Zaica - Zig AI Coding Assistant — multi-provider LLM agent CLI with tool calling

Showcase
, ,
1

I’m excited to present pre-alpha release of Zaica - AI coding assistant.

7 Likes
2

6 posts were split to a new topic: Clarifications about AI Policy

4

Does Zaica interfacing with Ollama mean, that you could configure it to run with local LLM only, so that you can train it to give responses for a particular zig version?

Currently if you try to get some zig code examples from your search engine, there is a good chance to get a perfectly good example that only compiles with some older Zig version.

Assuming it is possible to run this locally and train the LLM, that might help newbies, and irregular Zig coders (that may have skipped a minor version or two since last using Zig and didn’t read the intermediate release notes). It would require combining Zaica with a description of the steps to set up Ollama, what model to use, and a list of compatible source code repositories (the compiler itself, but also other projects checked to build with that version) and instructions on how to use them to train/extend the model.

Such a setup might not be as responsive (depending on the hardware available to you), but it I for one rather have a slow response that compiles than a fast one that requires zig 0.12.

1 Like
7

Nice work!

1 Like
8

My biggest pain with any local or open LLMs is that they can’t predictably use newest std for Zig 15.2. No matter how hard I try to force them or even provide ready answer thay always fall back to earlier versions of Zig. I tried aider, md files guidance etc. and many open models that could fit on my 4090 but no luck there. Only frontier models, Opus 4.6, GPT 5.4 and Gemini 3 Pro seem to get it right. Anyone had more luck? What am I missing?

1 Like
9

I recently tested cloud versions of GLM 5.1 and Deep Seek v4 Pro and they both managed to vibe code my own benchmark test of making a hello world rainbow triangle in Zig 0.15.2 with static webgpu lib. Normally almost all models fail even with some RAG help as Zig 0.15.2 and webgpu are not strongly present if any in traning data. GLM was 3x faster then Deep Seek but 2x slower then Opus 4.6. GPT 5.4 and Gemini 3 Pro which all manage to do it within 5 minutes. No other model can do it so far and I tested many.