Omlx

oMLX serves models locally on Apple Silicon through an OpenAI compatible API, and the whole server sits in the macOS menu bar rather than a terminal window. Continuous batching lets several requests share the machine, and the KV cache is tiered across memory and SSD so earlier context stays reusable when a conversation changes part way through. You can pin the models you use daily, let heavier ones swap in on demand, and cap context per model. Installation is a .dmg with in app updates, a Homebrew formula, or a source checkout, and it needs macOS 15 or newer with Python 3.11 to 3.13. The code is Apache 2.0, and optional MCP support and custom kernels for a few newer model families install separately.