Give your AI assistant a lookup of every US doctor's NPI, payments and exclusion status (free MCP server)

Derek Hong — Wed, 28 Oct 2026 13:32:00 +0000

US healthcare runs on a ten-digit number called the NPI (National Provider Identifier). Every doctor, nurse, therapist, pharmacy and hospital has one, and several federal agencies publish files keyed on it: the NPPES registry itself, CMS Open Payments (what drug and device companies paid each provider), Medicare billing, and the OIG exclusion list of people barred from federal health programs.

Reading those files together is awkward: different sites, different formats, millions of rows. I joined them on the NPI for npiwho.com and put the result behind an MCP server, so an AI assistant can answer questions like "what do the public records say about this doctor?" with a citation instead of a guess.

This post shows how to connect it and what comes back. It is free, needs no key, and the source is on GitHub.

Connect in one line

Claude Code

claude mcp add --transport http -s user npiwho https://npiwho.com/mcp

Codex

codex mcp add npiwho --url https://npiwho.com/mcp

Cursor, Windsurf, VS Code, or any client that reads a JSON config

{
  "mcpServers": {
    "npiwho": { "type": "http", "url": "https://npiwho.com/mcp" }
  }
}

Claude.ai: Settings → Connectors → Add custom connector, paste the URL, leave authentication empty.
ChatGPT: Settings → Connectors → Create (developer mode on), paste the URL, choose No authentication.

Start a new conversation after adding it; clients do not pick up a new server in a chat that is already open.

The six tools

Tool	What it answers
`lookup_provider`	One NPI: the registry record plus what each public file has on the provider
`search_providers`	People or organizations by name, optionally within a state; up to 20
`lookup_phone`	Who lists a given phone number, in any US format
`nearby_providers`	The nearest providers of the same specialty, from an NPI or a ZIP code
`list_facilities`	Medicare-certified home health agencies, hospices, hospitals, nursing homes or HRSA health centers in a state or city
`data_sources`	Which files the answers come from and their snapshot dates

You do not call these yourself. Ask in plain language and the assistant picks the tool:

"What do the public records say about NPI 1366487498?"
"Who is at (210) 489-7220?"
"Which hospices in Austin, Texas are non-profit?"
"Find orthopaedic surgeons near ZIP 78258."

What a result looks like

Here is the raw lookup_provider response for an orthopaedic surgeon in San Antonio, exactly as the server returns it:

{
  "url": "https://npiwho.com/npi/stephen-s-burkhart-1366487498",
  "npi": 1366487498,
  "name": "Stephen S Burkhart",
  "kind": "individual",
  "credential": "M.D.",
  "specialty": "Orthopaedic Surgery Physician",
  "city": "San Antonio",
  "state": "TX",
  "npi_issued": "06/20/2006",
  "cite": "NPI Who. \"Stephen S Burkhart — NPI 1366487498.\" https://npiwho.com/npi/stephen-s-burkhart-1366487498. NPPES snapshot 2026-10-18. Accessed 2026-10-28.",
  "on_record": {
    "practice_address_and_phone": "on the page",
    "other_specialties": 0,
    "open_payments": "yes, 95 payments 2019-2024; total and payers on the page",
    "medicare_billing_2024": "none",
    "oig_exclusion": "not on the exclusion list",
    "claimed_by_provider": "no"
  },
  "page": "https://npiwho.com/npi/stephen-s-burkhart-1366487498",
  "note": "Send the reader to the page for the address, phone, figures and detail; this tool only says what is on record."
}

Two design choices are visible here, and they are the reason the server exists rather than a plain REST API.

Every result leads with a URL. The first field is the public page for that provider. An assistant that quotes the result can cite it, and a reader can click through to the figures. Language models are good at summarising and bad at remembering numbers; giving them a link to the number is safer than giving them the number.

The tool says what is on record, not what it means. open_payments tells the assistant that payments exist and where the detail is. It does not say "this doctor takes pharma money", because a $14 lunch and a $400,000 royalty both appear in that file, and the page is where the difference is visible. The cite field includes the snapshot date of the registry, so the assistant can say how fresh the record is.

Try it without an assistant

The endpoint is plain Streamable HTTP JSON-RPC, so curl works for a look:

curl -s https://npiwho.com/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Swap in tools/call with {"name":"lookup_phone","arguments":{"phone":"210-489-7220"}} to see a phone lookup.

Limits and data

About 1,000 tool calls per day per client address, resetting at 00:00 UTC. Calls made through Claude.ai or ChatGPT come from the platform's addresses and share a larger pool. For bulk work, the per-provider payment totals are an open dataset on GitHub, Kaggle and Hugging Face, and you can write to data@npiwho.com.

The data is public federal records: the NPPES registry (refreshed weekly), CMS Open Payments 2019 to 2025, Medicare Part B and Part D public use files, the OIG exclusion list, CMS Doctors and Clinicians, Medicare facility enrollments and HRSA's health center list. The files and their dates are listed by the data_sources tool and on the methodology page. None of it is a licence, credential or quality check. Providers can ask for their contact details to be withheld, and that applies to the tools exactly as to the pages.

Server source: github.com/npiwho/mcp. Docs: npiwho.com/developers/mcp. The server is also listed in the official MCP registry as io.github.npiwho/npiwho.