npi-providers-mcp-server

Public Hosted Server: https://npi-providers.caseyjhand.com/mcp

Overview

The US healthcare provider directory from the NPI Registry. Search doctors and organizations by name, location, and specialty in plain language, decode any NPI to its provider record, and browse the NUCC specialty taxonomy.

Tools

Resources

All resource data is also reachable via tools; the resources are convenience twins for resource-capable clients.

The bundled npi-providers-guide skill is available at skill://npi-providers-guide/SKILL.md. It covers specialty cohorts, provider identity, and the limits of exhaustive directory searches.

Related MCP server: npi-verify

Capability reference

npi_search_providers tool

• Inputs: name_search or first_name / last_name (individuals), organization_name (organizations), provider_type, city / state / postal_code (practice addresses only), and a plain-language specialty resolved through the bundled NUCC taxonomy, or an exact taxonomy_description (never both). At least one criterion is required, state alone is rejected, and individual and organization criteria can't be mixed. A trailing * needs at least 2 leading characters; postal_code takes 5 or 9 digits or a 2–9 digit prefix ending in *. Positive limit (clamped to 200, default 10), skip 0–1000. Optional use_first_name_alias controls upstream first-name variants; omitted defaults to true for a non-wildcard first name. When supplied for a wildcard/no-first-name search, the option is reported as inapplicable

• Output: compact rows (npi, name, primary taxonomy, primary practice city / state / postalCode, status), with matchedLocation / matchedOtherName naming the practice location or other name a row matched through, and resolvedTaxonomies / appliedTaxonomyDescription echoing the specialty resolution. appliedFirstNameAlias reports the effective boolean only for a non-wildcard first name. No true total is reported: a full page names nextPage, and the terminal window (skip 1000, limit 200) returns continuationPostalCodes that continue the search

• Typed error reasons: no_search_criteria, conflicting_specialty, mixed_provider_criteria, unresolved_specialty, invalid_search_field

npi_get_provider tool

• Accepts npis: a single NPI or up to 100 raw entries, each exactly 10 digits. Deduplicates, checks every check digit, then looks up the first 10 distinct valid NPIs; optional unprocessed lists the remaining valid NPIs in order. With NPI_CLAMP_LIMITS=false, the raw array maximum is 10

• Partitions the batch into found (decoded records: taxonomies with license, practice addresses, credential, identifiers, FHIR/Direct endpoints, status — only LOCATION address rows for individual providers, so their mailing address is withheld), notFound (confirmed absence), errored (upstream failure — retry these), and invalid (failed the check digit, never looked up)

• Typed error reasons: invalid_npi_format (every NPI failed the check digit), none_found (every NPI looked up is a confirmed absence); an upstream failure surfaces as that underlying error instead. Error messages and structured error data retain every invalid NPI and any unprocessed remainder

npi_lookup_taxonomy tool

• Modes: resolve (query → matching active codes), get (exact code → full entry with NUCC's notes), browse (walk grouping → classification → specialization, filterable by grouping and NUCC section). Section describes the taxonomy category, not a provider's NPI entity type; use search's provider_type for that. resolve / browse take positive limit (clamped to 50, default 20) and skip (0–1000) and disclose truncated; wrong-mode arguments are rejected before defaults apply

• Every entry carries status (active / inactive) and an inactive code's replacedBy; resolve excludes inactive codes, while get and browse return them. A resolved entry's specialization (or classification when absent) is the exact value npi_search_providers.taxonomy_description accepts

• Definitions and notes contain readable line breaks and list bullets in place of NUCC HTML markup; literal markers and source text remain intact

• Typed error reasons: no_match, missing_argument, wrong_mode_argument

npi://provider/{npi} resource

• Returns the same decoded record as npi_get_provider, for one NPI, as application/json — only LOCATION address rows are kept for individual providers

• npi must be a well-formed 10-digit NPI; invalid_npi_format when it fails the NPI check digit (no registry request), no_record when the registry has none (deactivated or never enumerated)

npi://taxonomy/{code} resource

• Returns the same entry as npi_lookup_taxonomy mode get (NUCC notes, status, and replacedBy included; inactive codes stay readable), as application/json; cached publicly for 24 hours

• code must match ^\d{3}[A-Z0-9]{6}X$; no_match when no entry exists for the code

Features

Built on @cyanheads/mcp-ts-core: stdio and Streamable HTTP transports, pluggable auth (none / jwt / oauth), swappable storage (in-memory, filesystem, Supabase, Cloudflare KV/R2/D1), structured logging with optional OpenTelemetry tracing.

NPI/NPPES-specific:

• Keyless — runs against the public CMS NPPES NPI Registry API (v2.1) with no API key or account

• Bundled NUCC taxonomy — the 883-code Healthcare Provider Taxonomy (v26.1) ships in the image and loads into an in-memory index at startup, so specialty resolution and code lookups work fully offline with no second upstream

• Specialty resolution turns a vague term ("heart doctor") into the precise taxonomy description the registry filters on, and echoes the match back for the agent to verify

• Detects the registry's quirk of returning HTTP 200 with an Errors[] body on validation failure and maps it to typed, recoverable error reasons

Agent-friendly output:

• Provenance on search — the resolved taxonomy and the exact taxonomy_description sent to the registry are echoed back, so agents can see what was actually searched and re-run with a different code

• Honest pagination — the returned count is disclosed as the page size, never a fabricated grand total; a full page names the exact next page, and a broad query that hits the 1200-match reachable ceiling gets the postal-code prefixes that continue it

• Graceful partial failure — npi_get_provider returns per-NPI found / notFound / errored / invalid rows instead of failing the whole batch

Getting started

Connect to the hosted instance, no install:

claude mcp add --transport http npi-providers-mcp-server https://npi-providers.caseyjhand.com/mcp

Or run it locally:

claude mcp add npi-providers-mcp-server -- npx -y @cyanheads/npi-providers-mcp-server

For other clients, add this to the MCP config file:

{
  "mcpServers": {
    "npi-providers-mcp-server": {
      "command": "npx",
      "args": ["-y", "@cyanheads/npi-providers-mcp-server"]
    }
  }
}

No API key is required. For Docker or a source build, see Running the server.

Configuration

No required variables — the server runs out of the box against the keyless NPPES registry. All variables below are optional overrides.

See .env.example for the full list of optional overrides.

Running the server

Local development

Requires Bun v1.4.0 or higher (or Node.js v24+). The NPPES registry is keyless and the NUCC taxonomy is bundled.

• Clone and install:

  git clone https://github.com/cyanheads/npi-providers-mcp-server.git
  cd npi-providers-mcp-server && bun install
  cp .env.example .env   # optional NPPES base URL and timeout overrides

• Build and run:

  # One-time build
  bun run rebuild
  
  # Run the built server
  bun run start:stdio
  # or
  bun run start:http

• Run checks and tests:

  bun run devcheck   # Lint, format, typecheck, security
  bun run test       # Vitest test suite
  bun run lint:mcp   # Validate MCP definitions against spec

Docker

docker build -t npi-providers-mcp-server .
docker run --rm -e MCP_TRANSPORT_TYPE=http -p 3010:3010 npi-providers-mcp-server

The Dockerfile defaults to HTTP transport, stateless session mode, and logs to /var/log/npi-providers-mcp-server. OpenTelemetry peer dependencies are installed by default — build with --build-arg OTEL_ENABLED=false to omit them.

Project structure

Development guide

See CLAUDE.md/AGENTS.md for development guidelines and architectural rules. The short version:

• Handlers throw, framework catches — no try/catch in tool logic

• Use ctx.log for request-scoped logging, ctx.state for tenant-scoped storage

• Register new tools and resources in the createApp() arrays

• Wrap external API calls: validate raw → normalize to domain type → return output schema; never fabricate missing fields

Data attribution

Provider data from the CMS NPPES NPI Registry (public domain). The bundled specialty codes are the NUCC Health Care Provider Taxonomy, © American Medical Association on behalf of the National Uniform Claim Committee (NUCC), redistributed unmodified beyond formatting under the NUCC permission. See NOTICE.

Contributing

Issues are welcome. Run checks and tests before submitting:

bun run devcheck
bun run test

License

Apache-2.0 — see LICENSE for details.