I dug out the AI shell that I used for myself. At that time
I dug out the AI shell that I used for myself. At that time, the concept of MCP was just becoming popular (it was still the era of DeepSeek V3 and R1). In order to use it smoothly and learn related concepts, I spent a month working on it. It does not use any off-the-shelf AI SDK or LangChain.
2x3090, Qwen 3.8 Flash Next
1 shot, deepseek harness 42min /goal: Create a Diablo 1 clone, based on Three.JS, in HTML+CSS+JS, low poly 3D/voxels, isometric projection. Cool, light, pastel, pleasant graphics. Add sounds, music, etc. It should be a fully playable game, with
The top AI models always cost like a high-end gym subscription.
The output can only be IDR 300 thousand per 1M tokens. But there is one that plays in a different league: DeepSeek V4. Open weight, MIT license, SWE-bench Verified 80.6% — almost the most expensive. LiveCodeBench 93.5%.
I've been paying for API credits every month just to keep my coding agent alive.
Turns out I didn't have to. It's called Free Claude Code. FCC. It's a local proxy. Plug in a free NVIDIA NIM key, point your agent at localhost, keep working. The part I didn't expect: it isn't
You have reached the end of the archive
All of deepseek