The reason to route these questions through a server at all is that a model will produce a plausible ICD or LOINC code from memory and be wrong in a way nobody catches. Here each answer names its source, its data vintage and the moment it was fetched, and carries a citation and a licence with it. Two practical notes before wiring it up: ICD-11 is the only part that needs credentials, and the SNOMED tools ship disabled because the public endpoint behind them was retired, so enabling them means running your own Snowstorm.
A server over the major medical terminologies that calls WHO, NLM, NIH and DataSUS rather than recalling codes from training data. It searches and resolves codes in each system, walks their hierarchies, maps between them, and validates a mixed batch of codes against their sources.
- ICD-11 searched, looked up by code or URI, walked up and down the hierarchy, listed by chapter, and queried for postcoordination axes when a composite code is needed
- LOINC: search, full details for a code with its component and property, the valid answers for a questionnaire item, and the structure of a panel such as a CBC
- RxNorm by brand or generic name, a concept by RxCUI with its synonyms and status, its active ingredients, its therapeutic and pharmacologic classes, and mapping both ways between RxCUI and NDC
- MeSH descriptors searched and resolved, with scope notes, tree positions and the qualifiers each one allows
- WHO ATC classification — the code for a drug by name, an ATC code resolved to its class name and level, and every substance belonging to a class
- Brazilian CID-10 in Portuguese, searched by text, looked up by code, and browsed by chapter with its constituent groups
- map_icd10_to_icd11 uses WHO's own transition tables bundled with the server; find_equivalent runs a ranked search for the same concept across several terminologies at once; validate_codes checks a mixed batch against their sources, which is what auditing a legacy database actually needs
- terminology_versions reports the version, release date, publisher, source URL and update cadence of every terminology it queries, and terminology_diff says what comparison data exists between two versions
- Every successful result carries provenance — source, canonical URL, data vintage, the real extraction instant, a ready-to-use citation and the licence — in structuredContent, with a compact footer for text-only clients
- Three prompts wrap the common workflows into one-click actions, and four resources hold reference content read in-process
Published as medical-terminologies-mcp and launched over stdio, or run as an HTTP server. Most of it needs no credentials at all: LOINC, RxNorm, MeSH, ATC and CID-10 work unauthenticated. The ICD-11 tools need WHO_CLIENT_ID and WHO_CLIENT_SECRET from the WHO ICD API, with WHO_ICD11_RELEASE_ID choosing the release and defaulting to 2024-01. The SNOMED CT tools are off by default because the public Snowstorm endpoint they used was retired and now answers HTTP 410; setting ENABLE_SNOMED_TOOLS also requires SNOMED_BASE_URL pointing at a Snowstorm instance you host yourself.
One command plus a key — npm install -g medical-terminologies-mcp, then supply credentials