Claude Code taught to work with local LLMs via llama-server and a Go launcher

⚡ The gist in 5 seconds - The author has reworked a Go launcher so that llama-server works in tandem with Claude Code as a backend for local models. - The solution targets large MoE models in the 35B, 120B, and 176B classes, running on limited hardware. - No ready-made build is published: the breakdown and the logic of the rework are in an article on Habr, and the configuration must be reproduced on your own. ### 🔍 What was found On Habr, a piece came out on how to connect llama-server to Claude Code using a reworked Go launcher. The author is the same developer who previously published articles about running large MoE models on limited hardware via llama-server and about comparative experiments with models in the 35B, 120B, and 176B classes. The essence of the rework is adapting the Go launcher (a wrapper for launching llama-server) so that Claude Code runs on top of a locally hosted backend instead of calling a cloud API. The specific implementation details, launch flags, and configuration are disclosed by the author in the article itself at the link: https://habr.com/ru/articles/1094534/, not in the accompanying description. ### 💡 Why it matters Claude Code works out of the box with Anthropic's cloud infrastructure, so pairing it with a fully local llama-server is a practical way to remove dependence on an external API and per-request costs for those who already run MoE models on their own hardware. The value of the material lies in working notes on gluing together two tools that are incompatible "out of the box," not in a standalone product: no ready-made artifact or official support for such a configuration is claimed in the source.