A native macOS app for running local models on Apple Silicon. Find a model, convert it to MLX, verify it, serve it on a loopback port, compare it against alternatives and wire it into your coding tools.
MLX Workbench is the desktop front end for mlx-agent. Every operation runs through the engine's scripts/mlx-agent … --json command line: each change is previewed before you confirm it, launched work leaves an engine receipt, and serving binds to 127.0.0.1.
The native Library keeps model readiness and memory fit beside the selected model. Screenshots show illustrative inventory.
Prepare and compare models
Requires an Apple Silicon Mac, Python 3.12 (uv recommended) and Apple's Command Line Tools (xcode-select --install). Git is needed when using a source checkout.
Download mlx-workbench-<version>.dmg from the latest release and drag MLX Workbench into Applications. In Setup Assistant, choose Use bundled engine…, confirm Prepare engine, then use Install Runtime… to install its Python dependencies. Preparing the engine copies the packaged sources; it does not install packages or download models.
You can also use a source checkout:
git clone --recurse-submodules https://github.com/cavi-ai/mlx-workbench.git
cd mlx-workbench
make installOpen the app and set Settings → Engine → mlx-agent checkout to the checkout's vendor/mlx-agent. The first-launch Setup Assistant walks through the same checks.
To build from source instead, run make start in the checkout after make install; it builds the app with Xcode and opens it.
- Library lists the GGUF and MLX weights already on disk, with whether each fits in memory now.
- Paste a Hugging Face link anywhere with ⇧⌘V. The intake window shows what the model is and how it converts before anything downloads.
- Prepare previews the conversion, its output size and fit. Confirm conversion starts the job; supported task types then run a canary before the model is marked verified.
- Run previews and starts a loopback server for the verified model.
- Clients points opencode, Continue, Zed or Aider at that server, with backups and rollback.
See Your first model for the full walkthrough.
| Area | Screens | Highlights |
|---|---|---|
| Workbench | Overview, Library | Model budget, fit per model, model details and lineage, LM Studio source |
| Lifecycle | Prepare, Compare, Run | Previewed conversion, quality gate canaries, measured comparisons in eight modes, always-on endpoints with load on request |
| Operations | Activity, Reclaim, Clients, Health | Job timeline, quarantine and Trash with restore, client config wiring with rollback, runtime readiness |
| Lab | Discover, Studies, Training | Hub candidates and task recommendations, controlled studies, text LoRA training with evaluation |
Qwen3.8-Flash-Next can also run on the optional Strata runtime from a prepared GGUF pair.
Hosted at cavi-ai.xyz/docs/mlx-workbench.
- Start: Overview, Installation, Your first model
- Workflows: Library, Add from Hugging Face, Prepare and verify, Run and endpoints, Compare, Clients, Reclaim, Discover, Health and alerts, Training and studies, Strata
- Reference: Configuration, State and files, Make targets
- Operate: Troubleshooting, Security model, Version and support
make test # Python compatibility and contract tests
make test-swift # XCTest suite for the native app
make docs-test # documentation contract testsSee CONTRIBUTING.md and Development. The docs are built from docs/mlx-workbench/source with make docs-build into the release asset mlx-workbench-docs-v<version>.tar.gz; the consumer contract describes it.
MIT. See LICENSE. mlx-agent is also MIT.