Aemps Cima

The descriptions work hard to route callers ('START HERE', 'PREFER OVER', 'step up only when needed'), and most tools do carve out a distinct niche. But several near-duplicates remain: ask_pipeworx / ask_pipeworx_beta / ask_pipeworx_grounded / deep_research / validate_claim are all universal-question routers with the same argument shape, and the prediction-market cluster (bet_research, polymarket_edges, polymarket_arbitrage, polymarket_kalshi_spread, release_calendar_markets) overlaps heavily. ask_pipeworx_beta is by its own description byte-equivalent to ask_pipeworx, which is a deliberately unresolvable choice.

Everything is snake_case with a stable vocabulary and clear prefix families (cima_*, polymarket_*, ask_pipeworx*, pipeworx_*). The only deviation is verb-first (resolve_entity, compare_entities, validate_claim) vs noun-first (entity_profile, company_facts, resolution_audit) naming, which is readable but not a single uniform pattern.

41 tools is heavy for a server whose name advertises one narrow domain (the Spanish AEMPS CIMA register), which only accounts for 5 of them. The remaining ~36 gateway/meta/prediction-market tools are a much broader surface than the stated scope implies, and the redundancy (beta router, three overlapping Polymarket scanners) means several slots are not independently earned.

The CIMA subset covers search, full record, structured SmPC/prospecto text, change feed and shortage feed — a solid lifecycle for that register. The gateway side covers discovery, routing, grounding, memory, subscriptions and feedback, with only minor gaps (no drug-interaction check across CIMA records, no direct ATC-hierarchy browse, no PyPI/Maven equivalents of scan_dependency).