I'm on a m1 max with 64gb ram, but i never use this vscode plugin before. Should I try?

Is this the one? https://github.com/ggml-org/llama.vscode it sems to be built for code completion rather than outright agent mode

It is RAG for your codebase, and provides code completion. The gain is the local inference, and is actually useful with smaller models.

The plugin itself provides chat also, but my gut feeling is that ggerganov runs several models at the some time, given he uses a 192gb machine.

Have not tried this scenario yet, but looking at my API bill I’m probably going to try 100% local dev at some point. Besides vibe coding with existing tools seems to not work that good for enterprise size codebases.

What languages do you work in? How much code do you keep? Do you end up using it as scaffolding and rewriting it, it leaving most of it as is?

Languages: JS/TS, C/C++, Shader Code, Some ESP Arduino code. Not counting all the boilerplate and CSS that I dont care about too much.

It very much reminds of tabbing autocomplete with IntelliSense step by step, but in a more diffusion-like way.

but my tool-set is a mixture of agentic and autocomplete, not 100% of each. I try to keep a clear focus of the architecture, and actually own the code by reading most of it, keeping straight the parts of the code the way i like.