› cavi-ai/
GitHub ↗

Serve

Serve launches local MLX servers through the same preview-confirm discipline as wire.

bash
python3 scripts/mlx-agent serve start --repo <repo> --runtime mlx_lm
python3 scripts/mlx-agent serve start --repo <repo> --runtime mlx_lm --confirm --preview-hash <hash>
python3 scripts/mlx-agent serve status
python3 scripts/mlx-agent serve stop --port 8080

Hard gates

  • Model present in the Hugging Face cache (never downloads).
  • Runtime executable already installed (never installs).
  • Port free and unclaimed by wired configs; loopback-only bind.
  • Every start writes a receipt: pid, argv, port, log path.

serve stop signals only receipt-owned processes, and only after verifying the live pid's argv matches the receipt — never a port-scan kill. serve status cross-checks receipts against live processes.

Recipes and adapters

mlx_lm (default :8080) and mlx-vlm (default :8083) launch directly; Ollama and LM Studio manage their own servers. --adapter-path serves a trained LoRA alongside the base on mlx_lm.

launchd

serve start --launchd installs the same plan as a launchd agent: a deterministic plist (managed com.mlx-agent.serve.<port> label) applied through the transaction flow. It prints the exact launchctl bootstrap command — loading stays with you.