@shaddyt/clinical-reference-mcp

Drug, prescription, and pharmacology reference tools for AI engineers, sourced from openFDA and RxNorm/RxNav.

An MCP (Model Context Protocol) server that gives AI agents and command-line callers a small, citable set of drug-reference tools backed by U.S. regulator data: openFDA for FDA-approved labels and adverse-event reports, and RxNorm/RxNav for drug naming and therapeutic-class lookups. It is built for AI engineers who want their healthcare-adjacent features grounded in authoritative sources rather than improvised by a language model. It is not clinical decision support, not an FDA-cleared device, and not for patient-facing use. The project exists because there is a real gap in safe, citable healthcare tooling for AI builders, and because someone with both a medical degree and a production engineering background is well placed to draw the line between "useful reference" and "irresponsible advice."

Try it without installing

Hosted demo: https://clinical-reference.shaddyt.space — pick a tool, enter a drug name, see live results from openFDA / RxNorm / RxNav. No install required. Or point any MCP client that speaks Streamable HTTP at https://clinical-reference.shaddyt.space/mcp.

Safety & scope

Returns regulator-published data only. Not for clinical use - for developer reference.

(Verbatim from src/lib/safety.ts — the same disclaimer text appears in every response payload, every HTTP response header, every CLI footer, and every MCP tool description.)

This server is:

• Not clinical decision support. It does not synthesize diagnoses, treatment plans, or prescribing recommendations. Tools return structured slices of regulator-published documents.
• Not FDA-cleared. No part of this software has been reviewed or approved by any regulatory body.
• Not for patient-facing use. No tool accepts patient identifiers, and no response should be shown directly to a patient as medical guidance.

Security reports go to the maintainer per SECURITY.md.

Quickstart

Three paths, each working from a clean install with no API keys or configuration.

Use with Claude Code

Add this to your Claude Code MCP configuration:

{
  "mcpServers": {
    "clinical-reference": {
      "command": "npx",
      "args": ["-y", "@shaddyt/clinical-reference-mcp"]
    }
  }
}

Restart Claude Code. The six tools become available to the agent.

Use with Claude Desktop

Add the same block to claude_desktop_config.json (locations: ~/Library/Application Support/Claude/claude_desktop_config.json on macOS, %APPDATA%\Claude\claude_desktop_config.json on Windows):

{
  "mcpServers": {
    "clinical-reference": {
      "command": "npx",
      "args": ["-y", "@shaddyt/clinical-reference-mcp"]
    }
  }
}

Restart Claude Desktop. The six tools become available in the conversation.

Use as a CLI

Every tool is also exposed as a subcommand of a standalone CLI, useful for shell scripting, manual lookups, or piping into jq:

npx -p @shaddyt/clinical-reference-mcp clinical-reference lookup-drug aspirin
npx -p @shaddyt/clinical-reference-mcp clinical-reference get-drug-label aspirin --sections warnings
npx -p @shaddyt/clinical-reference-mcp clinical-reference check-interactions warfarin aspirin

Add --json for the raw response envelope; add --no-disclaimer to suppress the human-readable footer (the JSON envelope still carries the disclaimer field).

Available tools

Tool	Purpose	Source
lookup_drug	Resolve a name to RxNorm canonical data	RxNorm
get_drug_label	Fetch FDA-approved label sections	openFDA
check_interactions	Surface label warnings mentioning multiple drugs	openFDA
find_alternatives	List other drugs in the same therapeutic class	RxNorm RxClass
lookup_adverse_events	Top reported adverse events from FAERS	openFDA
get_dosing_reference	Published dosing text from FDA labels	openFDA

Every successful response carries the disclaimer text plus a citation block listing the upstream sources used to produce it.

Architecture

The codebase is layered. Outbound HTTP funnels through one chokepoint at src/lib/http.ts, which is the only place in the project that calls fetch. Above that sit per-source clients (openfda.ts, rxnorm.ts), then normalization (normalize.ts), then pure tool handlers under src/server/tools/, then the transport-agnostic MCP server factory in src/server/server.ts, and finally three entry points: stdio, Streamable HTTP, and the CLI.

              ┌──────────────┐  ┌──────────────┐  ┌──────────────┐
              │ stdio bin    │  │ HTTP bin     │  │ CLI          │
              └──────┬───────┘  └──────┬───────┘  └──────┬───────┘
                     │                 │                 │
                     └────── server.ts (registry + dispatch) ──────┘
                                         │
                          tools/ (pure handler functions)
                                         │
                       lib/ (clients, normalize, respond, safety)
                                         │
                              http.ts (single fetch chokepoint)

CONTRIBUTING.md walks through the structure in more detail.

Data sources & attribution

• openFDA — the U.S. Food and Drug Administration's public API for FDA-approved drug labels and FAERS adverse-event reports. Public domain, freely redistributable.
• RxNorm — the U.S. National Library of Medicine's normalized vocabulary for clinical drug names. Public domain, freely redistributable.
• RxNav — NLM's API surface over RxNorm and RxClass (therapeutic-class lookups). Same terms.

Formal attribution lives in NOTICE. This project is independent of, and not endorsed by, the FDA or NLM.

Versioning & stability

Semantic Versioning. Currently 0.1.0.

Tool names, tool input schemas, and response envelope shapes are stable from 0.1 forward — breaking changes here will land in a major version. Internal modules under src/lib/ are not part of the public API and may change in any release; the public library export is intentionally narrow (safety constants, citation builders, response helpers, types). See src/index.ts for the full public surface.

Contributing

Issues and PRs welcome. One scope rule worth flagging up front: PRs that introduce clinical reasoning, treatment recommendations, allergy logic, or prescription validation will be declined. This server wraps regulator-published data; it does not synthesize clinical decisions. Full process notes in CONTRIBUTING.md.

License

Apache License 2.0. Attribution and third-party notices in NOTICE.

About the author

Dr. Shadrack Omary is a Medical Doctor and Senior Full Stack Engineer based in Tanzania. He is Co-Founder & CTO at Hisa Hub, an AI-driven market intelligence platform, and Engineering Lead at Zandbox. The combination of clinical training and production engineering shaped this project: tools for healthcare AI builders, designed by someone who has both modeled clinical workflows in the EHR and shipped TypeScript at scale.

More work at shaddyt.space.