How to Use Claude Code for FREE (2026) — Transcript
Full transcript
- 0:00Hey, this is a video on how to use
- 0:01Claude Code completely free. So, Claude
- 0:02Code is the number one coding agent in
- 0:03the world right now, but it also costs
- 0:05between 20 to 200 bucks a month. It has
- 0:07pretty restrictive rate limits, as we
- 0:08know, and some people complain about
- 0:10quality fluctuations. In this video, I'm
- 0:11going to show you a free, fully
- 0:13open-source alternative method called
- 0:14free Claude Code that literally lets you
- 0:16use the exact same Claude Code
- 0:17interface, so the same terminal and the
- 0:19same commands, but with significantly
- 0:21more affordable back-end models that
- 0:23allow you to achieve somewhere between
- 0:2480 to 90% of the quality of Opus 4.7,
- 0:26but for between 2 to 5% of the cost. So,
- 0:29very high leverage. Let me show you how
- 0:30to do it. So, this is an app called
- 0:31Habitual that I just put together, and
- 0:33uh it's just a simple habit tracker app.
- 0:35What's funny is this would have cost me
- 0:37something like 5 to 10 dollars in in
- 0:38Anthropic credits. With Deep Seek Flash
- 0:41V4, which is the model that I was using
- 0:42inside of this free Claude Code uh
- 0:44configuration, I was able to build this
- 0:46for like 3 cents. So, that's not
- 0:48hyperbole. I literally did so for like
- 0:50several hundred times less money than I
- 0:52would pay to Anthropic. And it's not to
- 0:54say that this is the exact same quality
- 0:55of app that I would get if I built, you
- 0:57know, using Claude, but uh you know, the
- 0:59the coding models that are available now
- 1:01on platforms like OpenRouter, Nvidia
- 1:03NIM, and then Ollama, all three sources,
- 1:05which I'm going to run you guys through
- 1:06how to set up on your end, are actually
- 1:08quite powerful, and you can do more or
- 1:09less anything you would do with Claude
- 1:10Code, um you know, with one of those
- 1:12models instead of a traditional Opus
- 1:14back-end.
- 1:15The thing to know is, you know,
- 1:16obviously at really high levels of
- 1:18quality, which I think like the Claude
- 1:19Opus models are getting to, even a 1%
- 1:22improvement in quality might mean like
- 1:23really, really different results. So,
- 1:25I'm cognizant of that. But, I think a
- 1:26universe probably exists where you can
- 1:27fire off a lot of things with like Opus
- 1:294.7 or another really advanced uh kind
- 1:31of frontier model, and then have, you
- 1:33know, Deep Seek, let's say, or G-line
- 1:354.7 do a lot of like the heavy lifting
- 1:37work behind the scenes, the code
- 1:39refactoring, and so on and so forth, uh
- 1:41to save you like 5 to 10x what you what
- 1:42you normally would have spent. So, what
- 1:44I have over here on the left is a
- 1:46actually have Deep Seek running. You can
- 1:47see it's doing this in the exact same
- 1:48Claude Code interface, as per usual. We
- 1:50even have like thinking blocks that are
- 1:52kind of coming out. Um you can actually
- 1:53see that it's referencing Deep Seek V4
- 1:55Flash via proxy, and that's going to be
- 1:57the approach I'm going to show you.
- 1:58But, um for all intents and purposes, I
- 2:00mean, not really much else to say. You
- 2:02talk to it the same way that you do any
- 2:04sort of like Claude Opus model. Let me
- 2:05show you guys how to actually set this
- 2:06up. It's super easy. It takes just a few
- 2:08minutes, and then I can show you guys
- 2:09some examples of a different models and
- 2:11how to how to work with them. So, the
- 2:12way that typically Claude code works is
- 2:14you have it set up here in a terminal.
- 2:16Maybe you have it in Anti-Gravity. Maybe
- 2:17you have it in like the Claude desktop
- 2:19app or something like that. And then you
- 2:20send it something. So, you say, "Hey."
- 2:22And you know, the second that you send
- 2:23it a request, what it'll do is it'll
- 2:25push it over to its API right over here,
- 2:28and then you'll be billed usually based
- 2:29off token usage directly. Or, you know,
- 2:31if you're smart, what you'll do is
- 2:32you'll get a plan like pro or maybe max
- 2:35or whatever. And what these do is they
- 2:37basically subsidize the token usage a
- 2:38little bit in exchange for Claude or
- 2:40Anthropic, the parent company, knowing
- 2:42that you're going to pay a subscription
- 2:43every single month. And that's all well
- 2:44and good. It's just this is a fair
- 2:45amount of money, right? I would much
- 2:47prefer to spend $0.005 per million
- 2:49tokens instead of $5 per million tokens.
- 2:52Now, what our solution is going to do is
- 2:54when I send something over to Claude
- 2:56code, instead of it just going directly
- 2:58to Anthropic, what we're going to do is
- 3:00we're going to force it to instead push
- 3:02to our free Claude code proxy. And uh
- 3:04this is going to be set up as a local
- 3:06server on our computer, hence why it's
- 3:07this localhost, you know, 8082. You can
- 3:09put it on whatever freaking server you
- 3:10want. But, what this proxy will do then
- 3:12is it's then going to route it to one of
- 3:13three possibilities. The first is
- 3:15Nvidia's NIM, which uh is free, you
- 3:17know, with an account. I think you can
- 3:18also pay for some. That basically
- 3:20leverages a bunch of Nvidia GPUs to send
- 3:22requests. These models aren't as good as
- 3:24this next one which costs some money. Uh
- 3:26there's OpenRouter, which uh you know,
- 3:27you can get models down to like 14 cents
- 3:29per million. That's quite a far cry away
- 3:31from $25 per million, right? And then
- 3:33finally, you have Ollama. And Ollama
- 3:35actually uses like your your own GPU.
- 3:37So, you might have a GPU rig set up or I
- 3:39don't know, maybe you just have a gaming
- 3:40laptop or something. You can um set up
- 3:42models to run directly on that
- 3:43infrastructure. And if you're really
- 3:45good, you can actually set them up to
- 3:46run way faster than traditional uh you
- 3:48know, Claude models cuz you're not
- 3:49competing with millions of other people
- 3:50to send those requests. So, what we're
- 3:52going to do right now is I'm going to
- 3:53show you guys how to set up this
- 3:54section, which is the free Claude code
- 3:55proxy, and then after I'm actually just
- 3:57going to run through all three of these
- 3:59options as well as, you know, all of the
- 4:00models sort of on these platforms, so
- 4:03you guys could see everything that you
- 4:04need to do in order to set them up
- 4:05yourself. And you don't need any
- 4:06technical experience setting up local
- 4:08models or anything like that. I'm going
- 4:09to I'm going to walk you guys through it
- 4:10and handhold you from start to finish.
- 4:12Okay, so everything starts with this
- 4:13repo right here called free-claude-code
- 4:15by Ali Sherer. And it looks like this
- 4:18has gained a ton of popularity in a very
- 4:20short period of time. Just scrolling
- 4:21down here to its star history, it was
- 4:23basically a nothing in February and
- 4:24March, and just now it's hit an
- 4:25inflection point. A lot of people have
- 4:26been sharing and stuff like that. What
- 4:28you guys need to know is not about this
- 4:29specific repo. There are a lot of very
- 4:30similar approaches that do basically
- 4:31exactly what this thing does. So,
- 4:33uh should you not want to use this one
- 4:35or want to use a different one or should
- 4:36this become unavailable at any point in
- 4:38time, like just you know, don't The
- 4:40point The purpose of this guide is that
- 4:41you guys should see it in practice and
- 4:43then know how to set it up in your end.
- 4:47Okay, so what we need is we just need
- 4:49this little quick start down here, and
- 4:50I'll walk you guys through everything.
- 4:52Um just open up a terminal. And so, this
- 4:54terminal instance is currently running,
- 4:55you know, deepseek-v4-flash. This is
- 4:57like the one that's actually live. Uh
- 4:58so, what I'm going to do is I'm just
- 4:59going to exit out of this.
- 5:01And then I'm just going to clear. Okay?
- 5:02If you don't know how to do that, let's
- 5:04see if you have another terminal window
- 5:05open. You can also just head over to
- 5:06shell, new tab. Uh if you're running uh
- 5:08Windows, then you're going to have to
- 5:09use like Windows-specific thing called
- 5:11PowerShell, which is a little bit
- 5:12different, but um this quick start guide
- 5:14is going to run you guys through how to
- 5:15do everything regardless. So, because
- 5:17I'm on macOS or Linux, all I'm going to
- 5:18do is I'm just going to copy these three
- 5:20commands in.
- 5:21It's now going to go through and then
- 5:22install everything that it needs to. In
- 5:24my case, I already have all of it
- 5:25installed cuz I've obviously set this up
- 5:26before. Okay, then what I'm going to do
- 5:28is I'll just copy this next command
- 5:29over, give that a paste. So, it's going
- 5:31to clone directly in. Now, uh when I say
- 5:33clone directly in, you'll see this has
- 5:35actually changed, and that's because one
- 5:36of the three commands here was to go
- 5:38into the directory, and so now we're
- 5:39actually in the directory.
- 5:41Uh if you're on PowerShell, you just
- 5:42copy this command over. Um same thing
- 5:44with that.
- 5:45Okay, and so from here on out, we
- 5:46actually have like the folder set up.
- 5:48Like I know it sounds interesting and
- 5:49weird, but we've we've done like 80% of
- 5:51the work. Um just by copying this
- 5:53configuration, we have more or less
- 5:54everything we need aside from just the
- 5:56API key of the provider that we're
- 5:58using.
- 5:59The simplest and easiest provider to get
- 6:01set up with is OpenRouter, so that's
- 6:02what I'm going to show you guys first.
- 6:04And then after that we'll chat NVIDIA
- 6:05NIM and then finally Ollama. Um
- 6:07OpenRouter's probably the simplest like
- 6:08plug-and-play drop-in replacement for
- 6:11whatever you guys are currently using.
- 6:12If you guys have any experience sending
- 6:13API requests or whatever you guys for uh
- 6:15this video, then that'll be very very
- 6:17simple to you guys. And then if not, you
- 6:19know, I'll show you guys how to use even
- 6:20cheaper ones. But basically what you
- 6:21need to do is you just need to go and
- 6:22grab your your your uh OpenRouter uh uh
- 6:25path. So, that's pretty easy. All I need
- 6:27to do is just go to openrouter.ai. Now,
- 6:29I actually already have an account set
- 6:30up, so I'm just going to open up my
- 6:32account. And then uh here are all my
- 6:34requests that I've been making. I'm just
- 6:36going to go to API keys, and then I'm
- 6:38just going to go create. I'll say for
- 6:40YouTube.
- 6:41Then, I don't know, I'm just going to
- 6:43expire this in like a day. That way, if
- 6:44I forget the API key, nobody can uh you
- 6:47know, bill or run me up a million
- 6:49dollars. I'm also going to copy it, and
- 6:50you have to copy right there, otherwise
- 6:52it's going to uh basically exit out.
- 6:54And then what I'm going to do, and uh
- 6:56this may be kind of intimidating if
- 6:57you've never done it before, but I'm
- 6:58just going to find that folder. So,
- 6:59free-cloud-code. And the way that uh you
- 7:01can do this really quickly here is just
- 7:02PWD. So, in my case,
- 7:04users/nicksurvive/hello-world-claude/free-cloud-code.
- 7:06That's pretty straightforward. So, I'm
- 7:07going to go to finder.
- 7:09And uh I already had this open earlier,
- 7:11so I'm going to go hello-world-claude.
- 7:14And once you're done with that, you'll
- 7:16see the folder. Um what you guys will
- 7:18probably notice right off the bat if you
- 7:19just have like a vanilla Mac install on
- 7:21our developers is these files here are
- 7:23sort of like um you know, significantly
- 7:25less opaque. They're sort of shadowed
- 7:27out. And you may have never seen files
- 7:28like that before. What these are, these
- 7:29are hidden files. Anything on a Mac with
- 7:31a period in front of the file name
- 7:33basically just defaults to hidden. This
- 7:34is a convention that they use just so
- 7:36that, you know, if developers build
- 7:38something cool and and maybe potentially
- 7:39a little dangerous, um average, you
- 7:41know, computer users don't really have
- 7:43to worry about that. Um, but what's
- 7:44really cool is you can actually just
- 7:45show or hide hidden files at any time.
- 7:47And I've totally forgotten the hot key,
- 7:48so I'm just going to look up like macOS
- 7:50hidden file hot key.
- 7:53And, uh, right over here you can see
- 7:54it's command shift period. So, I'm going
- 7:56to go back over here, I hold command
- 7:57shift, press period, and boom, it just
- 7:59like removed all the hidden files, and
- 8:01then it's showing them. It's removing
- 8:02them, it's showing them. So, yeah, if
- 8:04you guys ever wonder about how to do
- 8:05that, um, just feel free to to look that
- 8:06up and it'll walk you through it. And
- 8:08the file you want to edit here is this
- 8:10on.env file right here.
- 8:12Okay, as you see, if I zoom way in on
- 8:14this, uh, once you open this up,
- 8:16there'll be placeholders for all of the
- 8:17API keys. Open router config, deepseek
- 8:19config, you know, LM studio config, and
- 8:22so on and so forth. Okay, so I'm going
- 8:24to paste the key in right over here, and
- 8:26then there's one more thing that you
- 8:27have to do. If you scroll all the way
- 8:29down to where it says open router, you
- 8:30also have to specify a model. And, uh,
- 8:32you might be wondering, where do I do
- 8:33that? If you scroll down, there's a
- 8:34little section where it says model,
- 8:36okay, right over here. What you have to
- 8:37do is you have to you have to choose the
- 8:39provider from open router. So,
- 8:40basically, if we want to use, um,
- 8:42deepseek four or something like that, we
- 8:43just go to browse all models.
- 8:45And then I'm just going to type deepseek
- 8:48v4 flash, this one right over here.
- 8:50There'll be this little copy right over
- 8:51here, deepseek {slash} deepseek, um, v4
- 8:54flash, okay. Once you pasted that and go
- 8:56open_router
- 8:58like this, and this should be the
- 9:00format, um, I've done this a couple of
- 9:03times. You see here it's open_router,
- 9:05that's where the provider, then deepseek
- 9:07deepseek v4 flash.
- 9:09If not, we'll figure it out. Okay, I'm
- 9:10going to save this, and once I'm done
- 9:12with that, um, I can actually now open
- 9:13up a terminal right over here, verify
- 9:15that I'm inside of this directory, okay,
- 9:17and we just did the editing on the, um,
- 9:19.env file, which is hidden, so you need
- 9:21to go ls -a if you want to see it.
- 9:24And then from here, we can actually just
- 9:25scroll up to step three, which is start
- 9:27the proxy, paste this in,
- 9:29and I may have this port set up. No, I
- 9:31don't, which is nice, cool. And now this
- 9:32is running. And now we just need to open
- 9:34up another terminal window. So, a lot of
- 9:36terminal windows here, let's just open
- 9:37that up.
- 9:38We should zoom way in, like I did a
- 9:40moment ago.
- 9:41I think that's about equal, right? This
- 9:43one's a little bigger.
- 9:44And then, uh we just run Claude Code
- 9:46with this proxy setup, okay? So, I'm
- 9:48going to copy this bash command, paste
- 9:51this in, and then, go enter.
- 9:54You can actually see we're running
- 9:55DeepSeek-v4-Flash via proxy down here,
- 9:58because we just made that request, and
- 9:59we sort of did that specification. Now,
- 10:01I'm going to go, "Hello."
- 10:02So, when you say, "Hello," um basically,
- 10:04this right over here is just a real-time
- 10:06log, where it's catching every incoming
- 10:08request. It's then sending that incoming
- 10:11request to, you know, DeepSeek or some
- 10:13other provider. DeepSeek is then
- 10:14catching that, and then they're sending
- 10:16it back. And so, you know, on the first
- 10:18initialization, it may take a little
- 10:19while, and that's more or less what
- 10:20happened here. The user just said,
- 10:21"Hello." Simple greeting. I should
- 10:22respond in kind with a concise, friendly
- 10:24greeting. "Hello, how can I help you
- 10:25today?"
- 10:26Now, what's really interesting is, when
- 10:27you ask it, "What model are you?"
- 10:30Because there's so much Claude stuff, um
- 10:33sort of baked into the the system
- 10:35prompt, it doesn't always know. And so,
- 10:37it's saying it's Claude Opus 4.6, the 1M
- 10:39context. When we're running as Claude
- 10:40Code, the CLI agentic interface. Um but
- 10:42what's really interesting is, despite
- 10:43the fact that it's really confident that
- 10:45it is,
- 10:46you know, if you go back to your Open
- 10:47Router, uh sorry, you got to go down to
- 10:49logs,
- 10:50you can actually see the request that we
- 10:52just sent. It's 5:39. I just sent a
- 10:53request over here to DeepSeek-v4-Flash.
- 10:56>> [gasps and sighs]
- 10:56>> And uh you know, you can actually see
- 10:58that the raw JSON down here with said
- 11:00request, as well. Like, this is the same
- 11:02request that we just sent a moment ago.
- 11:05And what's cool is, I mean, this one
- 11:06here,
- 11:07uh let's see, the input token was like
- 11:0930,000, so that's the whole system
- 11:10prompt of Claude Code. And then, the
- 11:12output was 47, which was literally like,
- 11:13"Hello, how can I help you today?"
- 11:15Uh so, we we do, unfortunately, have to
- 11:16feed in the whole system prompt to
- 11:17Claude Code. But I mean, that is
- 11:18literally what Claude Code is, right?
- 11:19It's just part of the deal.
- 11:21But yeah, now you literally have, like,
- 11:22an another model set up here. We're
- 11:24using DeepSeek-v4. Awesome. Build me a
- 11:28simple habit tracking app in a
- 11:32directory, let's call subdirectory
- 11:34called habit tracker.
- 11:37Make it super straightforward local.
- 11:40This is a demo.
- 11:41And I just want to give you guys like an
- 11:42actual live play-by-play. Because, you
- 11:44know, I think local models, well, not
- 11:46local models, sorry, non-frontier
- 11:47models. So, like DeepSeek instead of,
- 11:49you know, one of the big players,
- 11:50Anthropic, Google, you know, OpenAI.
- 11:54You know, like it's not as good and it's
- 11:55probably not going to be as fast. You
- 11:57can actually see the number of tokens
- 11:58that it produces over a time period.
- 12:00You know, like represented down over
- 12:02here and you'll see the speeds vary from
- 12:03time to time. This one over here was 30
- 12:0523.6, this one was 59.4. So, like, you
- 12:08know, it actually does take a little
- 12:10longer to do some things.
- 12:11Counterintuitively, it takes a lot less
- 12:13time to do other things cuz it's like
- 12:14different architecture, right? These
- 12:16models, and you know, DeepSeek in
- 12:17particular, I think is is out of China.
- 12:19These ones are like built with very
- 12:20different architecture in mind and they
- 12:22work a little bit differently. So, it's
- 12:23just worth it to like actually see
- 12:24what's going on and that's what I wanted
- 12:25to do here.
- 12:27What's really cool though is the whole
- 12:27Claude interface works basically the
- 12:29exact same as You hear how we have like
- 12:30the transfiguring? We got the cute
- 12:32little messages. It's clearly doing some
- 12:34thinking.
- 12:35You know, we can press option-enter to
- 12:37send a multi-line message. We can do a
- 12:38lot here. And now you can see that we
- 12:40actually have the app. So, the app lives
- 12:42at nicksari.com/habittracker. So, what
- 12:43I'm actually going to do is I'll just
- 12:44say open it in Chrome.
- 12:47And let's see if it can it can get that.
- 12:50So, just hit open it in Chrome and then
- 12:51boom, we have it right over here.
- 12:52>> [sighs]
- 12:53>> Hey, what's up? So, I'm going to add it
- 12:55and now I'm just going to kind of check
- 12:57my little habits. Boom, just did it.
- 12:59Fantastic and 07. Because of this, there
- 13:01are other strategies that you should use
- 13:02as well. Like, for instance, a common
- 13:04strategy that I like using is I just
- 13:06create a new instance every 50,000
- 13:08tokens or so because I find that quality
- 13:10goes down every 50,000 or so tokens. Uh
- 13:13that, you know, it's just not as smart
- 13:15or as capable. Obviously, you're trading
- 13:16off that intelligence and capability for
- 13:18money. So, what I did here is I just
- 13:19opened up a new uh you know, instance
- 13:21and then I said, "Hey, here's a file.
- 13:23This is way too simple. Make it look
- 13:24sci-quality at a serif font. Make it a
- 13:26much higher brow feel. Uh now it's
- 13:28running. So, as it correctly identified,
- 13:30the JavaScript is pretty clean there. Uh
- 13:32so, what it's going to do instead is
- 13:33just uh redesign both the HTML and the
- 13:35CSS, which is pretty neat. You can
- 13:37actually see the updates occurring
- 13:38basically live. I'm just refreshing this
- 13:40over and over and over again. Uh you
- 13:42know, it's already correctly established
- 13:44a nicer and sexier font, uh which is
- 13:47nice. So, I don't know. I mean, this
- 13:48isn't like an actual habit, right? So,
- 13:49what am I going to put here? I'll say,
- 13:51"20 push-ups." You know, and here I have
- 13:53my little thing. Nice colors, too. And
- 13:56you know, after another 5 or 10 minutes,
- 13:57I'm sure you could see how you could
- 13:58build something like this just as a
- 14:00demo. I should note that I did this
- 14:02entire build in like entirely using
- 14:03DeepSeek, whereas uh as I mentioned, I
- 14:06think probably the best thing to do if
- 14:07you really wanted a good combination of
- 14:08both cost and effectiveness would be to
- 14:10use a smarter model as an orchestrator
- 14:12and then have it send requests to
- 14:13smaller DeepSeek models with more
- 14:15limited context, like let's say 50 or
- 14:16100k. I'm just to ensure that like
- 14:18everybody's doing what everybody knows
- 14:20how to do the best. Okay, so that was
- 14:22OpenRouter. Let me show you guys now how
- 14:23to do NVIDIA NIM. Uh it's pretty
- 14:25straightforward, same idea. Just scroll
- 14:26down here to where you actually have
- 14:28your uh NVIDIA NIM API key section,
- 14:30model section. And then what you have to
- 14:31do is actually have to go sign up to
- 14:33NVIDIA's NIM platform. Now, in my case,
- 14:36I think I'm probably already signed up.
- 14:38May May May already have me signed in
- 14:40here. So, let me see. Yeah, it did. Um
- 14:42what I'll do, just to show you guys sort
- 14:43of what that process looks like, is just
- 14:44sign out. Okay, now all I have to do is
- 14:46sign in with my email address, which I
- 14:48should be able to do here. The sign-up
- 14:50flow is probably going to ask you to put
- 14:51your phone number in your email address,
- 14:53all that stuff. Um I'm just going to
- 14:54assume that you guys have phone numbers
- 14:55and are willing to do that. I also just
- 14:57don't want to share mine. And then from
- 14:58here, just click generate API key, and
- 15:00then I'll say "For YouTube."
- 15:02Okay, and then the expiration here,
- 15:04again, I'm just going to do I don't
- 15:05know. Why don't we do 12 hours?
- 15:08Then click generate key, and now you
- 15:09have the actual API key right here,
- 15:10which is awesome. Next thing you got to
- 15:11do is go to models really quickly. Um
- 15:13when you click on models, assuming that
- 15:14NVIDIA actually loads. Holy moly, that's
- 15:17slow. Um anyway, it'll open up this
- 15:18little models page, and then you can
- 15:19actually select whatever specific one
- 15:21you want.
- 15:22So, as you can see, you know, DeepSeek
- 15:24over here, so is GLM 5.1, GLM 4.7 right
- 15:28now is free, so it doesn't even cost you
- 15:29any money. That's what I'm going to be
- 15:31using just because I think that's
- 15:33probably like the simplest one
- 15:35to get set up with. And then, uh, just
- 15:37because this is taking a while, I'm just
- 15:39going to go Z A I / GLM
- 15:41-4.7. Okay, now we're actually in a
- 15:44little test page and we can actually
- 15:45have a conversation with it. We can
- 15:46actually say, "Hey."
- 15:48Um, this is now going to send the
- 15:49request to the model. The model's going
- 15:50to go through its whole thinking,
- 15:52uh,
- 15:53whatever, and then, you know, so I guess
- 15:55we just verify that that works. Not
- 15:56super relevant for us. All we have to do
- 15:58now is if you go back to this quick
- 16:00start, you'll see that it says
- 16:00nvidia_nim/z-ai/glm4.7.
- 16:04So, I think we actually probably already
- 16:05have this set up. I'm just going to
- 16:06double-check. I just go back to finder
- 16:08where you had that .env. Okay. Then from
- 16:11here, um, just scroll down until you
- 16:13have the NVIDIA API key, which we do.
- 16:15So, I'm just going to paste in my NVIDIA
- 16:17API key. Zoom in a bit for you guys.
- 16:19Then scroll down to again where it says
- 16:21model, and this is going to be where,
- 16:23uh, you'll put in the NVIDIA one, the
- 16:25one that we just had a moment ago. So, I
- 16:26think that's probably the one that we'll
- 16:28do. I'll use NVIDIA NIMs GLM 4.7 because
- 16:32it's free, and, you know, I want to show
- 16:33you guys how to do a free model.
- 16:35We'll paste that in. And now I'm just
- 16:36going to hold command and then press S.
- 16:39That's going to save it. Uh, you can you
- 16:41can see there are a lot of other, you
- 16:42know, different types of models and so
- 16:43on and so forth. And now what we can do,
- 16:45you know, this is currently actually
- 16:47doing a build right now, um, but as you
- 16:48can see, the tokens have stalled. And
- 16:50the reason why the tokens have stalled
- 16:51is cuz I just like changed the the file.
- 16:53What I What I'm going to do is I'm just
- 16:54going to delete both of these, okay? And
- 16:56I'm doing that by holding control and
- 16:57then pressing, uh, C.
- 17:00I'm going to clear these out.
- 17:01>> [sighs]
- 17:01>> And now I'm just going to run the server
- 17:03again. That's 8082. And then over here,
- 17:06I'm just going to go Anthropic off
- 17:07token, Anthropic base URL
- 17:08localhost:8082/cloud.
- 17:11Okay, so now I'm going to go, "Hey."
- 17:13It looks like the reason we're getting
- 17:15this is speed. I think that's because I
- 17:17have fast mode on, so let me just turn
- 17:18that off.
- 17:20Okay, cool. So, we just went hey, and
- 17:21then it's sending a request over to
- 17:22Nvidia now.
- 17:25So, now it's saying, "Hey, how can I
- 17:26help you today?" Fantastic. What you'll
- 17:28find is if you're using things that
- 17:30aren't supported by these models, like
- 17:31for instance, these models don't have
- 17:32fast mode, even though that's sort of
- 17:33like enabled within Claude Code, you
- 17:35send the request, you'll receive an API
- 17:37error being like, "Hey, what the hell's
- 17:38a speed thing? I don't have that." So, I
- 17:40just make sure you turn any of that any
- 17:42and all of that stuff off.
- 17:44Uh and then I'll say, "Hey, what's your
- 17:45working directory?" Just cuz I want to
- 17:47double-check and see what it's at.
- 17:50And uh you'll also notice that this is
- 17:51still broken. It says deep sea before
- 17:53flash, which is kind of funny. Uh this
- 17:55one says hello world Claude free Claude
- 17:57Code, awesome.
- 17:59Um you know,
- 18:02"How do we change the status line?"
- 18:05Let's just ask it that. Maybe you can
- 18:07actually do that. So, they can see here
- 18:09that that response took a little while,
- 18:10and then it asked about changing the
- 18:12status line. Um this thinking is
- 18:14occurring here in just one giant block
- 18:16as opposed to streaming, which is
- 18:17interesting. So, what it's doing now is
- 18:19it's spawning the status line setup
- 18:21agent, okay, which is something that's
- 18:22actually built into Claude Code,
- 18:24interestingly enough.
- 18:25But again, we're using a different model
- 18:27brain
- 18:28to power it. And so, that's what's
- 18:30really cool about this harness, like
- 18:31it's interoperable, and you can swap
- 18:32whatever other agents are inside of this
- 18:34container. Okay, and here it's now
- 18:35asking me what the model name is. For
- 18:37model name, I'm just going to say use
- 18:38GLM 4.7 for model name, and then
- 18:42everything else the same.
- 18:44And let's see if it can actually go do
- 18:45it, if it has enough context and
- 18:47intelligence to do it natively on its
- 18:48own. Okay, and then the last thing we
- 18:50need to do is we need to set up Ollama.
- 18:51So, Ollama's actually pretty
- 18:52straightforward. Just head over to
- 18:53ollama.com, and then there'll actually
- 18:55be a one-liner that you can just paste
- 18:56into your terminal. You can also
- 18:58download it through like more
- 18:59conventional means, um but I'm just
- 19:00going to use the shell, cuz that's easy
- 19:02for me.
- 19:03These guys are pretty cute. They're a
- 19:04lot more visual, so all their stuff will
- 19:05be really adorable. Once you're done,
- 19:07just type in Ollama,
- 19:09and it'll actually like get Ollama up
- 19:11and running on your computer. So, as you
- 19:12can see here, there's now chat with a
- 19:13model, launch open claw, launch claw to
- 19:15code, launch open code more, right?
- 19:17There's a lot more down here. Whatever
- 19:19files you want. Um just for examples,
- 19:21I'll go chat with a model, and then what
- 19:23do we want? I'm going to really simple
- 19:24one.
- 19:25Uh you can see that a couple of these
- 19:26are cloud, right? So, this one's
- 19:27actually in the cloud. That one's in the
- 19:28cloud. That one's in the cloud. That
- 19:29one's in the cloud.
- 19:30But, I want something like really,
- 19:31really, really basic and really simple.
- 19:33I think I'll probably just use Gemma 4
- 19:35because this one's 12 gigs, and I can
- 19:37probably download it through the video.
- 19:38So, I'll just click Gemma 4, download
- 19:39Gemma 4.
- 19:40Now, it's going to pull the manifest,
- 19:41and actually have to do the the
- 19:42download. Now, my internet's pretty
- 19:43fast, so I should be able to download,
- 19:45you know, this whole thing, which says
- 19:47it's around 10 GB or so in a couple of
- 19:49minutes. I'm just going to circle back
- 19:50when it's done. We'll also go back here
- 19:52to models, and then what we wanted was
- 19:56Oh jeez, what did we just download? I
- 19:58think it was Okay, awesome. So, we
- 20:00actually have Gemma 4. And the reason
- 20:01why I'm doing this is because I just
- 20:02want to get the uh the code. And it
- 20:03looks like the code is just Gemma 4
- 20:05itself, all lowercase. So, we should be
- 20:06able to do that pretty straightforwardly
- 20:08when it's done. K, and then I think
- 20:11we're now actually talking to it. So,
- 20:12I'll just say, "Hey."
- 20:14And yeah, you can see the thinking
- 20:15process is now being written. Uh what's
- 20:17cool is this is now entirely local on my
- 20:18computer. Um are you
- 20:21on the cloud or what? I don't actually
- 20:23know if it's going to know.
- 20:24>> [gasps]
- 20:25>> It's asking a philosophical question.
- 20:26Yeah, no, it clearly doesn't actually
- 20:28know where it's running. Um that's
- 20:29pretty funny. Anyway, regardless, uh we
- 20:31have basically everything that we we
- 20:32need set up, which is cool. So, I'm just
- 20:34actually just going to exit out of this.
- 20:35I think it's control D.
- 20:37K, once you're done with this, I'll just
- 20:38click control C. [snorts]
- 20:40K, and I'll clear it. Okay, and then
- 20:42once we're done, um we just need to
- 20:43write Ollama pull, and then it'll pull
- 20:45Gemma 4.
- 20:47That will actually like go and grab
- 20:48Gemma 4. I mean, in my case, I already
- 20:50have it on my computer cuz we just
- 20:51downloaded it. But, uh from there, we
- 20:52just go Ollama serve. Uh and I think the
- 20:55reason why it's served is it I think it
- 20:57like auto serves um when you do that
- 20:59now. It might not have used to do that.
- 21:01Alternatively, we might have just
- 21:02already been running it, which is great.
- 21:04Um that means we can now go through and
- 21:05then update the last thing that needs
- 21:07updating and then open up the file and
- 21:09then look for Ollama. You can see down
- 21:11here it says Ollama config local
- 21:12provider no API key provided. Then it
- 21:14gives me the base URL. And if you guys
- 21:16notice the base URL localhost colon
- 21:1811434 is the same as the 11434 here. And
- 21:21the localhost, another way to say that
- 21:23is just 12700
- 21:25one. So what this is saying is like,
- 21:26"Hey, you're currently running the
- 21:28server for Ollama on this."
- 21:30Which is obviously awesome to us. Okay,
- 21:31and then the last thing we have to do is
- 21:32we have to actually grab
- 21:34um whatever this is, Gemma 4 latest.
- 21:37Okay, and then copy that over and then
- 21:39paste it in right here.
- 21:41And we always have to prepend with
- 21:42Ollama if it's if it's listed up here.
- 21:44So Ollama and then Gemma 4 colon latest.
- 21:47Save that and then now we can actually
- 21:48go and run the app again. So we just ran
- 21:50the app and then uh I think it's over
- 21:53here.
- 21:54So we connect to that. And now we should
- 21:56be having a conversation with um Gemma
- 21:594. So I'll say, "Hey." Let's see how it
- 22:00goes.
- 22:01You can see this is sent I think to a
- 22:03different endpoint. So this tells me
- 22:05it's probably right. What's really
- 22:06interesting is I just sent that and my
- 22:08computer just started getting really
- 22:10hot. The reason why is because all of
- 22:11this was occurring locally on my on my
- 22:14own computer. I can actually hear it
- 22:15humming now, which is really funny.
- 22:18Um that's because it's literally passing
- 22:20through that request and doing all the
- 22:21transformer multiplication uh matrix you
- 22:24know inversions or whatever the hell,
- 22:25like literally on my own computer, which
- 22:28is really cool to me.
- 22:29And it's now dying down because it's not
- 22:32actually doing the the run anymore.
- 22:34So that took a fair amount of time just
- 22:35to get this response, right? The reason
- 22:37why is because when you run on local
- 22:38models like Ollama, um it's it's really
- 22:40slow like to be honest because it's
- 22:41running on your own hardware, right? Uh
- 22:43it's differently if you're running on
- 22:44like dedicated GPUs like with OpenRouter
- 22:46or um even like Invidia NM.
- 22:48But here like I I am literally running
- 22:49on my MacBook and my MacBook is not
- 22:51optimized for this at all. So hopefully
- 22:52I'm still capable of recording without
- 22:54uh skipping frames or whatever, but
- 22:56yeah, that's uh that's how you run it
- 22:57locally on your on your computer as
- 22:58well.
- 22:59And I should note that this is a big
- 23:00model, you know? it's like uh you know,
- 23:0210 GB with 128k contacts running
- 23:05locally. You know, if you don't have a
- 23:06pretty good computer, you'll probably
- 23:08have to use a dumber model. So, um I
- 23:10don't know, you can go newest here to
- 23:11sort by like the the good ones, but a
- 23:13pretty reliable one was the Llama series
- 23:15of models for a while. So, uh you know,
- 23:17if you're running something a little bit
- 23:18smaller, maybe a Llama Let me see, Llama
- 23:203.1, 3.2.
- 23:23These ones are okay.
- 23:24Uh this guy said it was Mythos, which is
- 23:26hilarious.
- 23:28You'll also see that uh models on Ollama
- 23:30are from different people.
- 23:32This is by Siddiqui
- 23:33uh Siddiqui Nashra, which is pretty
- 23:35interesting. And so, typically what
- 23:36they'll do is they'll like modify the
- 23:38model a little bit, they'll change it,
- 23:40they'll fine-tune it, they'll just make
- 23:41it make it different. So, feel free to
- 23:43take a look see through Ollama's um
- 23:46database of models and see if there's
- 23:47anything maybe that like works for you.
- 23:49A payroll-specific model,
- 23:50accounting-specific model, and and what
- 23:52have you. One more thing that I'll
- 23:53mention is um earlier I talked about
- 23:56having Claude orchestrate conversations
- 23:58between Deep Seek and then you know, a
- 24:00smarter model in order to do stuff. And
- 24:02so, what I've done here, if I just get
- 24:03my ugly head out of the way,
- 24:05is um is I've done that. I've uh
- 24:07basically said, "Hey, send a message to
- 24:09Deep Seek using the flow we created
- 24:10earlier, then tell me what it says
- 24:12back." And so, that's what it's doing
- 24:14right now. It's actually basically doing
- 24:15all this hosting stuff totally on its
- 24:17own.
- 24:18And as you can see, when you just send a
- 24:19message directly to Deep Seek, it says,
- 24:20"I am Deep Seek, the latest version of
- 24:22the AI model developed by Deep Seek
- 24:23company."
- 24:24Um you know, okay, great. Orchestrate it
- 24:26to build an app, guide it, and suggest
- 24:29it
- 24:30through the steps. App description,
- 24:34I don't know, simple calorie tracker,
- 24:37but MVP needs to appear super luxe and
- 24:40high-end for the purposes of a demo.
- 24:44And um you know, what's happening right
- 24:45now is like my much smarter model, Opus
- 24:484.6 in this case. I know it's saying GLM
- 24:504.7 down here, but that's because we
- 24:51actually just updated the status line
- 24:53manually, which you shouldn't probably
- 24:54do, by By way. I I hardcoded in GLM 4.7
- 24:57just as proof of concept.
- 24:59Uh but we're doing is we're using this
- 25:00much smarter Opus model to basically
- 25:01communicate with DeepSeek through the
- 25:04proxy server that we just set up. So,
- 25:06it's actually using, you know, Claude
- 25:08code under the hood, and then we're
- 25:08going back and forth. And as you can
- 25:10see, our simple prompt here was
- 25:13understood, digested, and then spat out
- 25:15by Opus. So, it's a lot more complex,
- 25:17and actually has some opinionated
- 25:18baked-in font choices and
- 25:20colors and stuff like that. And um it's
- 25:22just sending all of this over to
- 25:24DeepSeek right now. And it's just going
- 25:26to go back and forth with DeepSeek until
- 25:27we have, you know, an okay-looking app.
- 25:29And as you guys can see, the
- 25:30orchestration sort of approach put
- 25:31together this app, which is kind of
- 25:32neat. Uh this is just a a base calorie
- 25:34tracker, but still looks pretty cool.
- 25:36Obviously, some minor issues here, like
- 25:38the spacing. I think if I build in a
- 25:40better orchestration loop, like I had it
- 25:42do screenshots and stuff, it probably
- 25:43would have sorted that out. It even has
- 25:45some like quick-add calorie stuff, which
- 25:47is wild. Then you actually have the the
- 25:48log here as well. So, you can get, you
- 25:50know, arbitrarily complex, more or less
- 25:52as complex as you want. Looks like it
- 25:53even did a little calorie tracker app,
- 25:55which is awesome.
- 25:56So, that's pretty cool. And um you know,
- 25:58when Anthropic ran the same sort of
- 26:00infrastructure, uh I think pairing Opus
- 26:024.6 with I think Sonnet 4.6, or maybe it
- 26:04was Opus 4.5 with Sonnet 4.5 sub-agents,
- 26:07cuz this is basically what we're doing,
- 26:08a sub-agent flow. Uh I think it
- 26:09performed like, I don't know, 15% better
- 26:12or something like that.
- 26:13So, what we're doing now is we're not
- 26:14just using sub-agents made by the same
- 26:16company. We're using sub-agents that are
- 26:17hundreds of times cheaper to do a lot of
- 26:19our intellectual heavy lifting. And then
- 26:20the Opus series is just going to take a
- 26:22look at it and be like, "That's good."
- 26:23or "That's bad." which is obviously a
- 26:25lot more profitable. Okay, so hopefully
- 26:26you guys appreciated that video. I had a
- 26:27lot of fun putting it together. Uh costs
- 26:29were a major concern of mine up until
- 26:30just a few years ago when I started, uh
- 26:32you know, making a fair amount of money
- 26:33with this stuff. So, obviously this sort
- 26:35of intelligence should be equally
- 26:37distributed, but for the near term, I
- 26:39doubt that it will be, simply because,
- 26:40you know, uh model companies got to eat,
- 26:43I guess. GPUs got to go burr.
- 26:45If you guys like this sort of thing, and
- 26:46you want to see how to maybe make some
- 26:48money as opposed to try and squeeze your
- 26:49wallet as much as you possibly can,
- 26:51definitely check out Maker School, it's
- 26:52my day-by-day accountability program
- 26:54where I basically show people how to
- 26:55take their skills with Cloud Code and
- 26:57then other no-code automation platforms
- 26:59as well like N8n and so on and so forth.
- 27:01And then you use that to sell AI agents
- 27:03and full workflow automation to
- 27:05companies to make a living. We have a
- 27:0790-day guarantee where you will get your
- 27:09first client within 90 days or I will
- 27:10pay you all of the money that you spent
- 27:11on my program back. It's been run
- 27:14through by over 10,000 people to date
- 27:16and I'm not going to sit here
- 27:17advertising any more than that but it's
- 27:19pretty cool. There's basically no
- 27:20downside to signing up. Aside from that,
- 27:22if you could do me a big solid, just
- 27:23like and subscribe to the video. Every
- 27:25little bit of engagement helps and I'm
- 27:26starting a big YouTube run here so
- 27:27that'd be quite valuable. And yeah, let
- 27:30me know what you guys want to see next
- 27:31because I make videos like this
- 27:32basically entirely based off of your
- 27:33requests and recommendations. This one
- 27:35was made directly by a request that I
- 27:37received a couple of days ago. Have a
- 27:38lovely rest of the day and I'll catch
- 27:40all y'all in the next video.
About this transcript
This page contains the full transcript of How to Use Claude Code for FREE (2026) by Nick Saraev, generated from the public captions YouTube serves with the video. The transcript has 6,843 words across 983 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 choose a caption track available on the video.
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, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.