MLX-Swift-LM Skill

This repo packages a skill for agentic coding tools (Claude Code, Codex, Droid, etc.) to help with the MLX-Swift-LM library. The skill provides usage guidance, best practices, and reference notes for local LLM/VLM inference on Apple Silicon.

Source library: https://github.com/ml-explore/mlx-swift-lm/

What this skill covers

• Local LLM and VLM inference with MLX-Swift-LM
• Streaming generation and tool calling
• LoRA and adapter training flows
• Embeddings and RAG-friendly patterns
• Concurrency, cache, and tokenizer usage

Repo layout

mlx-swift-lm/skill.md          # Skill definition, triggers, and overview
mlx-swift-lm/references/       # Topic-specific deep dives
README.md                      # This file
LICENSE

References included:

• references/model-container.md
• references/kv-cache.md
• references/tokenizer-chat.md
• references/concurrency.md
• references/tool-calling.md
• references/supported-models.md
• references/lora-adapters.md
• references/training.md
• references/embeddings.md

Install (generic)

1. Copy or symlink the mlx-swift-lm folder into your tool's skills directory.
2. Restart or refresh your tool so it discovers the new skill.

Common locations:

• Codex CLI: $CODEX_HOME/skills (default: ~/.codex/skills)
• Claude Code: default ~/.claude/skills
• Droid: default ~/.factory/skills

Use the skill

Skill name: swift-mlx-lm

You can invoke it directly, or rely on triggers like: mlx, mlx-swift, mlx-lm, apple silicon llm, local llm swift, vision language model swift, lora training swift

Example prompts:

• "Use the swift-mlx-lm skill to add streaming generation to my Swift app."
• "With the MLX-Swift-LM skill, show how to run Qwen2-VL with an image."
• "Use swift-mlx-lm to explain LoRA adapter loading for MLX-Swift-LM."

Update

Pull the repo and re-sync the mlx-swift-lm folder in your skills directory.

Contributing

Edit mlx-swift-lm/skill.md for the main skill instructions and add new topic docs under mlx-swift-lm/references/.