MLX Agent overview

mlx-agent discovers, verifies, measures, wires, serves, converts, and trains local MLX-optimized models on Apple Silicon — for your coding agent. It queries the Hugging Face Hub live, matches models to your machine's memory and installed runtimes, proves behavior with deterministic runtime probes, measures real tokens-per-second on your chip, and emits confirmation-gated configuration for the agent you use.

What it is not

• It never downloads model weights. Discovery and verification read metadata and already-served models only.
• It never installs provider CLIs or model runtimes. You install them; the pack detects them.
• It never mutates configuration without a reviewed preview and an explicit --confirm --preview-hash.

The pipeline

1. Discover — scout ranks live Hub candidates with real download sizes, license and gating, reasoning detection, and KV-cache-aware fit.
2. Verify — adopt runs deterministic role probes (coding, reasoning, vision OCR, embedding, tool-use) against your local runtime before recommending.
3. Measure — bench records TTFT and tok/s on your chip as runtime_measured evidence.
4. Wire — wire applies configuration through a preview-confirm-receipt-rollback transaction.
5. Operate — serve, doctor, fleet, and watch run the fleet day to day.
6. Produce — convert, lora, and fuse quantize, train, and fuse your own models.

Providers

First-class adapters ship for Claude Code, Codex CLI, Gemini CLI, and OpenCode, plus portable AgentSkills packages. Every provider exposes the seven capabilities: /mlx-scout, /mlx-adopt, /mlx-wire, /mlx-bench, /mlx-doctor, /mlx-watch, and /mlx-fleet ($mlx-agent: prefix on Codex).