MLX Agent overview

mlx-agent helps you choose and operate local models on Apple Silicon for a coding agent. It combines Hugging Face discovery, local role verification, performance measurements, configuration transactions, and model lifecycle commands in one provider-neutral Python core.

Requirements and boundaries

The core requires Python 3.9+ and uses the standard library. MLX execution requires Apple Silicon and the appropriate runtime. Optional backend installation uses Python 3.12 dependency locks; provider CLIs are installed separately.

Discovery, adoption, and benchmarks do not download weights. Explicit intake fetch downloads a reviewed model source; backend install installs declared optional dependencies. Both require a reviewed preview and confirmation.

Wire and Fleet change configuration. Serve launches owned processes. Convert, LoRA, and Fuse create outputs from local inputs. Those jobs have reviewed starts and receipts, but are not configuration rollback transactions. Research and Blueprint write guidance packs without training or scaffolding applications.

The pipeline

1. Discover — discover ranks Hub candidates using size, license, gating, reasoning detection, and KV-cache-aware fit estimates.
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 models, fleet, and watch manage local model processes, health, configuration, and Hub changes.
6. Produce — intake, backend, convert, lora, and fuse acquire sources, manage dependencies, quantize, train, and fuse local models.

Your first local model covers making weights available and starting a server before verification. For domain research or project design, use Research or Blueprint. See Evidence and sizing for probe and fit limits.

Providers

First-class adapters ship for Claude Code, Codex CLI, Agy, and OpenCode, plus portable AgentSkills packages. Every provider exposes the seven capabilities: Claude Code and OpenCode use /mlx-* commands, Codex uses $mlx-agent:mlx-*, and Agy and AgentSkills use the named mlx-* skills.