MLX Workshop is a local-only macOS application.
The app accesses only model and run-workspace folders selected by the user. It may read model configuration, tokenizer, tensor metadata, and weights needed for inspection, conversion, and qualification. It writes plans, candidates, journals, logs, manifests, and evidence only inside the selected run workspace.
Host snapshots record bounded technical facts used for resource review: macOS and tool versions, Apple chip and memory capacity, free disk space, power/thermal state, and sanitized labels for relevant active ML workloads. Process arguments are used only for local classification and are not retained.
Exact local paths and secret-free command arguments are stored in run evidence for reproducibility. Users should avoid placing credentials in model or workspace paths.
MLX Workshop has no telemetry, analytics, advertising, account system, crash upload, cloud sync, model upload, or remote registry integration. The protocol-v1 workflow does not send models, prompts, logs, paths, or host facts over the network.
The bundled Python dependencies include general-purpose networking libraries because they are dependencies of MLX-LM and Transformers. MLX Workshop's allowlisted beta commands use local filesystem paths and do not invoke download or upload operations.
Delete the app to remove it. Delete the run workspace to remove generated candidates and evidence. Saved folder bookmarks and preferences can be removed with:
defaults delete com.cozylabs.MLXWorkshopQuestions may be filed in the public repository without attaching private model paths, logs, weights, or prompts.