oMLX Mac setup with Claude Code, OpenCode and Pi

The presenter explains why he prefers oMLX for running a local LLM on his Mac, then walks through installation and connections to coding tools. Built on MLX LM, oMLX adds a server, model management and a dashboard. He describes a cache that keeps hot blocks in RAM and saves cold blocks to SSD, with reusable prefixes surviving server restarts.

Setup starts with a stable DMG release from GitHub. The walkthrough shows how to find an MLX model on Hugging Face, paste its URL into settings and download it. ModelScope is another download source. The dashboard provides server endpoints and quick-start commands for a coding assistant such as Claude Code.

The demonstration uses an M5 Max with 128 GB RAM. During the session, the presenter reports about 1.5 GB of disk cache and roughly 80 GB of RAM use, illustrating how context can increase memory demand. He also changes the context setting and notes that the model needs reloading for it to take effect.

The comparison reflects his preferences: he finds LM Studio easy to start with but wants a lighter application, and describes Ollama's MLX support as still rolling out. Although he demonstrates Claude Code, he prefers OpenCode to conserve context. The final setup connects Pi through a provider base URL and API key.