Discover

Discover offers Browse candidates and Recommend for a task. Both ask the engine about Hugging Face Hub candidates. Neither downloads weights; downloads go through Add from Hugging Face. Switching modes preserves their current drafts and results without starting either operation.

Browse candidates

Discover runs the engine's discover command for a role and lists ranked candidates from remote catalog metadata.

Pick a Role. The picker offers only the roles the engine accepts.
Set the number of candidates per role. Any searches every supported role, so the total can exceed this limit. Fast uses repository names instead of inspecting each model; sizes and memory-fit estimates may be missing. Turn it off and select Discover again to request inspection.
Discover.

Stop discovery cancels only the current search. Search controls stay busy until that request exits; Discover then starts a fresh search. Previous successful candidates remain visible while searching, after a failure, or after stopping, labeled Showing previous search results with their original role, limit and inspection mode. Editing the controls does not relabel those results. A successful search replaces them, including when it finds no candidates. Clear removes the displayed results when no search is running. Results are kept while switching modes or routes, but are not saved across app launches.

Each candidate shows its license, a fit bar for this Mac and an In Library mark when it is already on disk. Actions: Add from Hugging Face…, Open on Hugging Face, Copy repo id.

Memory fit is an estimate, not a runtime check. Add from Hugging Face… opens a separate review before downloading; use Recommend for a task to check installed candidates through local runtimes.

The separate Catalog metadata section shows cached metadata. Refresh Metadata is disabled when no metadata provider is configured; the reason appears beside the section. Role-based discovery and cached entries remain available. This cache does not determine which models are installed; Library scans do.

This mode runs the engine's adopt flow for one role. It builds a shortlist and checks installed candidates through local runtimes, using generation and role probes where supported. Verification can load a model into RAM. A recommendation's evidence strength shows whether it came from runtime checks, metadata or heuristics; it is not a Workbench conversion verification verdict.

Pick the requested Role. Fast uses heuristic-only discovery enrichment; installed candidates still receive runtime verification prompts. Offline metadata uses cached discovery and makes no metadata network requests; local runtime probes still run.
Verify and recommend starts the checks. Refresh reads the recommendation's saved status.

Stop checks cancels Workbench's current verification request. Controls stay busy until its process exits. Saved results are kept and may be partial or still say running; Refresh reads that saved state, while Verify and recommend starts a new workflow. A local runtime may finish an in-flight prompt and keep its model loaded after Workbench stops checking.

The recommendation shows its fit on this Mac and a next step: Run… for a model you have, or Add from Hugging Face… for one you do not. Each request keeps its own state file, and a failed refresh keeps the last known result, labeled Last known state. Existing saved Adopt navigation opens this mode in Discover.