Hi guys, I’ve created an open source MCP server that plugs straight into Codex, which allows you to create in 5 minutes what would normally require 2/3 hours of coding. In addition to that it saves you 80% of your costs, while still getting to use GPT-5.6 
This is Codex itself doing the calculations for me …
What is the best way to use with VS Code and the Codex extension?
Only 95.3? Noob 
Combined with Qoder you get 96.6 with Magic
I actually don’t know. I’ll play with it tomorrow to see if it’s even possible. However, the whole point is that the orchestrator LLM never sees the code (it’s “outsourced” code-gen to my LLM, which actually is better than both 5.6 and Fable on its domain, which is “backend dev stuff”)
So I’m not really sure if it even makes sense for backend dev, since the IDE (VS Code) wouldn’t even be able to show you the code, besides the frontend code of course.
However, there might be some nice combinations for frontend devs, not knowing backend dev, needing backends - So I’ll test this tomorrow …
Thanks Thomas. I was kind of thinking along the same lines but never know until ask
(post deleted by author)
I posted proof of work and proof of billing. The difference is layer depth. Most of these tools optimise one task, one prompt chain, or one workflow. TPE sits underneath that as persistent state infrastructure, so the savings compound across the lifetime of the system. I can still bolt surface optimisers like MCP, Qoder, Magic, or Codex wrappers on top. They are additive layers. They do not replace the persistence layer.
So yes, I stand by it: rookie numbers.
You realise I’ve fine tuned my own LLM, right? Currently, I’m running it on GPT-4.1-mini, generating backend code, in a DSL requiring roughly 15% of Python’s token consumption. This implies I pay roughly 0.15 cents for an HTTP endpoint currently, but I could easily make it drop to zero if I wanted to (only paying some small amount of electricity). Implying you cannot, not even in theory “outperform” me, because everything I do requires 15% of your tokens …
And on success ratio within its domain (domain is “backend development”), it outperforms every single SOTA model that exists in the world - Including GPT-5.6 and Fable …
I’ve already tried fine tuning Gemma 4 (LORA / 26B version), and it performs scarily good there too - Allowing me to install it in my own hw, from my own living room, on a $2,000 computer if I wanted to, resulting in roughly the same price as having a couple of old lightbulbs turned on 24/7 in monthly costs …
A friend of me is using PI combined with some locally running LLM, I think GLM-5.2 or something, having attached Magic’s MCP server to it. He’s running and managing half a dozen clients in his company, exclusively from his own hw, minus the HL generator that he’s connected to his harness being responsible for actually generating code.
His total price currently is literally $0 per month …
Then comes the other parts of the framework, such as being the deployment server too, implying from prompt to production for a 10 API endpoint thing takes about 6 minutes with Fable, 9 minutes with Qoder, 3 minutes with Codex, and 5 minutes with PI (100% “free” solution, zero tokens besides those you’re running on your own hw) …
I’ve asked every single frontier model to calculate the difference in “additional setup time costs”, and both GPT-5.6, Fable, and GLM-5.2 claims these 3 to 9 minutes would become 2/4 hours with “vanilla” models, straight out of the box of either Claude, Codex, or Qoder - Without “deployment tools” of some sort…
Even with “deployment” tools, deployment for Hyperlambda is literally “saving the file”. Deployment with Python, C#, Rust, “whatever”, requires pipelines, compilation, and the whole nine yards.
Time savings are roughly around 95 to 99 percent, which comes in addition to 100% “free” if I run it on my own hw.
Come back when you’ve created 600,000 fine tuning files, in a DSL, that’s 20 times faster than Python, 2x faster than Rust, delivers more functionality for 5 to 20 percent of LOC, that can run on stock hw you can buy at WallMart - And then maybe we can continue debating … 
The screenshots show one of my 600,000 files. The one I attached scrapes a website semantically for specific tags, and inserts values into some database. I’ve got 600,000 of these, covering web scraping, headless browsers, everything HTTP, everything database, everything file system, HTML, etc, etc, etc …
Then comes the fact that my stuff is MIT licensed, 100% free and gratis, and my LLM is even so ridiculously inexpensive on tokens I haven’t even bothered to setup a payment wall for it yet - Implying you can download Magic, connect it to a local LLM, and literally create a thousand backend APIs for literally ZERO dollars …
According to 4 different SOTA models, it’s the single largest DSL dataset that exists on earth, and dwarfs GPT5.6, Fable, and GLM-5.2 on “backend API code”. Salesforce has something similar, but they gave up after 60,000 files. I’ve got 10x as many.
However, if you want to compete with me - Great; Create your own DSL, create half a million fine tuning rows for it, and then come back and let’s have a debate. It’s not even possible for you in theory to outperform me before you do - Simply put because everything I generate is 15% of the size of whatever you generate. Something that’s been scientifically proven by Nvidia, and roughly half of Fortune 500’s research initiatives into the subject …
Like I said, noob!
However, when that’s said, I’d love for you to try to hook up your stuff with mine, to see if it gives further improvements
Magic isn’t “the end of everything”. It connect to whatever you’ve got from before, including Claude, Codex, Qoder, Cursor and Grok Build (these are the ones I’ve tested it with) …
