Crush

Crush runs coding tasks against Rapid-MLX with an openai-compat provider. These commands target Crush 0.98.1.

Quick start

Install Crush 0.98.1. Start the server and keep it running while you configure and use Crush.

$ rapid-mlx serve qwen3.5-4b-4bit
$ rapid-mlx agents crush --setup
$ crush run --model rapid-mlx/qwen3.5-4b-4bit "Use a tool to inspect this directory, then summarize it."

Setup writes ~/.config/crush/crush.json, or $XDG_CONFIG_HOME/crush/crush.json when that variable is set. CRUSH_GLOBAL_CONFIG takes precedence and uses its directory instead. Use --base-url http://127.0.0.1:PORT/v1 for another port, --model to specify the server alias, and --dry-run to preview the file.

Connection

The generated JSON registers the rapid-mlx provider as openai-compat, with a /v1 base URL and a model entry containing its context window. It also selects that model as Crush's large model. For a keyed server, export RAPID_MLX_API_KEY before setup and before starting Crush; the file stores an environment reference.

Crush 0.98.1 prefers the newer crushrc format. Its configuration guide still supports JSON. A project .crushrc or global crushrc can override the JSON model selection. Check the model shown by crush models if another provider is selected. Setup sets a five-minute request timeout for the large agent prompt on smaller Macs.

Check it

$ rapid-mlx agents crush --test

This runs the integration checks and a headless Crush request against the local model.