← Back

Launching codelazr

I made codelazr because I wanted an alternative to expensive AI coding websites. Maybe you can relate! I wanted to see if "lower-tier," "open" AI models from OpenRouter (with some nudging) could produce results similar to "frontier" coding systems and models. I wanted to feel more "secure" in my coding pipeline - more diversified (so I'm not relying on just 1 AI provider, with all my eggs in 1 basket). So I made this coding/chat UI first. Then I started testing some different models.

The results at first were quite abysmal (as expected)! Our initial testing, the models on average only completed 54% of tasks successfully (or even at all!). Then we started tweaking, optimizing, and adding little tricks here and there in the code. We made Plugins to catch edge cases. We added Commit checks that call a heavy-hitting AI to review code diffs before they're proposed for commit. And after all that testing (over 900 runs so far), we got our 54% task completion up to 96%!

But a finished task isn't the same as a correct one. So I kept pushing on that question too - sure, the model says it's done, but did it actually get it right? That Commit check I mentioned turned out to be the answer: a second, cheaper AI reads every diff before it lands and flags anything that looks off. It costs about a penny a check. One specific bug-fixing test went from 67% actually-correct to 100% actually-correct just from adding that one habit.

Then I got greedy. Toy test files are one thing - would any of this hold up on a real, popular open-source project, not code I wrote myself to make the job easy? So I forked dotenv (the little `.env` file library that half of Node.js quietly depends on), gave it a real missing feature to build, and ran it across 7 different budget models.

12 out of 14 runs got it right - 86%, across all seven models. On real code, with real existing conventions to follow, not a sandbox built to flatter itself. The two misses weren't even code mistakes; they were the same model hitting our task time limit. We found that gap, fixed it, and re-ran the batch to confirm.

We're past 900 test runs now, and the number keeps climbing because we keep testing - not to convince you, but because I genuinely wanted to know whether "cheaper and more open" had to mean "worse." Turns out, with the right nudging, it doesn't have to.

If you've ever felt boxed into one AI provider, or wondered whether the $20/month tools are really doing something the open models can't - that's exactly the question I started with. See the actual numbers on /stats, or just come try it.

So what actually is codelazr?

It's a web harness with tweaks and stuff to turn weaker OpenRouter models into a full, high-tier coding platform. Your instruction goes to a model, but the response doesn't just get shown to you raw - our system processes it, and (optionally) a second, cheaper model checks the first one's work before it's proposed as a commit. LLMs checking each other's homework, basically. On top of that, a handful of small mechanisms catch the specific ways weaker models tend to trip - a stalled to-do list, a claim that something was done when it wasn't, a task that just hangs - and handle those gracefully so the task still finishes, instead of quietly failing on you.

What I'm actually trying to do here