jan.ai with some local model

This. Llama.cpp with Vulkan backend running in docker-compose, some Qwen3-Coder quantization from huggingface and pointing Opencode to that local setup with a OpenAI-compatible is working great for me.