Related Servers

Most tools map to a distinct resource (providers, models, hardware, pulls), but several pairs overlap meaningfully: search_available is explicitly a subset of search_models (source 'catalog'), diagnostics partially duplicates health_check + system_resources, and ensure_model overlaps with the pull_model/load_model/fit_check combination. complete vs delegate is also a judgement call for an agent, since both run inference and only the description clarifies which to pick. Descriptions do a lot of work to disambiguate, keeping this above a 2.

The dominant pattern is snake_case verb_noun (pull_model, list_models, search_models, remove_model, suggest_model), but it is mixed with noun_noun (model_info, system_resources, health_check, fit_check) and bare single-word tools (complete, delegate, embed, benchmark, diagnostics). Everything is readable and there is some within-family consistency (list_*, pull_*, search_*), but the overall set does not follow one predictable convention.

22 tools is on the heavy side for a local-runtime manager, and a few do not clearly earn their place: search_available is redundant with search_models, and pull_status/pull_cancel are arguably one job-status tool. The remaining surface (lifecycle, discovery, hardware, inference) is defensible, so this sits at borderline rather than excessive.

The model lifecycle is well covered end to end: search/suggest, pull with background jobs, load/unload, fit check, benchmark, delete, plus inference (complete/delegate/embed) and hardware/diagnostics. Minor gaps remain, notably no way to add or edit provider configuration, no model copy/alias/rename, and no way to cancel long-running non-pull work such as benchmark. These are workaround-able rather than blocking.