Subscribe
Share
Share
Embed
$32,700 in savings. 10 months of runway. non-technical founder building a real SaaS to $10K/month using only AI tools - or going broke trying.
this playlist is my full journey from $1,940 MRR to $10K MRR, filmed in real time with real numbers. the wins. the fuck-ups and the dollars. no code experience. no safety net. just AI tools and a deadline.
new videos every week. subscribe to see how it ends.
Someone copied this post from me word for word. She even lifted the exact same screenshot. She gave me zero credit and never showed anyone how the Codex trick actually works. I'm the one who built it. So today you get the whole thing, duplicate the Codex app, run any model inside it free. Okay, so here is what you're getting in the next few minutes. You'll have a second copy of Codex sitting in your dock. It looks like Codex, it opens like Codex, except this one runs whichever model you point at it. GLM, Kimi, Quan, DeepSeek, whichever. Or a model running locally on your own machine for free with no API key and nothing leaving your laptop. But why does this even matter? Because right now, if you use Codex, the normal Codex app, you're locked into OpenAI's models. You don't get a choice. Some of these other models are cheaper, some are genuinely better at specific things, and some cost you nothing at all. But the app won't let you near them, at least out of the box. So we're going to change that. On screen, no magic, no trust me, just the actual build. One thing before I start, every command, the repo and the copy paste prompt that wires this whole thing up is free inside my school community, Codex Club. Link is in the description. It's just www.school.com forward slash Codex Club. And I'll point you there at the end so you're not just scribbling notes and watch the build first. Now, here is the thing that a lot of people get wrong, and I want to clear that up before we touch anything. There is a GitHub repo everyone links when this whole idea of the Codex app and putting your own models into it comes up. This guy, Anil, said it in my comments. The repo that is on GitHub is just the logic in the back, not the UI. And he's completely right. The repo is the plumbing. It's not the app. If you clone this repo and stop, you've got a working engine, but like no car around it. So that's the part that people skip. And it's the part that actually makes this feel like codex instead of just a script in a terminal. So let me show you what we're actually doing. We're not copying the codex.app and renaming it. That doesn't work. What we do instead is three pieces. One is that we duplicate the desktop app into its own isolated home directory. So this copy, it's got its own config, its own identity, its own settings, and it's completely separate from the real Codex app. So your normal Codex app stays exactly as it is. Nothing about that changes. This new one just lives in its own little world. Two, and this is the load bearing trick. So pay attention here. We run a small local proxy on your machine. Here's why that matters. So the newer Codex only talks one wire format now. It expects what's called the responses API. That's the only thing it speaks. They speak a different dialect, which is the standard OpenAI chat completions format. And those two don't line up. So as we've got in this diagram, the proxy, it sits in the middle. It listens locally. It takes what Codex says and then translates it into the format that the other models understands, then sends it off and translates the answer back. That's it. So that little translator is the whole reason this works. And without it, the app and the model are talking past each other. And then finally, three, we hand that duplicate app a model catalog and point it at the proxy. And as you can see, I've got the duplicate open right now and I can click on the model picker over here and they're all there. GLM, Kimi, Quen, DeepSeq. I can add other ones if I wanted. I'm running a completely different model catalog inside a Codex looking app and that's the whole experience and that's the part that the repo alone doesn't give you. Now, Ignacio asked, can it be done on Windows? And the answer is yes, absolutely. But I'll be straight with you about how. The duplicated app trick, which is just the second icon in your doc, that's a Mac thing. On Windows and Linux, you don't get the renamed desktop app in this version. So what you get is honestly just as powerful. So you drop your models into a config file and each one has its own profile. Then you launch Codex with codex dash dash profile and the name of the profile you want. profile is a different model and you're ready to go and if you're running a local model through olama you add the dash dash oss flag so codex dash dash profile your local model and then dash just oss and it talks your machine instead of a remote api so it's the same outcome any model inside codex but on mac you get the pretty second app and on windows and linux you get profiles and a flag both completely work don't let anyone tell you that it's mac only it's not it's just the interface is a bit different Now, Henry left this comment, please do make a guide on how to config this. I tried, but failed horribly. And it's probably because he hit one of these three spots. So let me save you the pain. Failure one is you copy the .app and just rename it. And they think this is a copy the icon trick. And if you just duplicate the application file and rename it, it finds home to the same place and nothing changes. A separate config, separate identity, and that's what makes the copy actually independent. Failure two is the wire format. And this is the big one. So people point Codex straight at OpenRouter or Ollama and it just errors out and it looks completely broken and they just give up. And the reason for this is that, remember what we spoke about earlier, Codex only speaks that responses format and those models speak chat completions. They don't match. Remember that step two, which is the translator. And if you skip the proxy, it will fail every single time and it'll look like the whole thing is fake. It's not, you've just missed that bridge part. And then failure three is pointing it at the wrong address. So people wire the app to the model's public URL instead of the local proxy running on their own machine. So the translator is sitting there doing nothing and the app's just ignoring it. The fix is the app points at your local proxy, local host, and the proxy points out to the model. So it's local first and then out. That's it. Isolated home, the proxy, point at the proxy. Nail those three and it just works. Miss any one and it will fail horribly, as Henry said. So now you know exactly where to look if you run into any errors. This guy raised a really important point, and I need to be completely honest with you, and this is obvious, but it's still something you obviously need to bear in mind. Duplicating the Codex app does not change who's paying or where your code goes. So what do I mean by that? Forking the app changes the interface, yes, but it doesn't really change the bill, and it does not change where your prompt lands. If you wire this up to OpenRouter, you are paying OpenRouter per token. So your prompts, your code, the stuff you're working on, that's hitting their servers and it's not free and it's not private. It's a different provider with a different invoice and your data goes to them. Now, the other important thing to note here is that yes, everyone says that open source models are cheaper and obviously it depends on which model you pick. But for the most part, these are the live prices per million tokens on open router versus what the cost of say GPT 5.5 is. However, if you're watching this, chances are you're not actually paying the API price for GPT 5.5 of 5.4 per million input tokens and 30 per million output tokens. It's much more likely you've just got an OpenAI subscription. And if you're watching this, it's probably the $100 or the $200 plan, which means that these open source models, when you're paying for them via open router, they might not always be actually cheaper than what the equivalent of inference is when you've got the heavily subsidized $100 per month plan. So then you might be wondering, all right, well, why even bother doing this in the first place? And this is where I've got two main points. Firstly, GLM 5.2, it's a very, very good model. People are saying it's like Opus 4.7, Opus 4 level intelligence, great open, great motion graphics for videos. So there's that. It also has a 1 million context window, which GPT 5.5 doesn't. the first point and then secondly we come back to the price thing if you're going to use something like deep seek this is ridiculously cheap as you can see here would actually be cheaper than using the heavily subsidized chat gpt plan but even then you're still paying and it's not private so how do you address that when you point this at a local model running on your own machine through olama nothing leaves your laptop There's no API key. There's no per token bill. There's no company seeing your code. It's genuinely free and it's genuinely private path. I can turn my Wi-Fi off. I don't know if I can do that while I'm screen recording with Descript, but you're just gonna have to take my word for it. But the hot butt as to whether or not you actually want to do that because you need some pretty serious hardware. And even then the time of this recording, unless you paid 10 to $20,000 on some serious machinery, you're probably not going to be able to get the greatest formant out of the local models that you can run. But that's a whole topic for another time. But just for example, this is QWEN Coda Next. And this model, I literally just said hey to it. I can hear my MacBook Pros fans whirring. And it took it about a minute to say hello back to me. So I pick on purpose. OpenRouter is more model choice. You pay and your data travels. And then Ollama, it's free. It's private. It runs on your hardware. But the performance is just nothing like what it would be if you're using one of the other options. So that is the whole build, duplicate app, isolated home, the proxy that bridges the wire format, the model catalog, the picker, Mac gets the second icon, Windows and Linux get profiles, free and private on Ollama, paid and cloud on OpenRouter. That's the real version, not the screenshot, the build. As I said at the start of the video, I'm not going to make you reconstruct all that from memory. So this exact repo, every command in order and the copy paste prompt that just sets the whole thing up for you. So you can basically just not worry about everything I just said. It's all free inside of Codex Club, www.school.com forward slash Codex Club. Link is in the description. Join, grab the prompt, paste it. Got your own copy running any model you want. Same thing I just built on screen. No paywall. It's free inside the community. I'll see you inside. Don't be a dropkick. Subscribe to the channel. Peace.