Run LLMs natively
on Apple Silicon

One command to spin up a coding agent on your local GPU. No API keys, no cloud costs, no latency.

Pick from 5 agent launchers: Claude Code, Codex, Opencode, Pi — or start a plain chat.

Model picker built in. Select, launch, code. Under 10 seconds from cold start.

Drop-in replacement for any OpenAI-compatible tool. Just point to localhost:6767.

Supports Chat Completions, Responses API, Anthropic Messages — one server, every protocol.

Hot-swap models without restarting. LRU cache keeps your most-used models in memory.

168+ models from the curated MLX registry. Human-friendly aliases, Apple Silicon optimised weights.

Multi-select download — queue up several models and grab them in one go.

Models are stored locally in ~/.ppmlx/models. No Docker, no containers, no overhead.

Auto-downloads missing models on first use. Just type the name and start chatting.

Streaming REPL with token stats, timing, and slash commands built in.

Switch models mid-session with /model. No need to restart anything.

Bring any HuggingFace model. Convert to MLX format and quantize to 4-bit in one step.

69% smaller models with minimal quality loss. Fits large models in unified memory.

Run your quantized model immediately — no extra setup, no config files.