Scored across 29 tools
Several boundaries are fuzzy despite detailed descriptions: fhir_read vs fetch (both read one resource by id), search vs fhir_search (both query records), sources_check vs wearables_sync_status (both survey connections), and fhir_validate vs fhir_propose_write (the latter also validates). The ChatGPT-connector vs native duplication is somewhat explained by descriptions, but an agent could still misselect.
A domain-prefix pattern exists (fhir_, curatr_, questionnaire_, action_), but ordering and form are inconsistent: verb_noun for most, yet noun_noun for action_status, care_gaps, guardrail_conformance, and noun_verb for sources_check. The bare search/fetch also break the prefix convention used by their near-twins fhir_search/fhir_read.
At 29 tools the surface is heavy and the count is inflated by overlapping pairs (read/fetch, search/fhir_search), but the domain genuinely spans FHIR CRUD, data quality, real-world actions, sharing, wearables, and conformance, so most tools have a plausible place. Borderline rather than clearly excessive.
The FHIR lifecycle is well covered: read, search, stats, lastn, compiled-truth with provenance, propose/validate/commit writes, permission evaluation, questionnaire populate/extract, and data-quality fixes. Minor gaps remain (no explicit delete, no patient enumeration), but core workflows and guardrail operations are present.