mlx-app-converter

Use when the user wants to convert an existing Streamlit or Gradio app to use MLX (Apple's machine learning framework for Apple Silicon) instead of HuggingFace transformers. Trigger phrases: "convert this app to MLX", "use mlx-lm for inference", "use mlx-vlm for inference", "make this run on Apple Silicon with MLX", "swap transformers for MLX", "port this to MLX", "convert this VLM app to MLX". Operates on an existing app file in the current working directory (`app.py`, `streamlit_app.py`, or `gradio_app.py`); rejects other inputs (notebooks, scripts, GitHub URLs, HF model card URLs). Scope (v2): LLMs (text-generation via `AutoModelForCausalLM`) and VLMs (vision-language via `AutoModelForVision2Seq`, `AutoModelForImageTextToText`, or a curated list of family-specific classes). Multi-modal apps (LLM + VLM in the same file) are supported. Audio modalities, streaming inference, and VLM family-specific classes outside the curated allowlist are out of scope (planned for follow-up versions).

By darylalim schedule Updated 5/1/2026