The archive is open: every release of the directory since May, kept, free
The US government publishes its list of health care providers as one big file, and it keeps only the newest copy. When a new version comes out, the old one is gone. If you want to know whether the list got better or worse, you need someone to have saved both.
We saved both. As of this week the archive is on Databricks Marketplace, and anyone can use it for free.
What is in it
Two releases so far, 2026-05-08 and 2026-08-20, as six tables: one each for practitioners, where they work, organizations, organization links, places and web addresses. That is 54,162,643 rows across the two. Every row keeps the original government record exactly as published. Each also has plain columns for the things people look up most, like the NPI, the state and the phone number, so an ordinary question does not need any parsing. The same code extracted both releases, so when you compare them you are comparing the data and not two programs.
Each table is split by release date. Comparing two versions is one line of a query, not a download.
A worked example is attached to the listing as a notebook. It walks through the biggest change between the two releases. The records saying where a clinician works more than doubled. The share of clinicians the directory can place moved only a few points, because most of the new records went to people it already described.
Two ways to read it
If you use Databricks, open the listing and click Get access. The tables land in your workspace.
If you do not, the same tables are available through OpenSharing (formerly Delta Sharing) with a small credential file and one Python package. No Databricks account is needed. Ask for a credential at ainpi.dev/databricks, which also has the reference architecture and the quickstart.
Four organizations connected in the first three days.
One warning before you compare releases
Practitioner records carry the NPI in their id, so they line up across releases. Web address and place records do not: CMS gives them new random ids on every export. If you match those on id, every single one looks changed, whether or not it did. Match web addresses on the address itself. The table notes say the same thing and the notebook shows it.
New on the site since the last update
Explore the directory by place. Pick a state, then a county, then a ZIP. See how many providers the directory lists there, how many it can say work somewhere, and how that splits by profession. All fifty states and DC.
Find listed care near you. Enter a ZIP or share your location and see what the directory has nearby. Built on the directory's own coordinates, with no mapping service behind it.
A primer. What each identifier means, what each record type holds, how they join, and which joins are safe. Written for someone capable who does not already live in this data. It ends with a scoreboard of six directory measures and where each stands today.
Related work. Other people measuring provider and insurer directory data from angles we do not cover, with links to what they publish. Also the CMS and HL7 venues where this work is discussed.
An answer engine for AI agents. The site now speaks the protocol AI assistants use to call tools. An agent can look up an NPI, check whether a provider appears in the federal exclusion cohort, read any finding, or pull a state's audit. Free, no key, rate limited.
Specialty by workplace. We asked whether a clinician's listed specialty changes depending on which organization the record is attached to. Of 445,527 practitioners the directory places at two or more organizations, 82,977 carry a different specialty at differently-named ones (18.6%). It is published as exploratory, because it answered a question that came up rather than one we registered in advance.
The homepage map is now measured. The treemap on the front page used to run on a placeholder seed. It now shows real numbers from the 2026-08-20 release for every state and specialty.
A correction
The Marketplace listing said that practitioner and organization records both carry the NPI in their id. That was true for practitioners and for the provider half of organizations. It was false for the other half: 2,199,519 of the 4,402,671 organization records in August are tax records with a random id. The listing is fixed. We caught it while checking this update, which is what the checking is for.
What comes next
We hold the releases and we will keep adding them as CMS publishes. The next step is to publish a standing set of measures with one point per release, so "is the directory improving" becomes a chart instead of a question. The set is being chosen now.
Which quantity would you want tracked release over release? Reply to this email, or open an issue on the repository. The ones people ask for get built first.
Check our work
- The listing: Databricks Marketplace
- The release-over-release numbers: release-deltas.json
- The notebook source: ainpi_archive_quickstart.py
- How the numbers are made: methodology
This is a measurement of a federal file, not a source of truth about any individual provider. Check NPPES, the OIG exclusion list and SAM.gov before acting on a named record.
— Eugene Vestel, FHIR IQ