This Stealth Model Makes Claude Code Free. Here's How.

0:00 Right now, I've got all these different Claude Code tabs running for completely free. Because right here, you can see that the model I'm using isn't Claude, it's Stealth Ox Alpha, but I'm using it

0:08 within the Cloud Code harness. So, today I'm going to show you guys how you actually can set this up so you can use free models inside of Cloud Code. And then I'm going to talk about if it

0:15 actually works and how good the stuff is. I'm going to show you guys everything and share my honest thoughts. So, let's not waste any time and just get straight to this one. So, right now

0:20 I'm inside of Herk 2, which is my AI operating system. It's my second brain with all my business context. And I've got chats running in here for completely free. So, let me show you exactly how

0:29 you actually get this set up. It's super simple. All you need to do is get an open router account and then you just need to change one file inside of cloud code. What you're going to have to do

0:36 here is if you're using cloud code in the desktop app, you can't do this in the desktop app. The reason is because the desktop app overrides these settings and forces you to use a cloud model,

0:44 which honestly makes sense. So, what you're going to do first is you're going to go to open router.ai. If you haven't used this before, it basically lets you have one API key that can route to tons

0:53 of different models. As you can see, there are 500 plus AI models on this site. If I go to the models, you can start to see that we've got certain ones that are paid, like Deepseek, which is

1:01 still very cheap, or this new one that just dropped today, which is what inspired me to make this video, called OX Alpha. As you can see, it's $0 per million input tokens and $0 per million

1:09 output tokens. You can also come in here and you can search for something like free, and you can see how many models in here are free. Now, obviously, this comes with some, you know, fine prints.

1:19 Maybe they've got rate limits, maybe they got a limit per day or whatever it is, but there's a lot of free models in here that you can play around with. And of course, they've got the ones that we

1:25 all love and know. So if I go to for example Claude, you can see that there's Opus 5. You can see that there's Fable 5. And also same thing with like GBT and Gemini. It's got all the models in here.

1:34 So that's step one is to create an account. And then all you're going to have to do is go to your actual account over here. You're going to go to credits. And this is where you need to

1:41 basically grab an API key. So you'll go to API keys and you'll go ahead and create a new key. And in just a sec, we will actually start to use that. Now, just to show you guys that this is

1:49 coming through free. If I go over here to my where is it? my activity. You can see that my top API keys, it'll show this stealth test one, I've used 61.6 million tokens just today by testing

2:01 this out. And if I go to my total spend and I come over here to today, you can see that for OX Alpha, it did cost me about 13 cents total after all of that. But this was after using it pretty much

2:10 all day and running through tons and tons of tokens, testing things out with it. Okay, so you see where to get an API key. You see an open router this model called Ox Alpha. It's from a company

2:18 called Stealth. So, if you want to read more about it, you can click onto here and go to the stealth model terms and you can read about this a little bit. But this is basically an anonymous model

2:27 provider. So, we don't know exactly where it came from. We don't know if it's Chinese or not. So, think about carefully like what data you're putting in here. If it's really sensitive data,

2:34 then you probably don't want to give it to an anonymous AI provider. That's not the point of this video. The point of this video is just showing you how to get it set up and how it works, my

2:42 honest thoughts on it. So, all we have to do is we are going to go into a project. We're going to do this in VS Code. So if you typically are a desktop user for cloud code, you're just going

2:51 to have to do download some sort of IDE or you could run this in the terminal if you want, but if you download VS Code, it's completely free and then you just open up the terminal in here or the

3:00 cloud code extension and you use cloud code, then you will be able to do this. So inside of your cloud, we typically have a settings file. So right here is my settings file and inside the settings

3:10 file, we have uh a section called env. So kind of like environment variables. Now, in here, what you need to put in is this stuff. I'm going to copy and paste this into the description of this

3:21 current YouTube video. You'll grab that. You'll put it right into your environment variables inside of your settings file. And all you're going to have to switch out is your enthropic o

3:30 token. And this is actually just going to be your open router key. So you'll go back into open router. You'll go to your activity like I showed you. You'll go to your API keys right here. Create a new

3:41 key. Copy that. Go back into cloud code and paste that in right there. And then if you want to use the stealth model, which is what I'm showing you right now, you basically just have to copy this and

3:51 put it in all of these slots. So basically the way I found that is back in here when I go to my model, I go to oxalpha. I just copied the name of the model right here, stealthox-alpha.

4:01 Now, let's say you wanted to switch out the model for a different free one. You would basically just have to find the free one you want. So, let's say we wanted to try something like um GLM 5.2.

4:10 We would basically just click on copy right here and then we would go into the settings file and every instance of stealth oxalpha. We would just switch that out for this model GLM 5.2 free.

4:20 You'd save all that and then you'd be good to go. So, I'm going to unsave that. And then when you open up a new session, it should show right here that you're using this model. And it should

4:30 also show down here if you have a status line set up. But that's how you know that it's actually working. And by the way, real quick, Open Router seems to switch out these models that are free

4:37 sometimes. So maybe by the time you're watching this video, the ox stealth alpha model isn't here anymore. It's not free. So what you can do is obviously you can come in here and search free,

4:45 but also Open Router has one called a free models router, which is the simplest way to get free inference because this basically just selects free models at random. basically whichever

4:54 ones are available. So, it shows you some of the models that it will constantly route to. So, if you use this whole method where you use this as the actual model inside of your settings, if

5:02 you go back into there and you were to actually just put open router free inside of all these um places, then it would be constantly every single request just routing to a different free model

5:11 on open router. So, I just wanted to throw that out there in case you're watching this video a while after I made it. So, anyways, let me just show you two quick things that I did once I

5:18 opened this up and show you how this worked. So, I did a /goal. I said, "Look inside my project that lives here." So, I told it to go to a different directory. I told it to create me a

5:26 simple but professional and onbrand landing page. It needs to have all the details of the different products and it needs to be clickable and on brand all that kind of stuff. Now, what happens is

5:35 it does the normal claw code loop, right? Cuz it can acknowledge the goal. It can think, it can search files, it can write, it can use all these tools. But what kept happening was I kept

5:45 getting this API error and it said upstream idle timeout exceeded. So I realized that this was happening because it was trying to like write so much and look through so much. So I told it to

5:53 work in smaller chunks and then it basically ended up being fine and it kept going and it wasn't an issue. But when you are using these free models on open router that tends to happen for

6:02 multiple different reasons. Maybe tons of people are hitting the API, maybe it's trying to do too much. Sometimes that just happens. But anyways, I finally got it to acknowledge why that

6:10 was happening and it kept going and it finished up, right? So it not only did it create the landing page, but it also did verification. So, it did technical verification, functional verification,

6:19 and visual verification, but it also took 6 hours. And I know for a fact this would not have taken that long with regular claude. It's just right now it feels very, very slow. So, anyways, here

6:31 is the actual site. It's not too bad, right? Like, it's got the logo, which is correct. It used the right packaging. So, I told it to use these. I actually wanted it to use the real product shots.

6:40 So, I wanted it to use like these images instead that look a little more real, but I accidentally pointed it to the packaging. So, it used the packaging, which obviously doesn't look like a real

6:47 can, but hey, this is what I told it to do, and that is what it did. So, I'll give it credit there. It also did use our brand guidelines. So, obviously the logo, the color scheme, the spacing, all

6:56 of the stuff, the typography, it used all of this, which was also really good. If I keep scrolling down, you can see that we have a little sliding bar here. We have two drinks, one can, coffee plus

7:05 protein shake equals perk form. All of this information is correct. We have these different products, and we can click into these. So, bold mocha, I can view this. I can buy a single can, a

7:14 12-pack. I can add it to my cart and it actually goes in my cart. It's functional, right? Like we have the roast, the sweetness, the body, we have all this. We can go back to the shop. We

7:23 can go into another flavor. Same exact thing. We can add these to the cart and we can see more facts about all of these. It even made the nutrition facts. So, it's not too bad, right? For front

7:32 end and for digging through all that, it's not too bad at all. Like I said, the issue here was just how long it took because this took 6 hours. And so that's kind of the theme right now that I'm

7:42 finding is if you wanted to build an app, build software, if you wanted to do some deep deep deep technical stuff, these models, these free models, these small local models are not going to do

7:52 it nearly as well as claude or GPT. It's just not. That's the truth. But for knowledge work, it will do stuff for you very well. The way I feel typically is Opus 5 or GPT 5.6 Soul or whatever

8:02 you're using to drive your day-to-day is probably overkill. Like most of the tasks that you're doing on the day-to-day are not that complex. It's just when you really need to think about

8:11 scoping out an entire project or having a model be an orchestrator of tons and tons of dynamic workflows and sub agents. That's where you're really going to feel a big divide where you need a

8:20 more powerful model like an Opus or a GBT. So, let me just show you a few more examples and that will make sense to you. Here's another one that I did a slash goal. I wanted it to go to my

8:28 YouTube channel data and pull a report for me about the past quarter and over the past year. I wanted to also give me a prediction based on my audience, my channel, my niche, what I should be

8:36 prepping for as far as content in 2027. predictions and I wanted this to be a final deliverable as a Google sheet. So, it went through the loop. It did all this stuff. It took This one took also 6

8:46 hours. So, it just took a long time and this would have probably only taken 20, 30 minutes with a regular clawed code. But the actual deliverable is not too bad. So, it was prepared on today's date

8:54 for this channel with this many subs, these sources, a data note. There's different tabs. And when you go through these, it's not bad, right? It was able to look through my data. And also, if

9:03 you'll notice, I'm in a project right now called Freecloud Code. And basically there's nothing in here. So it had to investigate, dig through my other projects and it was able to find that I

9:13 had all of these like scripts and API keys already written. So it found those and it used those to actually access my YouTube data because in here there's not even like a claw.mmd that says who I am,

9:22 right? This is basically just an old claw.md that I set up about Gemma 4 that has nothing to do with what we're doing now. So, I will give it credit for that as well for being able to use the

9:31 harness and use its intelligence to figure out how to actually get the job done, especially because I gave it a SL goal. Anyways, the output's not bad. It found all this stuff. Here was the

9:39 quarterly report. It pulled per month views, watch hours, subs, the average viewers, the top videos, all this kind of stuff. It even went into um different sources. It went into different

9:49 countries, audience, devices, all this kind of stuff. It did the exact same thing for the year. It looked at every single video. So all 4 like69 videos that I've put out on here, maybe 468 and

10:00 then a 2027 outlook. So it did a good job. Like I said, it's not a terrible model for knowledge work. It did a decent analysis. It pulled all the data. It it was a little bit like pragmatic

10:09 and solved its own problems, but it took way too long for something like this. Like arguably, I could have done this manually about the same, maybe even a little quicker. And now let's jump back

10:19 over here. I went to my actual Herk 2 project, and you can see that I said, "Hey, here's a YouTube video. I need to process this for me. I need you to put this skill in the database. So, you

10:28 know, I have this sheet that I have all my resources in. And it was able to look through it, right? It found the skill right away, which means that because it's a model, even though it's a foreign

10:37 model in the cloud code harness, it still knows to look through your cloudmd. It still knows to invoke your claude skills, stuff like that, your memory. So, it found that skill, it used

10:45 it, it did all the right things. But this has been going for way too long. This normally takes normal claude like 5 minutes, if that. Anyways, it found these skills. It wrote this stuff. It

10:54 just worked a lot. And then it said, "Okay, so I stopped trying this retry loop and I'm handing this to you. Everything local is done, but I'm blocked by cloud code infrastructure

11:03 issue." So it got blocked for some reason. But what else I noticed that's kind of cool is that it created like a cron right here. So it tried to create a cron job for itself to basically like

11:11 retry this every I don't know what this is, maybe 10 minutes, maybe 20 minutes. That's kind of cool that it wanted to retry, but ultimately it stopped and said, "Hey, I'm blocked." But it took 44

11:20 minutes to get there, almost 45 minutes to get there. So it's just very slow. Also in the second one, I asked it to go to school. It pulled 166 support needed threads from my community AIS plus and

11:30 it found the three common threads. 30 posts about this, 15 posts about this, 35 posts about this. But this took 25 minutes. Not too bad, but this normally happens way quicker with regular Claude.

11:40 But it was able to invoke my regular skills and do my regular knowledge work. And then you can see over here another instance where this keep is happening. I told it to do some research on oxal

11:48 alpha stealth model and I told it to create me an HTML of what happened. It hit the error. I told it to keep going and it hit the error again. So sometimes it just runs into these issues. But the

11:58 cool thing is the good news is it can invoke skills. It can use cloud codes fetch. It can use cloud codes web search. So really it can do things that work inside the cloud code harness. It's

12:08 just going to be a little bit slower and you're probably going to run into some issues like this. But hey, I honestly do think it's really cool. If you run into a situation where you're out of your

12:15 credits for the day or for the week and you need to keep working on some light things, then chuck in one of these models. Like, you might as well just play around with it a little bit. As

12:22 long as you're being safe about the data you're sending to these models if you don't know where they're coming from or maybe you go to, you know, open router and you switch out the models. Let me

12:29 just show you real quick how that would work. Let's say I wanted to use this one. Like I said, I copy this. I'm going to go back into this one. I'm going to just change out the model right here.

12:37 And I'll only do it once right here. I'll do it here. And then if I save that and then I open up a new terminal right here, it now should have updated. This one should be using ZI's GLM 5.2 free

12:48 with API usage billing. Now, the reason you want to switch all these other ones out too is because if your session invokes different sub aents and it wants to make like a sonnet sub agent, it

12:57 would do that with stealth. Or if it did an open sub agent, it would do that with stealth. So that's why you want to switch out all of those when you want all of them to actually be the same

13:05 model. So anyways, that's how this works. I hope that this opened your guys' eyes and I hope that this all makes sense. Go ahead and give it a try. Let me know if you're able to find some

13:12 good use cases with this type of stuff, but it's definitely worth at least a shot. And if you guys want to see how you can do this type of stuff with completely free local models that you

13:19 would run like right down here on your machine, then let me know. I'd love to make some more videos on that type of stuff as well. But that's going to do it. So, I hope that you learned

13:25 something new or you enjoyed the video. And if you did, please give it a like. It helps me out a ton. And as always, I appreciate you guys making it to the end of the video. And I'll see you on the

13:32 next one. Thanks ass.