Noirlab Datalab

Several clusters have genuinely fuzzy boundaries: ask_pipeworx, ask_pipeworx_beta, ask_pipeworx_grounded and deep_research all route the same 6,475 tools with only subtle differences (stable vs experimental vs grounded vs multi-facet), and entity_profile/company_facts/compare_entities/recent_changes all answer company questions with different scopes. The descriptions do explain when to prefer each, and most other tools (resolution_audit vs resolution_diff, subscribe vs unsubscribe) are cleanly separated, but an agent could easily misselect within the router and company-data families.

Mostly snake_case with a roughly verb_noun shape (ask_pipeworx, compare_entities, resolve_entity, validate_claim), but there is a large noun-only faction (company_facts, entity_profile, resolution_audit, kalshi_weather_edge) and inconsistent prefixes like datalab_query/datalab_tables/datalab_cone_search mixing verb and noun forms. Readable and predictable enough to navigate, but not a single consistent convention.

39 tools is well past the heavy threshold for what is, functionally, one gateway plus a handful of astronomy tools. The router family alone accounts for four near-duplicate entries and the Polymarket family for five overlapping scanners, so a meaningful share of the surface does not earn a distinct place.

The surface covers a lot of the lifecycle: discovery (discover_tools, suggest_questions), routing/lookup, grounding, verification, memory (remember/recall/forget), subscriptions (subscribe/list/unsubscribe/recent_alerts), and feedback. Minor gaps exist — no generic fetch/read of an arbitrary URL or record and no explicit update path for stored memory — but agents can work around these.