Model intake

Intake classifies a Hugging Face source and identifies an available MLX backend. Resolution and architecture analysis read metadata. Fetching weights is a separate confirmed operation.

Resolve a source

python3 scripts/mlx-agent intake resolve <repo-or-hugging-face-url> --json

Review the verdict, model type, task, backend installation state, components, and warnings. A registry match is conversion guidance, not evidence that the model has run. Sources can identify a repository, revision, subfolder, or file. Use --revision to select a revision explicitly.

The mlx_conversions field lists up to eight MLX models that the Hub records as quantized from the source repository, ordered by downloads. Each entry includes repo, downloads, and library_name. The lookup is omitted for sources already classified as MLX and for subfolder sources. A failed lookup leaves the list empty and adds a warning without changing the resolution verdict.

These are metadata matches, not verified runtime results. Review each candidate's revision, license, size, and backend compatibility before fetching it.

Review and fetch

python3 scripts/mlx-agent intake fetch <repo> --revision <revision> --json
python3 scripts/mlx-agent intake fetch <repo> --revision <revision> --confirm --preview-hash <hash> --json
python3 scripts/mlx-agent intake status --json

The preview identifies the repository, revision, file patterns, destination, and executable arguments. Fetch requires huggingface_hub in the selected environment. It starts a receipt-tracked job; inspect status and the log before using its output. Review license, gating, storage, and source revision before downloading.

Use --file <relative-file> for one GGUF file, --hf-cache <cache-root> for a different cache, or --local-dir <absolute-directory> for an explicit destination. If resolve supplies a ported model_type, pass it with --model-type to use the declared file selection. Default snapshot fetching excludes formats MLX does not consume, including GGUF; choose a GGUF file explicitly.

Keep arguments identical between preview and confirmation. --receipts-dir selects the base for fetch receipts; use that same base for status. A returned PID indicates a started job, not a finished download.

Convert or serve

Already-usable MLX weights can be served. Original checkpoints use Convert; optional backends may need an explicit installation. GGUF conversion quality is limited by its source.

Analyze an unsupported architecture

python3 scripts/mlx-agent intake port-analysis <repo> --json
python3 scripts/mlx-agent intake port-plan <repo> --endpoint http://127.0.0.1:8080/v1 --model <served-model-id> --json

port-analysis reports deterministic metadata facts without downloading weights. port-plan sends architecture context to an already-running local model and writes a Markdown proposal. Review this generated guidance before implementing a port; it does not add a backend or establish architecture support.