oMLX: LLMs in your menu bar
oMLX is a menu‑bar LLM runner built for Apple Silicon with cold/hot KV cache that spills to SSD, a Web UI for model management, and compatibility with OpenAI/Anthropic—installable via Homebrew and aimed at heavy Mac‑based AI workflows. — it’s a practical example of local LLM tooling that sidesteps cloud latency and privacy concerns. (x.com)
The project is published as jundot/omlx on GitHub and shows roughly 6.1k stars, ~489 forks and about 490 commits on the main branch, licensed under Apache‑2.0. (github.com/jundot/omlx) Packaging includes a notarized macOS app distributed as a DMG with in‑app auto‑update, and the packaging code enforces a minimum macOS 15.0 for MLX ≥ 0.29.2. (omlx.ai; github.com/jundot/omlx/tree/main/packaging) A Homebrew formula is included in the repo (Formula/omlx.rb) that declares build dependencies on rust and python@3.11, restricts installation to arm64 macOS, and defines a service to run the server via brew. (github.com/jundot/omlx/blob/main/Formula/omlx.rb) The releases page shows a recent pre‑release v0.2.19.dev2 with changelog notes fixing a Metal GPU crash and an idle GPU spike, with the pre‑release published in the last 24 hours. (github.com/jundot/omlx/releases) Repository code and docs reference an upgraded mlx‑lm dependency (v0.31.1) and framework‑level support for serving LLMs, embedding models and rerankers simultaneously from a single process. (github.com/jundot/omlx/tree/main/omlx) Community threads highlight agent workloads as a primary motivation: a Hacker News post by the author frames caching for shifting context, and a recent GitHub issue (#313) documents a user running Qwen3.5‑122B on a 128GB Mac and observing an SSD‑first caching behavior. (news.ycombinator.com/item?id=47080277) (github.com/jundot/omlx/issues/313) The Homebrew formula's install steps explicitly create a Python 3.11 virtualenv, build certain Rust‑backed Python wheels from source, and add linker flags to avoid Mach‑O header problems during packaging. (github.com/jundot/omlx/blob/main/Formula/omlx.rb)
