AR6

Workshop / LLMarium

LLMarium

The local models of a Mac in the menu bar: which MLX engines run, what each process keeps in memory and who launched it, whether a model fits before it loads, and one click to open it in a coding agent.

State
16 Oct 2026, private and in development (0.1.0, nine commits from 6 to 13 October); launching a model from the menu bar is written but not yet committed, and no open-source licence has been added. m
commits per week, 12 weeks · 8 in the last 7 days · as of 14 Oct
Runs on
macOS 26 or later, with oMLX, MTPLX or mlx-serve installed; measured on a single Mac
Built with
Swift and SwiftUI, no third-party dependency; a menu bar app, a command-line tool and an MCP server
Data
Local files, keys in the Keychain, a local entry point on 127.0.0.1; reads public model lists and benchmarks
Code
Private repository, open source announced

Known limits

  • Three MLX engines only; Ollama, llama.cpp and LM Studio are left out on purpose. m
  • Measured on one machine, and the two recorded benchmarks were disturbed by another tool using the same engine. m
  • Some proofs are missing: mlx-serve has never been launched for real, and the switch away from LiteLLM is written but not done. m
  • Not distributable yet: French interface, ad hoc signature, no icon, no licence. m

Sources and questions

Register · 9 entries

  1. correction

    The local model launcher is reframed and renamed LLMarium (6 October). The launcher was reframed on 6 October and renamed LLMarium on 10 October. c

  2. correction

    Models in memory are listed with the process that launched them (13 October). Listing each model in memory with the process that launched it was committed on 9 October. c

  3. milestone

    Models in memory are listed with the process that launched them. m

  4. rename

    The local model launcher, until then “Modèles locaux” with the command lml, becomes LLMarium after four waves of names. m

  5. decision

    LLMarium keeps to three MLX engines and takes five ideas from neighbouring tools: a GPU memory ceiling, a speed estimate, an idle stop, a model trash and a radar sorted by use; Ollama and llama.cpp stay out. m

  6. milestone

    A local model opens in four coding agents from the app, Claude Code, Codex, Pi and OpenCode, each finishing a tool-call task; the fifth, DeepSeek Harness, is configured but not yet run. m

  7. decision

    A local entry point is built without any dependency to replace a LiteLLM gateway: it rewrites only the model name and streams answers through untouched; the switch itself is not made. m

  8. decision

    The local model launcher is reframed and renamed LLMarium. m

  9. measure

    Claude Code and Codex each finish a tool-call task, reading a secret file and answering with its token, on a local oMLX model and on a local MTPLX model: four pairs out of four. m