokens. It used
19:41to show us the model prices so we could
19:43choose; you choose based on money,
19:44right? Yeah, ah, here's the table. It
19:53tells you here what models there are;
19:55for example, it says DeepSeek V4 Flash
19:57is like this, and weekly tokens might
19:59be limited; a token is roughly a
20:01combination of input and output amounts
20:03. Input token, output token, and
20:06roughly what it generates for you is
20:07somewhat related to the token. If you
20:10pour a lot of things into it and say
20:11read all of this and tell me what's in
20:13it. You've poured a lot of tokens into
20:15it. If it gives you a long answer, it's
20:16providing output. Usually, the output
20:19token is more expensive. DeepSeek isn't
20:21a bad thing. Although today it's
20:23emailing people saying we are soon
20:24going to make this more expensive.
20:27We'll be miserable, but you can use
20:28different things. MiniMax or whatever;
20:30I don't have expertise in which one is
20:32better than the other. And reviews are
20:34constantly coming out. I sort by price;
20:38some of them are free, meaning if you
20:40say sort by pricing, you’ll see some
20:42are free, but you can’t use them
20:44directly there; when you go, it
20:45doesn’t give it to you right now, and
20:47some have limitations; for example,
20:49they say your usage limit is reached.
20:55In short, it’s mostly for viewing and
20:57testing, and often those that are free
20:59keep a lot of your data. You can come
21:02here to Personal, go to your Workspace,
21:04and create a new key here or create a
21:06new workspace. I’ll go into mine and
21:14here I can say I want a new API key,
21:16I’ll name it "Amghazi" as usual, set
21:18an expiration so it stops working after
21:21a day, and you can say consume a
21:23maximum of 4 or 5 dollars. You have to
21:30put a little money in the credit loop
21:32so it can be used later. It gives you
21:34this as "Your New Key." I’ll copy
21:36this; obviously, anyone who has this
21:38can call your API and spend your money,
21:40but since I set it for one day, it will
21:42be gone by the time the video is out.
21:45Let’s go back here; this is still
21:47working. Downloading the headless
21:52Chrome shell essentially allows it to,
21:54like Google Chrome, do things...what do
21:56they call it? Connect to the web.
21:59It’s like running a Google Chrome and
22:01connecting to the web; it’s much
22:03better than "curling" directly. What
22:06else should I tell you? Nothing else;
22:08we’ve already created our keys here.
22:12Hermes has used 4 dollars here so far,
22:14that’s good; 4 dollars of usage is
22:16practically free. I usually don’t use
22:19this anymore, obviously. Well, let’s
22:25let it move forward; I’ll see you
22:27later. It moved forward! It didn’t
22:30want me to see you later, it wanted me
22:31right here. It asks how should I set it
22:33up? It has up, down, and quick setup;
22:35it’s a bit tricky, sometimes it’s
22:37space, sometimes it’s enter. I say I
22:39want a full setup, I go on it, it says
22:41press enter. Full setup means I want to
22:43set up everything. It asks how do you
22:46want to connect to your agent; your
22:48chat model...not agent, your agent? It
22:51has various cool methods that connect
22:53through its own site. I don’t want to
22:56do that; they built the agent and now
22:57they want to make money off me, this
22:59News Portal or whatever its name is. I
23:02say, for example, I connect to
23:04OpenRouter; you can say I want a custom
23:06connection to my own model on my
23:08computer, on my own system. And there
23:12are a bunch of other things too. For
23:14example, if you have Claude, you can
23:15say connect directly to Claude if
23:17you’ve already paid for Claude. I say
23:19I want to connect to OpenRouter. It
23:21asks what is your OpenRouter API key?
23:23I’m saying I just copied it, I’ll
23:25show you too; it’s this one. It says,
23:29how many models can you use there?
23:31Which one do you want to use? I can
23:33even say, oh, it’s listed the prices
23:35here too; good. I can say here that, uh
23:38, OpenAI GPT-55 Pro is 180 dollars, wow
23:41. Those ones are 50 dollars too.
23:46Don’t laugh! He’s laughing, we’re
23:49done for with these guys. Soon
23:51they’ll make all these expensive and
23:52only they will be making money. Instead
23:55of writing a to-do list, you and I are
23:56telling our agent, "I need to call him
23:58tomorrow." Well, look, here it says
24:01space or enter to select; alright. I
24:05chose that free one. It says now if I
24:07want to run a terminal, what should it
24:12be? I say run it directly on this
24:14machine; I put you in a virtual machine
24:16so we'd be comfortable. It says what do
24:18you want to connect to? You can connect
24:20to many different things, like Telegram
24:22, it goes along quite nicely. To Matrix
24:25, a bunch of stuff, Home Assistant;
24:27just be careful not to get banned by
24:29IRC. Email, I can connect you directly
24:31to email and a whole bunch of other
24:33things. Start small, then say you want
24:35other things; Telegram is a reasonable
24:37basic start. It says, do you want it to
24:39be automatic? I say yes, let it be
24:41automatic. It says open this link and
24:44proceed from there. I won’t go
24:48further with this now; how can I stop
24:50this? I hit Ctrl-C and it cut off
24:55completely. The good thing is that
24:58everything is moving along and starting
25:00up now, no danger anymore. I was able
25:01to set the config; why did I actually
25:03do that? You see the previous ones
25:10where it says this is it; I say keep
25:12that same previous one, it says keep.
25:14Were you connected to Nvidia which was
25:16free? I say yes, and it’s very good
25:18too. It says, what should your terminal
25:20backend be? Keep current local;
25:22actually run the commands. Remember,
25:24here we wanted to connect it to
25:25Telegram, which I haven't done; we'll
25:27open it somewhere later. It says scan
25:30this, make a bot; it’s very
25:32straightforward and tells you what to
25:34do; if you couldn't, forget it for now;
25:36I hit skip to cancel. We go to the next
25:40step; it says do you want to configure
25:44the CLI? I say yes, do it. What happens
25:47to Vim? It says, "These are the things
25:49I have now," as it puts it, "what
25:51interfaces should I have?" Video
25:54analysis, image. Create it, generate
25:56video. All of these make calls to that
25:58other one and there’s no problem at
26:00all. Whatever you don't say, it’ll
26:02connect itself again later. When I used
26:05the defaults there, it asked what
26:07browser do you want to connect to? It
26:09had installed a headless Chrome as a
26:11local browser; so, it was like it had
26:13its own browser. This part is a bit
26:15sensitive; those who want to do much
26:17cooler things in the browser usually
26:19use other services. Let it download
26:21whatever it needs for now. Vision
26:24backend, auto; if you give it a photo
26:26and ask what's in it, or turn on the
26:28camera or webcam to see what's there;
26:30though this virtual machine doesn't
26:31have a webcam. Back here, these are all
26:37done. It asks for new things every time
26:40. I had the vision backend, let's go
26:43back here. What happened? Skip, cancel.
26:47I was just wandering around, it's fine.
26:51I hit skip and it canceled, said let's
26:53continue with the rest now. The backend
26:56is nagging, why does this grow with you
26:57? D Skip, cancel, skip, cancel, skip,
27:08cancel. I keep hitting skip, maybe we
27:13can get out, leave me alone. Done. It
27:20gives you reasonable explanations here;
27:22for example, it says if you want to do
27:24these things again later, you type
27:26Hermes. Setup, you enter its setup
27:31again. API keys and settings, config,
27:36you see the configuration, we open the
27:38config editor. You run git install for
27:42some tasks like messaging and so on. We
27:46can try it, it's not bad; Hermes git
27:53install, let's see what happens. Failed
28:05to connect to user scope; well, it says
28:07we don't have this. This time, I’m
28:23growing with you. Yeah, I said let me
28:29review the documentation first; it’s
28:31bad to just copy-paste without reading
28:33it. Definitely look at the
28:36documentation; there’s a quick start
28:38that says what you can do. In Windows,
28:42when we chose the provider, what should
28:44we start from? It gives some
28:46explanations itself and explains where
28:48your settings are; this is important,
28:50all your settings are in/home/hermes. I
28:54personally always keep a copy of that
28:56directory so that if something happens
28:58to this virtual machine, or if a new
29:00container comes along, I can solve my
29:02problem. Or if a more technical
29:04container system comes, I can recover
29:06the data. The better way is that I have
29:08a mount in my home directory too. I
29:11have a mount on my own computer that is
29:13actually writing to my home folder. As
29:16a result, all the data you give to
29:17Hermes goes in there. Hermes is living
29:20in there. So, if you keep that file or
29:23directory and delete Hermes to install
29:25another one, it will have everything
29:26from before. This is an important thing
29:29. And it says when you want to run it,
29:31type "hermes" and you can go into your
29:34CLI or TUI, which is actually
29:35recommended. I wasn't used to it, let's
29:37try it and see what it is. I always
29:39used to type this when it finished. Now
29:41when you type "Hermes", you go into
29:43your chat, which I messed up; let me
29:45type it again (I changed the window
29:47size and it got confused). Here you can
29:49talk to it: "Hello". It thinks for a
29:56bit, then replies: "I am the Hermes
29:58agent, an intelligent AI, back from
30:00that model; we did some pretty
30:02significant work with it." With a
30:04forward slash, you can give it some
30:06commands, including the "quit" command.
30:09It says it’s better these days to
30:11move forward with the "TUI", which is
30:13its so-called "text-based user
30:14interface." I hadn't run this before,
30:17it's my first time, so let's see what
30:19it does. Check your environment, see
30:31where you're running, and get a better
30:33feel for it. They say they've "queued"
30:38this, let's see what it does now. There
30:44is a project that I usually used to
30:46install in the past, called "Hermes Web
30:48UI" or "webgui", this is it. When you
30:55go with this, it essentially gives you
30:57a convenient web interface for the
30:58tasks you want to do with your dear
31:00Hermes. Its installation is also very
31:03easy. If you are in there, it’s
31:05always good to give it a star if you
31:06think it's good; it's a very good habit
31:08. I wasn't logged in on this browser,
31:10but I'll give it one later (I probably
31:12gave it one, just not here). It says
31:16how you can install it; very
31:18straightforward: "git clone" such and
31:21such, "cd", bootstrap it, or run launch
31:23shell, or whatever. But since it's not
31:27part of the main project, I won't
31:28proceed with it. However, I used this
31:30project a lot; "Hermes webgui" a lot,
31:32which means these things are all new.
31:35Meaning, it’s not like I’ve been
31:37using this my whole life. See, here
31:40it’s explaining what I've been doing:
31:42first in the terminal I typed "pwd" to
31:44see what I'm doing, I typed a command,
31:47then I typed "ls", and read a file. It
31:49asks about your "soul", telling it who
31:51you actually are, Soul. A friend had
31:58also said I should explain "md" files;
32:00it's completely valid, they are "
32:02markdown" files, very easy to read, but
32:04they've become very numerous these days
32:06; fortunately, they are good files. "
32:08Soul" asks who you are, user "md". It
32:13says who your Hermes is, user "md" asks
32:16what the user I'm using is. This is how
32:19they say it has memory, has a
32:21personality, and such. Let me test that
32:23question again; look, it went through
32:25everything and said: "I did these tasks
32:27, I typed these commands in the
32:28terminal, I checked my memory, etcetera
32:30and so on." And I said okay, the user
32:33asked me this. At the end it said: "I
32:39checked my environment and I'm very
32:41settled; where I'm living is in this
32:43directory, I'm here, the main files I
32:45found, my soul identity is in here, the
32:47config and scale prompt settings and
32:49other things have allowed me to access
32:51all the tools, have memory storage; I
32:53really feel at home." That's very
32:59interesting. I can start talking to it
33:05about other things here; for example,
33:07tell it to check what it's doing every
33:08day or give it forward-slash commands.
33:13I mean, I can say "test your model,"
33:15and it says: "Oh, my model was with a
33:17mixture of agents, I was connected to
33:19OpenRouter; it was very good, which one
33:21do you want me to use?" I have these
33:26again, I can say "escape" to go back.
33:32You have the agent here, you can start
33:34talking to it right away; for example,
33:36let me say Friend "X is super cool
33:50Record" And for example 450, or all
33:56computer people like 42. I'm
33:59practically telling my agent some
34:01things; now, I said something really
34:03silly here, but the poor thing tried to
34:05keep it in its memory; it said: "User's
34:07friend X is super cool." Long after,
34:10you can ask if you know user X. And How
34:16is the weather today? It will probably
34:19try to check where we are. The
34:21interesting possibility is that it
34:22doesn't know where my physical location
34:24is. Let's see how it can solve this.
34:28Maybe it asks me, maybe not. While it's
34:31doing these things, see, the browser
34:33says it opened weather to see if it has
34:35access to the weather. How did it
34:42figure out? That I'm in Vancouver? This
34:45is an interesting question. S It starts
34:52here, because it can plan and rotate
34:54itself. You can tell it: "Check my
34:56emails." It tells you: "I don't have
34:57access to your email, do you want me to
34:59connect it? I can do that." Open it,
35:01but it depends on how much you trust it
35:03. Aha! It says: "I used weather, it was
35:09able to guess the location based on the
35:11IP I was connected to..." That was
35:13interesting and That's it! You can even
35:19tell it: "Check the weather for me
35:21every morning at 8 o'clock, tell me"
35:22for example, or tell it: "Connect to my
35:28Telegram?" This becomes a collection of
35:32things; now in the next lesson, we'll
35:34try to go a bit deeper and see what it
35:36has that can work interestingly now;
35:38look, it checked its skill lists, It's
35:42looking. It knows something about
35:44Telegram; someone previously wrote that
35:46if you want to connect to Telegram, the
35:48user must open this, a bot is created,
35:50it does this and that. Now it has
35:53become very easy; it says to connect to
35:55Telegram, it needs you to set up the
36:02Telegram gateway; open your Telegram,
36:04chat with BotFather, send "new bot",
36:06and then it gives you something, send
36:08that token to me. It is practically
36:10pushing itself forward. You and I know
36:12we could have moved this forward even
36:14more easily; you and I knew we could
36:15have come out of this environment. Now
36:18I have disconnected Hermes, and Hermes
36:24let me run the setup. The setup is
36:26happening again here. OpenRouter is
36:28fine, keep this. This is fine, keep
36:31this. Here I could have told it again
36:34to connect me to Microsoft Teams or
36:36connect me to Telegram and move forward
36:38. I think I have said almost everything
36:42I wanted to tell you in this part. You
36:45can now have your own agent. If it is
36:48running in Telegram, you can work with
36:49it. You start it. A good test is always
36:52to come out, restart your virtual
36:54machine once, and see if it comes up by
36:56itself or not. If not, tell it to fix
36:58the problem, and there is no need to
37:00sit behind this computer anymore. You
37:02start talking to it on Telegram, and
37:03it's very simple and straightforward.
37:06But let's take a very quick look here
37:08too. Look at the files it has created
37:10here; it has created many files this
37:12quickly. For example, skills, for
37:15example, that soul file I mentioned.
37:21Where is "cd/home/user"? So look at the
37:28soul; it says Hermes agent, whenever it
37:30wants to run, it reads this and
37:31understands its own personality. I
37:36expected it to have created a
37:37subdirectory for me, but it didn't; oh,
37:44excuse me, it created a user, if it
37:46wants to remember something about me,
37:48it remembers it there. Ah, it kept it
38:00in memories; here, for example, it has
38:04remembered that I have a friend named
38:06Xaxis who is super cool, and their
38:08magic number is 42. If you told it
38:10other things too, as I am saying here,
38:12you could even tell it to create a
38:14directory named friends, and every time
38:15I talked about a friend, save it in
38:17there; it can easily do this in its own
38:19file. As a result, you now have your
38:21own agent. The point wasn't just about
38:24running something; the point was that
38:26we have now given the computer to the
38:28computer itself. What interesting
38:30things come to your mind? In the next
38:32lessons, I first want to take a look at
38:34the concept of scale and such, and then
38:36examine the concept of MCP to
38:38understand more technically what MCP is
38:40. You've essentially mixed being very
38:43non-technical with being very technical
38:45. It feels great once you connect this
38:47to your Telegram. Let me know if you
38:48have any problems. I'm setting up my
38:50personal agent on my Telegram. Stay
38:52smiling, goodbye. Yeah.

About this transcript

This page contains the full transcript of ایجنت هرمس رو توی کامپیوتر خودش نصب کنیم by Jadi Mirmirani, generated from the public captions YouTube serves with the video. The transcript has 5,940 words across 845 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.