Claude Code Router + Google Gemma 4 31b = 100% Free App Builder (NUTS)
Chapters21
The creator notes Gemma 4 is underrated, pointing out its strong rankings and performance that users may be underestimating.
Gemma 431B plus Claude Code Router proves you can spin up a free, impressive SAS app quickly, with real-world setup live on screen.
Summary
Income stream surfers’ video showcases how a 31B Gemma model paired with Claude Code Router can build a complete SaaS project from start to finish, all with free tooling. The host demonstrates setting up Claude Code Router, confirming Gemma is the active model, and navigating environment variables and deployment steps. He walks through integrating Clerk for authentication and Convex as the hosted backend, then spins up a Next.js app and wires in deployment details. The walkthrough includes practical gotchas like starting the dev server, fixing a sign-in issue, and validating user creation in Convex. Throughout, he segments real-time observations on performance, noting the model’s speed and capabilities compared to traditional low-memory models. He also critiques design capabilities, praises the AI-powered SEO features, and speculates on the model’s potential to imitate outputs from Claude code. The sponsor segment plugs Harbor SEO.ai, touting white-labeling, pricing, and a live results tab. In sum, the video positions a free, Linux-friendly stack as a serious alternative for rapid prototyping and small-scale deployment.
Key Takeaways
- Gemma 431B running inside Claude Code Router can be verified as the active model by checking the {slash} model open router setting and a subsequent sign-in prompt.
- Setting up Clerk with a Convex backend can be accomplished using environment variables for the JWT issuer domain and template, enabling a working user-authenticated dashboard in minutes.
- A basic Next.js app can be scaffolded and connected to Convex for data, functions, and files, creating a live backend with real-time user data flow.
- The presenter demonstrates troubleshooting steps (e.g., a sign-in error and an incorrect useClient reference) to demonstrate practical debugging in a free, local-setup workflow.
- The video highlights an AI-powered SEO engine as a notable feature, drawing a parallel to Claude code’s capabilities and suggesting training data leakage theories while keeping a skeptical stance about originality.
- Harbor SEO.ai is pitched as a white-labeled, revenue-generating tool with tiered pricing (29, 49, 99 euros/month) and a live results tab that tracks content performance.
Who Is This For?
Essential viewing for developers and bootstrappers who want to prototype a SaaS app quickly using free models like Gemma 431B and Claude Code Router, plus integrated auth and backend services.
Notable Quotes
""The official model is only rank 48, rank 38, and rank 28. And then the free one for some reason isn't ranked for anything. I think people are sleeping on this model and they don't quite understand just how good it actually is.""
—Sets up the core premise that Gemma 431B is underrated and powerful.
""We're going to be making a complete SAS project from start to finish using only Gemma 431B and also Claude code inside Claude code router.""
—States the build goal and tools used.
""This is just a completely free AI that you can run on your computer. Not sure people understand how much of a step up this is.""
—Highlights the value proposition of a free, local model.
""Harbor SEO.ai... white labeling in its very best form... you can charge whatever you want from it.""
—Promotes the sponsor as a monetizable tool extension.
""This is // the bones of an entire project done very, very quickly by a completely free setup.""
—Emphasizes the speed and cost-effectiveness of the workflow.
Questions This Video Answers
- how to run Gemma 431B with Claude Code Router step by step
- can I build a SaaS app for free using Gemma 431B and Claude Code Router
- what is Convex and how do you integrate it with Clerk for auth
- what are the limitations of 31B models for SaaS prototyping
- is Harbor SEO.ai suitable for white-labeling a client portal or agency tool
Claude Code RouterGemma 431BConvexClerkNext.jsOpen RouterJWT issuer domainHarbor SEO.aiSaaS prototypingAI code assistants
Full Transcript
Look at this guys, the official model is only rank 48, rank 38, and rank 28. And then the free one for some reason isn't ranked for anything. I think people are sleeping on this model and they don't quite understand just how good it actually is. In today's video, I'm going to try and show you guys how good of a model this really is. We're going to be making a complete SAS project from start to finish using only Gemma 431B and also Claude code inside Claude code router. Let's jump into things. Okay, so on this Claude code chat, I basically have Claude code router and I'm just getting Claude code to set up Claude code router for me.
The reason I do that is because I'm not a dev, obviously. So, I'm just using this to do this for me, right? So, I'm just getting it set up, etc. So, I believe I can just run CCR start here. And then this should load with the correct model. So, let's see. So, yeah, I just run CCR code. Let's go model and this doesn't look right. Okay, so for some reason this model thinks that it's Claude 3.5 Summit, but the fact that it says 3.5 Summit makes me think that it's Gemma. So, I went to go and check open router and it looks like we are using Gemma inside open router.
So, I'm pretty sure we are good to go here. So, just so you guys know, you have to do {slash} model open router and then set the model. Okay, so the final check here is I just let a couple minutes pass. I'm just going to say hi here. And then we're going to refresh. I want to see if there's one at 51. There it is, 151. So, we're confirmed to be using Gemma. I don't know where these 29,000 tokens are coming from. But yeah, let's jump into things. So, I'm just going to do {slash} clear and I'm just going to quit out here and we're going to get started.
So, the first thing we do is we search Convex start Next.js on Google and we just click the first one here and we just press copy here. And we just press paste. Obviously, this is just me doing it myself, right? So, we'll put this inside my app, select Next.js, select select Clerk. While that's doing that, let's just go to clerk.com and start to set up a new project. So, just at the top here, you just want to click on uh here and then create new. And we'll call this Gemma Claude test and then press enter to create.
So, this is just going to create this for us. And you want to go to configure, you want to write JWT right here, and add new template. Convex with a lower C and then just uh change the template to Convex. And then hit save. That's pretty much all we have to do. So, you do need this environment variable. So, let's just go. We just need to save these environment variables somewhere. So, we'll just go to docs.new here. And we'll just save that. That is the JWT issuer domain. And then we'll go to overview. And then copy these other ones here and put them here.
These are just test variables, so don't worry about me showing them. Okay, good. So, from here, we want to do CCR code like that. And then I'm just going to make sure that the model is still set to model to the correct model. So, there we go. That should just do that now. So, we'll just say, "Please set this project up so Clerk is fully set up. When they log in, they are redirected to {slash} dashboard and a user is created in Convex." I will do the .env, just give me the placeholders. I just realized I forgot to start the uh project.
So, wait, I just need to leave here and do CD my app and then npm run dev. I just forgot to do this. This sets up the Convex project. So, create a new project. We'll call this Gemma Claude test. Cloud deployment, US, fine. What this does is it creates the actual Convex account. So, I'll just show you what that looks like in a second. Okay, so this opens up Convex. This is your back end, right? It's a hosted back end. So, you have data, database, functions, which is the back end like edge functions on Superbase, files, just, you know, images, things like that.
Schedules is cron jobs and logs are basically the server logs, so you can see what's going on. Okay, perfect. So, from here, we just want to do CD {dot} {dot} and then we want to do CCR code again and then paste the prompt that we got from before. I don't know why it's giving me this little quick soul thing. I really regret getting this, to be honest with you. Okay, so uh I don't really need plan mode. It's not that complicated. So, don't use plan mode. It's not that complex. So, I just want to see how well this model does with setting up a fairly simple project.
It shouldn't need plan mode, right? Okay, so this is pretty cool. It's actually read the skills. It's already updating, so it's already going. That was pretty fast. What we're going to do is we're going to open up this project inside Antigravity. So, Gemma Claude code like that and then oh, wait, we need to open Antigravity first. The reason I'm using Antigravity is just my IDE of choice. You don't have to use Antigravity. It doesn't really matter what you use. Um so, let's go back here. Oh, no, sorry, here. And go. So, Gemma Claude code. And just drag that over like that.
Beautiful. Now, we want to do the .env here. So, we go back to this page on the Google Doc. Control A, control C. Paste that. This is not an env variable for the client. It's actually for the server, right? So, what we can do is we can go here, go here. CD my app and then oh, CD my app and then npm run dev. This should tell us that the env variable is missing. There it is. So, we can actually open this link like that. And then set the JWT issuer domain like that and press save.
That's now done. Beautiful. So, it says it's fully set up already, but there's an issue. So, let's just fix this issue. I'd be surprised if it was done already, but I mean, I guess it's possible. So, let's go to data. There is a user schema, so that is correct, which is pretty impressive, actually. Okay, and that did actually run. That was extremely fast. So, let's go to localhost. Let's see if it's managed to set up Clerk properly. I'll be surprised if it has, but I mean, it's not impossible, for sure. Doesn't seem to be loading right now.
I think this is the problem. I don't know if this is an actual problem or if it's just a a warning. Okay, so we have this. Let's see if it's done it properly. Nope. Okay, there we go. So, we're now on the sign-in. Let's just sign in here. Let's see if this works, cuz I mean, even Opus messes this up sometimes. So, it'll be pretty interesting to see if this can actually do this first time. Be very, very impressive if so. And yeah, this is like the basic, you know, bones of an entire project done very, very quickly by a completely free setup.
Potentially. Okay, so that didn't work. But that's okay. Uh this is just a um small coding error, it looks like. Yeah, so it's just the use client thing at the top of the file. Happens quite a lot. Okay, so this just says loading. This is the dashboard, I believe. So, I think that has worked. Let's just see if a user was created. No user was created, so that's a mistake. No user was created in the dashboard. Um Also, add a basic dashboard instead of it just saying loading. Okay, so you can see the user is now being created.
It was a little bit slow, but it managed to get there in the end. We now have a dashboard that you can only see when you're authenticated. I mean, that's pretty impressive stuff for a completely free AI that you can run on your computer. I'm not sure people understand how much of a step up this is compared to normal open source low like memory models. Like normally, if you have anything less than 200 bill, then it's just not going to work. This is a 31 billion parameter model. I have to say, guys, whatever Google is cooking over there, they are doing an impressive job, for sure.
Now, I still wouldn't use this over Claude code, obviously, but if I was, you know, not able to purchase a Claude code $200 subscription, this is what I would be doing right now. Probably get myself a fairly cheap laptop that can run the 31 bill model on Ollama and just build. You could build on an on a flight. You can build, I don't know. It's really interesting stuff, guys. You don't have to worry about usage costs, etc. Okay, so yeah, I'm not really impressed with its design capabilities, I have to say. This isn't using any plugins or anything like that.
Yeah, it's not great. What is really interesting is this AI-powered SEO engine thing here. This is exactly like Claude. So, it's very, very interesting. I don't know if this is Claude code doing this, but actually, I did this with open code yesterday, as well, and it did the same thing. So, for me, not accusing anyone of anything, but I reckon they have taken the outputs from Claude and trained this model with those outputs. Overall, guys, I have to say, this is pretty damn good for a model that you can run on most computers. So, I think people need to go and check this out.
Just before we end the video, guys, a huge shout-out to our sponsor, me. This is Harbor SEO.ai, and boy, do I have some amazing ideas for this project. Yesterday, I added the client portal link, which basically means you can send a link from your website to craftdigitalagency.com, not Harbor, and send them Harbor as if it were your tool. This is coming very, very soon. This is white labeling in its very best form. So, basically, you can send this to someone, and you can charge whatever you want from it, and you can say, "Look, I made you this SEO content generator." A little bit like something like GoHighLevel, but, you know, I don't know.
I I I hope that this tool is even better than GoHighLevel. But, if you're just interested in writing, we made the SEO content generator 100 times better using Sonic 4.6. We added new features such as the cluster feature, which allows you to focus on topical authority. The discovery tab is getting better every single day. Trying my best to make this the best content writer and just SEO tool on the market. And it's one of the very few tools that actually has a results tab, where we will show you the results that we're getting. As of today, you can get Harbor SEO.ai in the description and in the pinned comment for these prices right here.
29 bucks a month gets you 25 articles a month. 49 bucks a month, or euros, sorry, gets you 50 articles a month. And 99 euros a month gets you 100 articles a month. I'm super super proud of what we've done here. Something is going crazy right now. Someone's article is getting a lot of clicks per day, I think. 116 pages published, 38,000 impressions, which is 300 impressions per page published, which is crazy. And 457 clicks, so almost four clicks per page published. Someone's content is popping off right now. I don't know whose it is, or where it is, but holy crap, is that getting a lot of content or clicks.
So, guys, go and check out Harbor. There's a link in the description and in the pinned comment. Thank you so much for watching. I think I'll leave the video there. I think I've shown that this model is capable of, you know, doing a basic build very, very quickly. Um and completely for free. So, yeah, thank you, guys. Thanks for watching. If you're watching all the way to the end of the video, you're an absolute legend. I'll see you very soon with some more content. Peace out.
More from Income stream surfers
Get daily recaps from
Income stream surfers
AI-powered summaries delivered to your inbox. Save hours every week while staying fully informed.









