Seven medical terminologies, one MCP server: a practical ...

If you've ever asked an LLM to "find the LOINC code for procalcitonin" or "list the active ingredients in Janumet," you've probably watched it confidently invent a code that doesn't exist. Medical terminologies are exactly the kind of structured, frequently-updated reference data that language models are bad at memorizing and good at looking up — if you give them the right tool.

medical-terminologies-mcp is a Model Context Protocol server that gives any MCP-compatible client (Claude Desktop, Claude Code, Continue, and others) unified access to seven medical terminology systems:

ICD-11 (WHO International Classification of Diseases, 11th Revision)
LOINC (Logical Observation Identifiers Names and Codes)
RxNorm (NIH normalized clinical drug names)
MeSH (NLM Medical Subject Headings)
ATC (WHO Anatomical Therapeutic Chemical, served via NLM RxClass)
CID-10 (Brazilian Portuguese translation of ICD-10, DataSUS V2008 — bundled)
SNOMED CT (Systematized Nomenclature of Medicine, optional, license required)

Twenty-six tools work out of the box with no authentication for LOINC, RxNorm, MeSH, ATC, CID-10, plus a bundled authoritative ICD-10 → ICD-11 mapping, a cross-terminology batch validator, and versioning + cross-revision diff. ICD-11 live lookup needs free WHO API credentials (a five-minute signup), bringing the default count to 31. SNOMED is gated behind an explicit feature flag and requires an IHTSDO license plus a self-hosted Snowstorm instance — more on why below.

The server also exposes 3 MCP Prompts that orchestrate tool calls into named user actions (find-medical-code, drug-info, cid10-portuguese-lookup) and 3 Resources for in-process reference content (server metadata, CID-10 chapter listing, per-terminology license disclaimers) — MCP clients render Prompts as one-click actions and read Resources by URI.

This post walks through three concrete clinical and research workflows where the server earns its place in the toolbelt.

Setup in 60 seconds

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

Drop that into your Claude Desktop config (%APPDATA%\Claude\claude_desktop_config.json on Windows, ~/Library/Application Support/Claude/claude_desktop_config.json on macOS), restart Claude, and 31 tools appear. WHO credentials are optional — without them the 5 ICD-11 live-lookup tools throw a clear configuration error and the other 26 work fine (including the authoritative ICD-10 → ICD-11 mapping, which uses bundled WHO data and needs no auth).

Prefer a hosted endpoint? The server runs on Cloudflare Workers at the edge — point any Streamable-HTTP MCP client at https://medical-terminologies-mcp.sidneybissoli.workers.dev/mcp and skip the local install. Also listed on Smithery, Glama, and mcpservers.org.

Use case 1: sepsis triage — finding the right LOINC for procalcitonin

A clinical scenario. An ED resident is documenting a sepsis workup and needs the LOINC code for serum procalcitonin to attach to the lab order in the EMR. They could open loinc.org, run a search, scroll through results, and copy the code. Or they could ask their LLM-powered scribe.

You: What's the LOINC code for procalcitonin in serum?

Claude (calling loinc_search):
## LOINC Search Results for "procalcitonin"
Found 23 total results (showing 10):

1. **33959-8** - Procalcitonin [Mass/volume] in Serum or Plasma
   Component: Procalcitonin | Property: MCnc | Timing: Pt | System: Ser/Plas
   Class: CHEM | Status: ACTIVE
2. **75241-0** - Procalcitonin [Mass/volume] in Body fluid
...

The LLM picks 33959-8 and proposes it. The clinician confirms or asks for details:

You: Give me the full LOINC details for 33959-8.

Claude (calling loinc_details):
# 33959-8 - Procalcitonin [Mass/volume] in Serum or Plasma

| Attribute | Value |
|-----------|-------|
| Component | Procalcitonin |
| Property  | MCnc |
| Timing    | Pt |
| System    | Ser/Plas |
| Scale Type| Qn |
| Method    | (none) |
| Class     | CHEM |
| Status    | ACTIVE |

What this replaces in practice: a 30-second context switch from EMR to a browser tab, every time a clinician needs to pin down a specific code. Multiplied across the day, across a team, that's real time.

The same workflow applies to any biomarker, vital sign, or observation: lactate (2524-7), troponin I high-sensitivity (67151-1), SpO₂ (59408-5). The tool is loinc_search for the discovery step, loinc_details for the verification step. Both are deterministic API calls against NLM Clinical Tables — the LLM doesn't memorize codes, it looks them up.

Use case 2: medication reconciliation — breaking down a combination product

A scenario familiar to any inpatient pharmacist or hospitalist: a patient is admitted on Janumet (a brand combo of sitagliptin and metformin), and the team needs to reconcile this against the formulary, check for class-level interactions, and document the active ingredients separately.

You: I need to break down Janumet for the admission med rec.
What are the active ingredients and their drug classes?

Claude (calling rxnorm_search, then rxnorm_ingredients, then atc_classify):

Step 1 — search for Janumet:
**860975** - Janumet (sitagliptin 50 MG / metformin hydrochloride 500 MG Oral Tablet)
   Type: SBD | Synonym: sitagliptin/metformin

Step 2 — get the ingredients:
| RxCUI | Name | Type |
|-------|------|------|
| 593411 | sitagliptin | Single Ingredient |
| 6809 | metformin | Single Ingredient |

Step 3 — get ATC classes for each ingredient:
For sitagliptin (593411): A10BH (Dipeptidyl-Peptidase IV Inhibitors)
For metformin (6809):     A10BA (Biguanides)
                          A10BD (Combinations of oral blood glucose lowering drugs)

Three tool calls, structured data, no hallucination risk. The LLM can compose this into a med rec note, flag that both ingredients sit in the same A10B (oral glucose-lowering) ATC bucket, and surface the ATC codes for any pharmacy system that uses ATC.

What this enables that bare LLMs don't: the chain brand → ingredients → ATC classes is a graph traversal across a real database. The model isn't recalling that sitagliptin is a DPP-4 inhibitor; it's looking it up against the same source the FDA uses.

Use case 3: evidence synthesis — building a precise PubMed search with MeSH

A scenario for a researcher running a systematic review on the impact of primary health care interventions on avoidable hospitalizations. The quality of a literature search hinges on hitting the right MeSH descriptors and qualifiers — not on free-text keyword guesswork.

You: I'm building a search strategy for a systematic review on
primary health care and avoidable hospitalizations. Help me find
the right MeSH descriptors and applicable qualifiers.

Claude (calling mesh_search):
| MeSH ID | Label |
|---------|-------|
| D011320 | Primary Health Care |
| D006297 | Health Services Accessibility |
| D004472 | Hospitalization |
| D006308 | Health Services Misuse |

You: Get me the tree position and allowed qualifiers for Primary Health Care.

Claude (calling mesh_tree, then mesh_qualifiers):
# Tree Numbers for D011320

### Health Care
- `N04.590.233.727`
- `N02.421.143.827`

# Allowed Qualifiers for D011320

| Qualifier ID | Label |
|--------------|-------|
| Q000208 | ethics |
| Q000379 | methods |
| Q000523 | organization & administration |
| Q000592 | standards |
| Q000656 | trends |
| Q000706 | statistics & numerical data |
| Q000941 | utilization |

*Use qualifiers with descriptors like: "Primary Health Care/utilization"*

That last detail — the seven allowed qualifiers, with /utilization and /statistics & numerical data highlighted — is exactly what a precise PubMed search needs. The model now has enough to draft ("Primary Health Care/utilization"[MeSH] OR "Primary Health Care/statistics & numerical data"[MeSH]) AND "Hospitalization/prevention and control"[MeSH] and explain why each leg matters.

The use case generalizes: any research informatics workflow that needs controlled vocabulary mapping (PubMed, Cochrane, OVID) benefits from mesh_search + mesh_qualifiers + mesh_tree. For systematic reviews specifically, the qualifier list is the part that's hardest to remember and easiest to get wrong.

What's new in 1.4.0 (data-integrity release)

Three additions that change the depth of what the server can answer:

map_icd10_to_icd11 now returns an authoritative WHO mapping, not a text-search heuristic. The 2025-01 release of the WHO transition tables is bundled (5.4 MB raw / 0.95 MB gzipped) covering 11,243 ICD-10 categories. 1,461 of those have multiple WHO-documented ICD-11 candidates — the tool surfaces the primary mapping plus all alternatives, with Foundation/Linearization URIs ready to navigate into ICD-11.
validate_codes is a new batch validator. Pass up to 50 { code, terminology } pairs and get back per-item { valid, active, title, replaced_by, source, error }. Designed for retrospective database analysis: flag codes that no longer exist, surface ICD-10 → ICD-11 replacements at scale, grade activity status where the source terminology exposes it (SNOMED's active flag, LOINC's STATUS field).
terminology_versions + terminology_diff for pipeline maintainers. The first lists current version, release date, publisher, and update cadence across all 8 terminologies. The second is guidance-only for terminologies without bundled history — except for ICD-10, where the bundled transition tables let it surface a real cross-revision summary (1:1 mappings vs splits vs average alternatives when split).

Together with the per-tool language parameter added to SNOMED + MeSH, that's four sub-tasks shipped as one data-integrity release. The earlier 1.3.0 added MCP Prompts (orchestration templates) and Resources (in-process reference content) — both visible in clients like LobeChat and Claude Desktop as one-click actions.

What the server doesn't claim to do

A few things worth being explicit about, because the README is honest about them and the LLM should be too:

map_loinc_to_snomed returns guidance, not a mapping. Direct LOINC ↔ SNOMED CT mappings live in UMLS Metathesaurus (license required) or the LOINC SNOMED CT Expression Association files (LOINC license required).
map_snomed_to_icd10 also returns guidance only today. The authoritative source is SNOMED International's ICD-10 Complex Map refset (447562003), which needs a Snowstorm instance with the refset loaded — planned as Phase 13.7 once that's tractable.
SNOMED tools are off by default because the historical public IHTSDO Snowstorm endpoint was retired (HTTP 410 Gone). Operators with an IHTSDO license and a self-hosted Snowstorm instance flip them on with ENABLE_SNOMED_TOOLS=true SNOMED_BASE_URL=....
None of this is a substitute for clinical judgment. It's a lookup layer for already-known codes, not a diagnostic tool.

Under the hood, briefly

For developers curious about the engineering: TypeScript on Node 20+, bundled with esbuild, built around a token-bucket rate limiter (5 req/s for WHO, 10 req/s for NLM, 20 req/s for RxNorm) and exponential-backoff retry with ±25% jitter. WHO OAuth tokens are cached using the actual expires_in from the API response, not a hardcoded TTL. Every default tool declares outputSchema and returns structuredContent alongside markdown — so MCP clients that consume structured data get typed objects, not parsed prose. The 313-test Vitest suite (unit + contract via nock + integration against live APIs gated by env flag) gates CI on PR. A daily integration cron catches upstream API drift — work that already caught three silent production regressions where the API had drifted and the client gracefully returned empty data.

Two transports: stdio (default — Claude Desktop, IDE clients) and Streamable HTTP. The HTTP path runs on Cloudflare Workers at the edge by default — single web-standard fetch handler, ~0.95 MB gzipped including the bundled CID-10 (DataSUS V2008) and ICD-10 → ICD-11 (WHO 2025-01) datasets, well within Cloudflare's 3 MB compressed script limit. Same source tree; per-isolate cache + rate limiter (KV + Durable Object swap available when traffic crosses the threshold).

It's MIT licensed. The medical terminology content has its own licenses, all linked in the README and surfaced as the info://licenses Resource.

Try it

npm: https://www.npmjs.com/package/medical-terminologies-mcp
GitHub: https://github.com/SidneyBissoli/medical-terminologies-mcp
Hosted endpoint: https://medical-terminologies-mcp.sidneybissoli.workers.dev/mcp (Streamable HTTP, Cloudflare Workers)
Smithery: https://smithery.ai/server/@SidneyBissoli/medical-terminologies-mcp
Glama: https://glama.ai/mcp/servers/SidneyBissoli/medical-terminologies-mcp
mcpservers.org: https://mcpservers.org/servers/sidneybissoli/medical-terminologies-mcp
MCP Registry: io.github.SidneyBissoli/medical-terminologies-mcp

WHO API credentials (free): https://icd.who.int/icdapi.

Issues, PRs, and use-case reports welcome — especially from clinical informatics teams, research informatics groups, and public-health data analysts using LLMs in real workflows. There's a real gap between "LLM as scribe" and "LLM with reliable terminology access," and closing it is what this server is for.

推荐订阅源

DEV Community