Medical Terminologies MCP Server

A Model Context Protocol (MCP) server providing unified access to major global medical terminologies:

• ICD-11 - International Classification of Diseases (WHO)
• LOINC - Logical Observation Identifiers Names and Codes
• RxNorm - Normalized names for clinical drugs (NIH)
• MeSH - Medical Subject Headings (NLM)
• ATC - Anatomical Therapeutic Chemical classification (WHO Collaborating Centre, served via NLM RxClass)
• CID-10 - Brazilian Portuguese translation of ICD-10 (DataSUS V2008, bundled)

🇧🇷 Leia em Português

See it in action

Ask your assistant:

• "What's the ICD-11 code for type 2 diabetes?" → icd11_search
• "Map ICD-10 code E11 to ICD-11." → map_icd10_to_icd11
• "What does LOINC 2339-0 measure?" → loinc_details
• "Qual o código CID-10 para infarto agudo do miocárdio?" → cid10_search

The answers come from authoritative sources (WHO, NLM, NIH, DataSUS) — real codes and mappings, not guesses from training data.

Features

• 33 tools: 31 terminology tools plus search/fetch for ChatGPT Deep Research
• 3 MCP Prompts that orchestrate tool calls into named workflows (find-medical-code, drug-info, cid10-portuguese-lookup) — clients render these as one-click user actions
• 4 MCP Resources for in-process reference content (info://server, info://cid10/chapters, info://licenses, info://stats) — sub-millisecond reads (except info://stats which round-trips to the StatsCounter Durable Object on the hosted endpoint)
• Multi-terminology support in a single server
• Cross-terminology mapping and search
• Provenance on every response (since v1.8.0): each successful tool result carries a machine-readable provenance block — source, canonical URL, data vintage, real extraction instant (cache hits keep the original fetch instant), ready-to-use citation, and license — in structuredContent.provenance + attribution, mirrored in _meta under com.sidneybissoli.medical/*, with a compact text footer for text-only clients. Multi-source responses (find_equivalent, validate_codes) carry one block per source; server-computed ranking fields are flagged as derived
• Built-in caching for improved performance
• Rate limiting to respect API limits
• Detailed responses with rich formatting
• Two transports: stdio (default; for Claude Desktop, IDE clients) and Streamable HTTP (the hosted Cloudflare Worker at https://medical.sidneybissoli.com/mcp, or your own instance of worker/)

📖 Article (in Portuguese): CID-10, CID-11 e o que muda para quem trabalha com dados do SUS — the V2008 structure in numbers, what the WHO transition tables are and are not, and the licences that differ between sources. Also published on the site, in Portuguese and English: sidneybissoli.com.

Who is this for?

This server is not a clinical-care decision tool — practicing clinicians have specialized assistants (UpToDate AI, OpenEvidence, EHR-integrated tools) for that. The actual audience is researchers, public-health analysts, clinical informatics developers, and educators who need programmatic access to authoritative terminology data.

What this server is not intended for — and what leaves your machine

• It is not intended for clinical decision support. It retrieves what the official sources publish. It does not diagnose, recommend treatment or dose.
• It is not designed to run offline. ICD-11, LOINC, RxNorm, MeSH and ATC are answered live by the public WHO and NLM APIs. Every query string you send is forwarded to those services. De-identify term lists before running them: no patient names, free-text notes or record identifiers. Two datasets are bundled and answered in-process, with no network call: CID-10 (DataSUS V2008) and the WHO ICD-10 → ICD-11 transition tables (map_icd10_to_icd11).
• Credentials are not required on the hosted endpoint. https://medical.sidneybissoli.com/mcp has WHO credentials configured. WHO_CLIENT_ID/WHO_CLIENT_SECRET are needed only when you run the server yourself, and only for the 5 ICD-11 tools.
• SNOMED CT is not served. It was retired in 2.0.0: no public Snowstorm host remains, SNOMED content needs a per-country license, and the off-by-default tools confused every catalog that listed this server. See SNOMED CT (retired in 2.0.0).
• It is not OMOP-shaped. Results carry the source's own codes, not OMOP concept_ids, and there is no concept_ancestor traversal. If your output has to join against an OMOP CDM, use an OMOP vocabulary service; use this server for general terminology work.
• Search is lexical, not semantic. find_equivalent and the *_search tools match words, not meanings (no embeddings).
• Large batches are paced by the upstream rate limits (see API Rate Limits). Thousands of distinct terms take minutes to tens of minutes.

Reproducibility: which version answered

A crosswalk or a coded dataset is only meaningful against a stated vocabulary version. Two things record it:

• terminology_versions lists the release each terminology is queried against. Call it at the start of a batch run and keep the output with your results. The ICD-11 release is pinned (default 2026-01, override with WHO_ICD11_RELEASE_ID), so a self-hosted run is repeatable until you change it.

• The provenance block on every response (structuredContent.provenance) carries retrieved_at (the real extraction instant; a cache hit keeps the original fetch time), the source URL, the citation and the license. Its data_vintage field carries the version when the source exposes one:

  Source	data_vintage in each response
  ICD-11	the pinned WHO release (e.g. 2026-01)
  ICD-10 → ICD-11 tables	the bundled WHO release (e.g. 2025-01)
  CID-10	V2008
  LOINC, RxNorm, MeSH, ATC	null, because these APIs do not state a release per response. Record terminology_versions + retrieved_at instead

Minimal batch log: one terminology_versions call at the start, plus the provenance of each result (source, data_vintage, retrieved_at).

Try the hosted instance (no install)

A public Cloudflare Workers deployment runs at:

https://medical.sidneybissoli.com/mcp

Connect via the MCP Inspector or any Streamable HTTP MCP client:

npx @modelcontextprotocol/inspector --transport streamable-http \
  --server-url https://medical.sidneybissoli.com/mcp

Or install via Smithery, which proxies the same endpoint through their gateway:

npx -y smithery mcp add sidneybissoli/medical-terminologies-mcp

The hosted instance has WHO credentials configured, so all 33 tools work without any setup on your side. For your own deployment (e.g. corporate network, different region, custom WHO credentials), see the Installation and Hosted on Cloudflare Workers sections below.

Installation

Global Installation (Recommended)

npm install -g medical-terminologies-mcp

Local Installation

npm install medical-terminologies-mcp

Configuration

Claude Desktop

Add to your Claude Desktop configuration file:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "medical-terminologies": {
      "command": "npx",
      "args": ["-y", "medical-terminologies-mcp"],
      "env": {
        "WHO_CLIENT_ID": "your-who-client-id",
        "WHO_CLIENT_SECRET": "your-who-client-secret"
      }
    }
  }
}

Environment Variables

¹ Required for ICD-11 tools. Get credentials at: https://icd.who.int/icdapi.

LOINC, RxNorm, MeSH, ATC and CID-10 need no configuration.

HTTP transport (hosted)

The server runs over stdio by default — that's what Claude Desktop and IDE clients expect. The Streamable HTTP transport is served by the Cloudflare Worker in worker/ (an instance of the maintainer's Fase 0 hosting template). The --http flag of the Node entry was removed in v1.6.0 — if you need a local HTTP endpoint, run the Worker locally:

npm ci && cd worker && npm ci
npm run dev     # wrangler dev on http://localhost:8787
# Inspector via HTTP
npx @modelcontextprotocol/inspector --transport streamable-http --server-url http://localhost:8787/mcp

Hosted endpoints (production and local alike):

• POST /mcp — JSON-RPC over Streamable HTTP (the MCP protocol). Stateless mode: each request is independent.
• GET /health — liveness probe returning { status, name, version, tool_count, uptime_s }.
• GET /status — version + deploy metadata. GET /metrics — aggregated per-tool usage.
• GET /stats and GET /stats/badge — public tool-call counter (since 2026-05-13) and its shields.io badge.
• GET /.well-known/mcp/server-card.json — static server card for registry scanners.
• CORS is permissive (*) so browser clients (e.g. the MCP Inspector web UI) can connect directly.

ChatGPT (Deep Research)

ChatGPT deep research (and company knowledge, and research workflows over the Responses API) only uses an MCP server that exposes exactly search and fetch — this server does, on top of the terminology tools. Point the connector at the hosted endpoint, no key required:

https://medical.sidneybissoli.com/mcp

search ranks the query across the bundled CID-10 (categories, subcategories, chapters), the terminology version records and a live fan-out to ICD-11, LOINC, RxNorm and MeSH (the same fan-out find_equivalent does; a source that fails is skipped) and returns { id, title, url }; fetch renders the document through the terminology's own lookup tool (cid10_lookup, icd11_lookup, loinc_details, rxnorm_concept, mesh_descriptor, terminology_versions) as readable Markdown with the canonical public page (WHO ICD browsers, loinc.org, RxNav, MeSH Browser), which is what ChatGPT cites. Both carry the same provenance block as every other tool — search one block per source that answered, like find_equivalent. In ChatGPT's developer mode (Settings → Security and login → Developer mode) any tool is callable — the terminology tools remain the ones to use for data.

Hosted on Cloudflare Workers (primary)

The production deployment is the Cloudflare Worker in worker/, config in worker/wrangler.jsonc, CI deploy in .github/workflows/deploy-worker.yml (auto-runs on every push to main).

To deploy your own instance:

npm ci && npm run build:worker-lib
cd worker && npm ci
npx wrangler login         # one-time, browser flow
npx wrangler deploy        # publishes to <name>.<account>.workers.dev
# Set ICD-11 secrets so those 5 tools work:
npx wrangler secret put WHO_CLIENT_ID
npx wrangler secret put WHO_CLIENT_SECRET

Note: worker/wrangler.jsonc pins the maintainer's account_id and custom domain route — remove/replace both for your own deployment.

Why Workers: zero cold start at the edge, $5/mo flat for 10M requests (free tier covers up to 100k req/day), and no VMs to size or restart. The template ships per-IP rate limiting and a usage-stats Durable Object; the upstream-facing cache/rate-limiter are per-isolate (PROGRESS.md Phase 11.9 Stage 2 tracks the KV/DO upgrade).

Listing on Smithery

After your Worker is live, register the URL on Smithery:

1. Visit https://smithery.ai → Publish → MCP (or https://smithery.ai/new).
2. Pick the URL submission path (Smithery deprecated container hosting in 2024 — URL is the supported flow now).
3. Paste https://<your-worker>.workers.dev/mcp. Smithery's gateway scans for compliance and proxies traffic.

Available Tools (33)

Official Portuguese (pt-BR) content

The server never machine-translates terminology content — but several sources publish official translations, and the tools expose them:

• CID-10 is natively Portuguese: cid10_search / cid10_lookup / cid10_chapter(s) serve the DataSUS V2008 dataset (the CID-10 the Brazilian SUS uses operationally).
• ICD-11 in official Portuguese: pass language: "pt" to icd11_search / icd11_lookup to search and read WHO's official pt-BR linearization labels.
• MeSH: pass language: "pt" to mesh_search / mesh_descriptor to request NLM's official translations where they exist.

If a source has no official translation for an entry, you get the source language back — never a machine translation.

ICD-11 Tools (5)

LOINC Tools (4)

RxNorm Tools (5)

MeSH Tools (4)

Crosswalk Tools (4)

ATC Tools (3)

WHO Anatomical Therapeutic Chemical classification, served through NLM RxClass (free, no auth). The WHOCC base itself requires a paid subscription, but RxClass envelopes the same code/name pairs.

CID-10 Tools (4)

Brazilian Portuguese translation of ICD-10 (DataSUS V2008). Bundled as a static dataset — no HTTP calls. The Brazilian SUS uses CID-10 V2008 operationally; for the international ICD-11 (current WHO revision), use the ICD-11 tools above.

Ask in your words, not the CID-10's. The CID-10 is worded in clinical Portuguese, and cid10_search matched your words against the code title as one verbatim phrase — so the everyday word returned nothing at all. Measured over the 14,496 categories and subcategories of the bundled V2008 dataset (2026-09-16), fixed since 1.12.0: every word must match (AND), and the everyday word is expanded to the CID-10's own (src/clients/cid10-vocabulary.ts, measured pairs only) — the response says so in vocabulary_notes, and zero results come with a way out.

What the V2008 dataset does not carry stays out and still returns zero — covid (U07.1 is from 2020), zika — because an alias for a code that does not exist promises what the source does not have. The same table feeds the Deep Research search index.

Versioning Tools (2)

Surface what version of each terminology this server queries against today — useful when running batch validation against a pinned release or when investigating an unexpected lookup miss after an upstream update.

ChatGPT Deep Research (2)

The OpenAI Deep Research contract — the only two tools without a terminology prefix (names fixed by OpenAI). See ChatGPT (Deep Research) above.

Example Outputs

The samples below are the actual formatted output the tools produce — the text body of the CallToolResult. Tools also return a structuredContent object matching each tool's outputSchema for programmatic consumers.

loinc_search — query: "glucose", max_results: 3

## LOINC Search Results for "glucose"

Found 1024 total results (showing 3):

1. **74790-7** - Glucose challenge (hydrogen breath test) panel - Exhaled gas
   Component: Glucose challenge panel | Method: -

2. **104708-3** - Deprecated Estimated average glucose [Moles/volume] in Blood
   Component: Estimated average glucose | Property: SCnc

3. **97510-2** - Glucose measurements in range out of Total glucose measurements during reporting period
   Component: Glucose measurements in range/Total glucose measurements | Property: NFr | Method: Calculated

total_count (1024) reflects every match in the NLM Clinical Tables index, not just the page returned. Bump max_results (max 50) to see canonical codes like 2339-0 (Glucose [Mass/volume] in Blood); the API's relevance ranking puts panels and derived measurements above plain blood-glucose at small page sizes.

rxnorm_ingredients — rxcui: "6809" (metformin)

# Ingredients for RxCUI 6809

Found 18 ingredient(s):

| RxCUI | Name | Type |
|-------|------|------|
| 6809 | metformin | Single Ingredient |
| 1007411 | chlorpropamide / metformin | Multiple Ingredient |
| 1043562 | metformin / saxagliptin | Multiple Ingredient |
| 1243019 | linagliptin / metformin | Multiple Ingredient |
| 1486436 | dapagliflozin / metformin | Multiple Ingredient |
| 1545149 | canagliflozin / metformin | Multiple Ingredient |
| 1664314 | empagliflozin / metformin | Multiple Ingredient |
| 729717  | metformin / sitagliptin | Multiple Ingredient |
| ...     | (10 more combinations)   | Multiple Ingredient |

For an RxCUI that is itself an ingredient (TTY=IN), the tool returns that ingredient plus every multi-ingredient (TTY=MIN) concept that includes it. Use this to enumerate combination products built around a substance.

mesh_descriptor — mesh_id: "D006973" (Hypertension)

# Hypertension
MeSH ID: D006973

## Scope Note

Persistently high systemic arterial BLOOD PRESSURE. Based on multiple readings (BLOOD PRESSURE DETERMINATION), hypertension is currently defined as when SYSTOLIC PRESSURE is consistently greater than 140 mm Hg or when DIASTOLIC PRESSURE is consistently 90 mm Hg or more.

## Tree Numbers

- C14.907.489

## Concepts

- Hypertension *(preferred)*

## Allowed Qualifiers

35 qualifier(s) allowed. Use mesh_qualifiers for details.

The scope note comes from the descriptor's preferred concept, not its annotation field (which is an indexer-facing note). Tree numbers are the navigable path into MeSH's controlled hierarchy — C14.907.489 places Hypertension under Cardiovascular Diseases → Vascular Diseases.

Common Workflows

• ICD-11 lookup: icd11_search with a clinical term → pick the result → icd11_lookup with the code for full details, or icd11_hierarchy to walk parents/children.
• Drug pipeline: rxnorm_search for a brand or generic name → rxnorm_concept for the canonical record → rxnorm_ingredients and rxnorm_classes for downstream analysis.
• Harmonize a column of free-text terms: terminology_versions once at the start → harmonize_terms in batches of up to 50 (each term with its domain) → accept exact, spot-check strong, send needs_review to a person → keep each row's provenance with the crosswalk. Write lab terms with specimen and property ("glucose serum"): a bare "glucose" matches over a thousand LOINC codes.
• Cross-terminology scaffolding: find_equivalent with a clinical term searches ICD-11, LOINC, RxNorm and MeSH in one call. Use it to bootstrap mappings; the pairwise map_* tools refine them.
• ICD-10 → ICD-11 (authoritative): map_icd10_to_icd11 reads the bundled WHO transition tables. It returns the primary ICD-11 code plus any WHO-documented alternatives, and null (never a guess) when the ICD-10 category is not in the table.

SNOMED CT (retired in 2.0.0)

Until 1.18.x this server shipped five SNOMED CT tools (snomed_search, snomed_concept, snomed_hierarchy, snomed_descriptions, snomed_ecl) plus map_snomed_to_icd10, all off by default behind ENABLE_SNOMED_TOOLS, and map_loinc_to_snomed (guidance only). Version 2.0.0 removed all seven, the Snowstorm client and the flag, and snomed is no longer an accepted value of find_equivalent, validate_codes, terminology_versions or terminology_diff (a request naming it gets a validation error listing the accepted values).

Why: the public IHTSDO Snowstorm host behind the tools has returned HTTP 410 since 2026-05; SNOMED CT content needs a license that depends on the user's country; usage was zero; and the half-present terminology misled every third-party catalog that described this server.

If you need SNOMED CT: pacharanero/sct serves it locally from your own licensed release, and FHIR terminology servers (e.g. tx.fhir.org, a public HL7 test server; CSIRO Ontoserver) expose $lookup/$expand — under your own SNOMED license.

Terminology Licenses

The MIT license covers the server code and server-maintained metadata only — not the terminology content served through it, and not the two bundled datasets (cid10.json, icd10-to-icd11.json), which remain under their own terms. The consolidated notice ships with the package as NOTICE.md; every tool response carries a per-source provenance block with the applicable license.

ICD-11 (WHO)

ICD-11 content is provided under the Creative Commons Attribution-NoDerivatives 3.0 IGO license (CC BY-ND 3.0 IGO), per the ICD-11 Terms of Use and License Agreement.

• Required citation: "International Classification of Diseases, Eleventh Revision (ICD-11), World Health Organization (WHO) 2019 https://icd.who.int/browse11. Licensed under the Creative Commons Attribution-NoDerivatives 3.0 IGO licence (CC BY-ND 3.0 IGO)."
• This server always serves ICD-11 codes and titles together with their URIs, verbatim; non-English labels are WHO's own official translations (never machine-translated)
• WHO may terminate the license at any time by notice (§4.7)
• API access requires registration at https://icd.who.int/icdapi

WHO ICD-10 → ICD-11 transition tables (bundled)

Format conversion (TSV → JSON, content unaltered) of the tables WHO publishes within the ICD-11 release. © World Health Organization, under the ICD-11 Terms of Use — not under this project's MIT license. WHO's guidance: the tables show correspondence between revisions and "are not intended for directly converting data from one revision to the other."

CID-10 V2008 (DataSUS / CBCD, bundled)

© World Health Organization; Brazilian Portuguese translation © CBCD / Faculdade de Saúde Pública da USP; electronic files published by DataSUS (Ministério da Saúde do Brasil). DataSUS/CBCD permission: developers may use the files with due credit and at no charge — this server serves them free with credit in every response. Not under this project's MIT license.

LOINC

This material contains content from LOINC (http://loinc.org). LOINC is copyright © Regenstrief Institute, Inc. and the Logical Observation Identifiers Names and Codes (LOINC) Committee and is available at no cost under the license at http://loinc.org/license. LOINC® is a registered United States trademark of Regenstrief Institute, Inc.

• Served via the free NLM Clinical Tables API; every code comes with its official display name
• Terms with third-party copyright are served with their notice passed through verbatim

RxNorm

RxNorm is produced by the U.S. National Library of Medicine; the RxNav APIs serve non-proprietary, public-domain RxNorm content free of charge.

This product uses publicly available data from the U.S. National Library of Medicine (NLM), National Institutes of Health, Department of Health and Human Services; NLM is not responsible for the product and does not endorse or recommend this or any other product.

ATC (via NLM RxClass)

ATC classification © WHO Collaborating Centre for Drug Statistics Methodology (https://atcddd.fhi.no/), retrieved via NLM RxClass and served verbatim. This server never redistributes the WHOCC ATC/DDD index.

MeSH

MeSH is a U.S. government work served free of charge under the NLM Terms and Conditions. The NLM statement above (RxNorm) applies to this channel as well.

Surface fingerprint, checkable from the registry

Every release publishes, in its MCP Registry entry, the sha256 of the surface this version serves (initialize, tools, resources, prompts) and which methods answer without a credential. A surface change without a version bump fails the build; the fingerprint in the registry lets a client check the same thing on its own side. Canonical form and procedure: SPEC.md of @sbissoli/mcp-surface. To check it yourself (Node 18+, no dependencies):

curl -sO https://raw.githubusercontent.com/SidneyBissoli/mcp-br-commons/main/packages/mcp-surface/exemplos/verify.mjs
node verify.mjs io.github.SidneyBissoli/medical-terminologies-mcp

API Rate Limits

This server implements rate limiting to respect API providers:

Development

Building from source

git clone https://github.com/SidneyBissoli/medical-terminologies-mcp.git
cd medical-terminologies-mcp
npm install
npm run build

Running locally

npm start

Testing with MCP Inspector

npx @modelcontextprotocol/inspector node dist/index.js

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

1. Fork the repository
2. Create your feature branch (git checkout -b feature/AmazingFeature)
3. Commit your changes (git commit -m 'Add some AmazingFeature')
4. Push to the branch (git push origin feature/AmazingFeature)
5. Open a Pull Request

Author

Sidney da Silva Pereira Bissoli

• GitHub: @SidneyBissoli

License

This project is licensed under the MIT License - see the LICENSE file for details.

Note: While this software is MIT licensed, the medical terminologies accessed through it have their own licenses (see Terminology Licenses above).

Acknowledgments

• WHO for the ICD-11 API
• Regenstrief Institute for LOINC
• U.S. National Library of Medicine for RxNorm and MeSH
• Anthropic for the Model Context Protocol

Support

If you encounter any issues or have questions:

• Open an issue on GitHub
• Check existing issues for solutions

Made with love for the medical informatics community