This content demonstrates a fresh installation and configuration of Hermes Agent on macOS, emphasizing its new computer control capabilities using a private model via the Venice API for enhanced user privacy.
Mind Map
클릭해서 펼치기
클릭해서 인터랙티브 마인드맵 전체 보기
Hermes agent continues to impress the
world and people are
migrating from OpenClaw, the
stars are adding up on GitHub, and Hermes
recently released computer use.
So now Hermes can control your computer.
But which models are you
using to control it and why?
In this video we are going to do a fresh
install of Hermes agent
using a private model with
the Venice API that does not store any of
your data so you can
use Hermes agent with
the privacy intact.
We will set up Hermes agent to manage our
Mac OS computer,
organize a desktop, create
reminders, even create a little chat
widget to interact with from the desktop.
Now remember this is all brand new stuff
so it might be a little
finicky on your machine
but this is a great skill to practice
learning to train your
AI to run your computer.
So in a few months when all this stuff is
normal you will be ahead of the curve.
Let's get started.
So to get started we just copy this
command here, click copy,
and then you're going to
open up a terminal on your machine.
Just search for the
terminal and you can find it.
I'm going to paste that there.
And this installer script
is going to do the rest.
You'll see all this computer looking
stuff show up on your
screen and that is installing
all the dependencies
that Hermes agent needs.
Such as Python, Git, Node, Node.js, etc.
So let that run for a bit and then it's
going to automatically run this
configuration command
the Hermes setup.
We don't have to run that.
It's going to do it on its own.
And here we are.
We'll just run with
the quick setup for now.
And we're going to add
Venice as a provider.
Now by the time you watch this you'll
probably be able to find
Venice here on the list.
But as of right now we
have to add a custom endpoint.
And that is hcpsapi. Venice.ai forward
slash API forward slash v1.
If you want to copy and paste that just
head over to
docs.Venice.ai and click the API
And you'll see just right here for setup
that URL that you can copy.
So we'll submit that.
And now we're going to need an API key.
You can get an API key right here on the
API section of the Venice user interface.
Just click generate new API key.
I'll call it Hermes agent,
generate key and copy that.
and then paste it into the terminal.
If you're new to Venice,
I recommend checking out the Diem token.
Once you stake this token,
you will get recurring $1 credit
in API access to Venice every single day.
This can be very
economically helpful with AI agents.
Next, we get the model list
and we see all the
available models from Venice.
Now, we can also see this
model list here on the docs.
If we just click Models and Text,
we can see all of these same models
and we can see if they're private,
if they are anonymized,
if they are intent encrypted, et cetera.
You can also see the context, length,
and the tools that this model can do.
So for example, Clodopus
4.7 can do function calling,
reasoning, vision,
and it's code optimized.
Gemma 4 Uncensored, however, is
uncensored and it's private
and it's much cheaper than Clodopus 4.7,
but you only can do
function calling and vision.
So ideally you find a model
that's pretty good at
all of these things.
Kimi Cape 2.6 currently
is a strong competitor
at 85 cents per million tokens input
and $4.66 per million tokens output.
It does all of the things,
function calling, reasoning,
vision, and code optimized,
and you get a 256,000
token context limit.
If you compare that price
to GPT 5.5, for example,
you are saving quite a bit.
This is $6.25 and $37.50.
So you're saving
quite a bit of money here
and you're going to have
similar results, believe it or not,
but don't take my word for it.
Try them out, mess
around with different models.
Kimi Cape 2.6 is a popular
one, DeepSeek version four,
as well as Qwen 3.6, as we
see this offers all of these
as well and it's even cheaper than Kimi.
So let's use that one, Qwen 3.6 27B.
If we see this here on the list,
we have that right
there and the number is 13.
So I'll type in 13.
We'll let it auto
detect the token length.
We'll call it Venice.
And now we have our provider set up.
So next we're
selecting the terminal backend.
Now what's neat about Hermes
agent is that you can select
where it will use the terminal
when it needs to run at certain commands.
So you can have it run
locally on our machine,
which is what we're going to choose,
or you can have it open up
an isolated Docker container
to use a terminal.
It can run on a cloud,
it can run on Vercel,
it can even run on a remote machine,
like a VPS or a Raspberry
Pi you have in your house.
So this is a cool options
for more advanced usage.
I'm going to skip the messaging platform
because the purpose of what I
want to use Hermes for right
now is to use it locally on my computer
as privately as possible.
When I use Telegram,
Discord, what's up, for example,
these messages are all stored somewhere.
They're not
necessarily intended encrypted
when I chat with it.
So you're adding a
middleman into the situation.
I want to use Hermes just
straight on my computer,
just with the AI
powered that connects directly
through Venice to the
model with zero data retention,
and that's it.
So we'll skip messaging.
And now it gives us a
summary of what we configured.
And we also see that if
we add some more API keys,
we can get more tools available.
This can be fixed because for example,
with the Venice API key,
we can do image generation,
and that's more configuration you can do.
What is cool about Hermes,
if we launch it now, is
that it is very capable
of working on itself.
So here we go.
It's opened up.
We see its tools.
We see its skills.
Now there's quite a
few tools and skills here
I'm not gonna need.
So a good way to get
started with Hermes Agent
is say, help me clean up the tools,
skills we won't use.
And now we can see it auto detected
the 256,000 token context length.
And we've already used 16,000 tokens
just in that first chat.
And that's because it's
loading all those skills and tools.
So that's why we wanna
optimize this right now.
The default Hermes Agent is very capable,
but it's not
necessarily super token efficient,
which is why we're gonna
start by cleaning that up.
So we get better performance.
You can see here, we
have the Hermes message
and its own little yellow box.
And then we see that it's
looking at the skill file.
It's using a terminal, and
then we get a message back.
And even our model here knows
these are all built in skills.
They blow your context
window and system prompt,
even if you don't use them.
So it wants to know what do we wanna do?
So I'm gonna say, I want
you to be my computer manager.
You're not doing my work for me,
but you're helping me stay organized.
My files, my desktop, et cetera.
Think of me as your
personal digital janitor
and organizational assistant.
Okay, so it's suggesting
now which tool sets to keep
and which to disable.
Now I do agree with most of these.
I do want skills though,
and I do want code execution.
Now you'll see here, computer use.
This, as of the time of your recording,
is new with Hermes.
And that's what we're gonna
play with in a little bit here.
I'm gonna keep skills.
I don't wanna disable
the skills tool sets.
This, I believe, enables
Hermes to create its own skills
when it needs them for
tasks that you delegate to it.
So yes, also keep skills.
And, all right.
And I accidentally
sent that off too early,
but I basically wanted yes
for each of these questions.
So now we're gonna see how Hermes agent,
powered by an open source model,
is gonna optimize itself
with no data retention by the provider.
So now we're getting a nice warning
that the remove command was blocked
because that could be a
security issue, right?
It's on our computer
and it's letting us know
where those directories are.
They're all inside of the .hermes folder.
So yes, I approve.
Now, this could be dangerous in general.
I am installing on a
pretty fresh machine.
There is no sensitive or personal
or just really
important data on this computer,
which is why I'm starting
fresh with Hermes agent here.
But it is recommended you install these
on their own devices,
a Mac mini, a Raspberry Pi,
something that doesn't risk
destroying your whole life
if it accidentally deletes something.
Now, there's no
guarantee that will happen,
but it isn't unheard of.
And on another note, if
you follow along here,
it's just really fascinating
to see an open source model
do this agentic work on
itself in a lot of ways.
The agent is working on itself.
It's just doing such a good job.
These open source models
really have come a long way.
So we went from 5.2
megabytes to 160 kilobytes.
97% smaller.
So we need to restart Hermes,
but first while we're
in this conversation,
let's just tell it to yes,
set up the cron job and the memory entry
because that will be forgotten about
after we restart and
enter a new conversation.
So let's just have it set that up
while it knows what we're working on.
So let's start a new session.
We do control C to close this.
And now we can just type in Hermes
and it will start it again.
And now we can see that
there's many less tools
and skills available.
So say, let's continue your setup
here to be my computer
manager and assistant.
So I can focus on my work
while you keep my machine clean.
And now might be a good time
to activate the computer use tool.
Also want to practice
your computer use tool.
Make a plan for our setup.
Now in general, making
a plan is very helpful
so your agent can stay on track.
Every agent follows plans differently.
And now we can see in the beginning
it loaded the planning skill
as well as the macOS computer use skill.
So it wrote the file for the setup.
So here we have the plan phase one,
audit the current state of the computer,
phase two, practice the
computer use tool, phase three,
establish organization conventions,
like rules we're gonna follow
and how we want this to work.
And then we'll set up
an ongoing workflow.
And we'll also make Hermes
agent load on the computer boot.
So when the computer
starts, it'll automatically load
and I don't have to worry about,
is it running or not?
Is it gonna forget
what it's supposed to do?
So let's answer its questions here.
How aggressive should cleanup be?
Always archive old stuff.
Any folders that are strictly off limits.
For now, no, but potentially yes, later.
Three, want daily
check-ins or weekly report?
Daily check-ins.
Four, organization
style, desktop minimalism,
organize screenshots and
keep folders organized.
What do you suggest?
And now you can see here is
we are all still only at
6% of the context length.
Remember, we were a lot higher than that
before we disable all the skills.
So we've got an efficient Hermes agent
getting started here.
Okay, here's what I
suggest for organization.
We'll organize the
screenshots in the pictures folder.
Oh, I forgot about that folder.
Downloads, anything
older than seven days,
moves to archives down, oh
great, that's a great idea.
But keep screenshots on desktop,
keep downloads archive in downloads.
Archives of folders
should be inside the root.
All right, so those are my preferences.
You might have other
preferences for yourself,
but yes, I like that.
All right, so here it goes.
Didn't require any
more confirmation for me,
just ready to go.
So let's think about this.
So far, we've had Hermes
agent with an open source model,
edit itself, configure
itself, modify its skills and tools,
create a plan, create
an organization strategy.
So it hasn't done anything
yet, but let's take a look.
Here's our desktop,
there's some screenshots.
We'll keep that open up here.
And this is a pretty new computer,
so there's not too much going on.
And we also have some
things in the downloads folder
that can also be removed.
Yes, do those two things.
I will eventually get to a point
where it'll remove the archive folders
because I don't know if we need those,
but at least it'll
keep things clean for now.
Now, we just saw the
downloads folder here.
Now we have an archives
folder and in the desktop as well.
Cool, those are now in
the screenshots folder.
Neat.
Now we're gonna practice computer use.
Oh, I don't even have to tell it.
I'm just gonna go for it.
Let's take a look and see how this goes.
We're not even at 10%
context window, this is great.
And it's working.
I've said this so many
times in this video already,
but don't sleep on
these open source models
that Venice provides.
They're so much cheaper
and they get the job done.
You don't need Opus
and GPT for everything.
So it looks like we
need to reset the session
to load computer use.
And I'll just, yeah.
This starts a fresh session.
Proceed.
Let's continue.
It's checking its memory now
from our most recent conversation.
I do like seeing
everything Hermes-Agen is doing,
which OpenCloud doesn't do by default.
You can actually shut
that off in the configuration
if you don't like it,
especially if you're
using WhatsApp or something,
maybe kind of weird to see all the notes.
So now it's loading
the computer use skill.
So one thing was missing from our setup
and Hermes-Agen took care of that,
installing the Kua driver.
And now we need to grant
it permissions on macOS.
So we take this command here, we copy it,
paste that into the terminal.
And now we're going
to see that Kua driver
would like to record this
computer screen in audio.
So yes, in order to use the computer,
it's going to need these permissions.
Remember, you're giving
your AI a lot of power here.
So make sure you're
doing this on a machine
that doesn't have a lot to lose.
Now we need to quit and reopen it.
And now let's see if it's working with
this second command.
It's to do it again.
The terminal needs to do it now.
Open the terminal back up.
Let's run that command again.
It looks like we need to do
the accessibility permission again.
So a lot of permissions
we got to give here, right?
And we got it, cool.
So now let's go back to Hermes-Agen
and we just have to type
Hermes into the terminal
to launch it.
Let's practice the computer use skill.
Let's see.
Now it has it in its memory
and knows what's been going on.
So it should know
exactly what we're doing here.
And now we can see that
we need to give permission
for Kua driver to use.
It looks like, all right.
Okay, so we're all ready to go,
but there's one quick thing
you may or may not have to do
depending on when you're watching this.
And if news research,
the Hermes-Agen team
has patched this for custom providers.
So over here in my reminders app,
which we're gonna use in just a second,
I have a prompt here.
And this is gonna be below.
You can find this prompt,
but you're going to need to
feed it to your Hermes-Agen.
And what it's basically
saying is to make a patch
in its own code so
that the computer use tool
with screenshots will not
overload the context window.
Because unfortunately,
this is just a bug in the code
that we have to fix.
So first thing we're
gonna send that through
and we can even say, check your code.
Do we need to do this?
And then that way you'll know if it's
already been patched
in the official repository or not.
This type of thing is
actually easier and easier to fix
because the agent itself can do it,
but we have to be aware of it.
And sometimes it can be a little
confusing or off-putting
when it came out of nowhere.
Okay, so it looks like
we already got it fixed.
So we're all good here.
Let's start a new conversation.
And here we are.
We're gonna say, let's
practice the computer use tool
using the reminders app.
And I'll even just mention
the current note in there
is to be provided to the audience
when this video goes live.
Otherwise I want to
create a system of notes
corresponding to my projects folder.
So I have a projects folder.
It'll scan all those projects
and then we'll just
create folders or something.
I've never used the reminders app.
We could also use Obsidian, for example,
to have maybe a little more flexibility
with markdown files.
But for now we'll just go with the
default apps on macOS.
So now it's using the
Apple reminders skill
and it's using the computer use tool.
And I'm keeping the
reminders window open up here
so we can see the changes it makes.
See, we've got the cursor moving.
That's fun.
These kinds of things
can be a little finicky,
but once it learns,
normally it does a good
job after forevermore.
Now you can see this
focus keeps getting stolen.
So a tip here is try
not to keep anything open
that doesn't need to be open,
especially while it's learning.
All right, we've got our reminders, cool.
So now we have lists of reminders.
We'll say yes to all
three of the suggestions.
I also want to create a
folder for video tutorials,
like the one I'm creating now
for Hermes agent computer use.
And we should set up cron jobs to ping me
and remind me as to how they're going
and update the to-do lists.
We should have to do lists per project.
So part of this for me
is really just exploring
how it works and what it does.
I don't generally use these apps
and I've never had an AI assistant
that just controls my computer.
So that would be my recommendation
as you mess around with this.
Just see what it does, see
what works, what doesn't work.
Remember, if you're not
satisfied with the results
with these open source models,
you can always use the more
expensive frontier models,
but you will be sacrificing your privacy
in the trade-off for performance.
So it's done a bunch of things here,
created the reminder lists,
which unfortunately I
can't actually select
while this is running,
which is a little bit annoying.
Okay, so I've got everything
open here to close this off,
but then what I would like to try to do
is add a widget over
here for Hermes agent.
So let's just prompt
Hermes to see if I can do this.
I'd like you to start on
boot every time I log in.
I'd also like there to
be a Hermes agent widget
in the left side of the desktop,
similar to the
calendar and weather widget.
So ideally the widget
there I can just chat with
and I don't have to open the terminal
and I don't have to use a messaging app.
Okay, so it's been working.
I had to install Xcode,
which Hermes agent walked me through,
and now it is creating
the Hermes widget over here
with what's just disappeared.
And it needs Xcode to build this
because that's how you
build basically software
for Apple products.
So I'm gonna stop it real quick.
The widget appears, but
it should be moved down
to not overlap the other widgets.
Can we put a chat input text box here
that automatically sends?
So now it just takes a
little bit of finessing,
but this is what is possible now.
And you could
theoretically create visual widgets
for anything you want
now with an AI agent
on your computer in Xcode.
And then with computer use,
the skill with Hermes agent,
which is only gonna get better.
I will definitely admit this is not
incredible right now,
especially on an open source model.
I have a feeling if you were
to use GPT 5.5 or Opus 4.7,
we'd probably have better results.
But this is also saving a lot of money
in API inference costs.
This is interesting too.
One of the options I
thought for routing the chat input
from here is just to
hit Venice API directly.
So that could be a whole nother project.
I mean, even without Hermes agent,
you could just put a
widget on your desktop
to just access any
model via the Venice API.
If you're interested in
seeing a video where I build that,
drop a comment down
below, because it could be fun.
And now we're just vibe
coding widgets on our desktop.
Okay, here we go.
The gateway is running.
We got the model.
Okay, so as it turns out,
you actually can't type in these widgets.
You can't even move them.
So we're just gonna say
in that case, forget it,
but show me a list of
whatever is planned,
scheduled by Hermes agent,
what it's currently working on,
etc, a dashboard UI with a
button to open up the terminal
with Hermes command to chat.
So it was going to just
create a titled window,
like an actual app window,
but I kind of want to stick
with the widget format here
so it can feel more
like locked into the OS.
And we'll check back in on that.
Okay, so here we go.
We've got our
sessions, our scheduled jobs,
and we really just have
the possibilities now.
We can see what's possible here.
If I close Hermes agent,
let's see, I'll click that button
and it opens up the terminal for me.
I'll have to fix that.
So it automatically opens up Hermes.
That's no problem, but there you go.
And in conclusion, we can
now have a fully private AI
running a really cool agent
framework on our computer,
act as a manager and
assistant on our computer,
and the possibilities are endless.
So I hope this video was inspiring.
Thanks for watching.
Share with us what you're
working on in the Venice Discord.
There's a link below.
And if you want more
tutorials on using AI, AI agents
with a focus on privacy,
hit that subscribe button.
텍스트나 타임스탬프를 클릭하면 동영상의 해당 장면으로 바로 이동합니다
공유:
대부분의 자막은 5초 이내에 준비됩니다
원클릭 복사125개 이상의 언어내용 검색타임스탬프로 이동
YouTube URL 붙여넣기
YouTube 동영상 링크를 입력하면 전체 자막을 가져옵니다
자막 추출 양식
대부분의 자막은 5초 이내에 준비됩니다
Chrome 확장 프로그램 설치
YouTube를 떠나지 않고 자막을 즉시 가져오세요. Chrome 확장 프로그램을 설치하면 동영상 시청 페이지에서 바로 자막에 원클릭으로 접근할 수 있습니다.
