What this is
This is an open working group, starting now. Members will implement and test Model Context Protocol (MCP) tools over National Provider Directory (NPD) data, draft specifications for the group to review, and check the tools against use cases that apps, payers, providers and vendors bring. Anyone can join, and joining needs no account.
This is independent open-source work by AINPI. It is not produced, reviewed, or endorsed by CMS or the NPD team.
The long-term aim is that a provider corrects their information once, and every organization that holds a copy is told.
Why MCP tools, not only FHIR APIs
What is the difference between an MCP tool and a FHIR API? A FHIR API exposes resources and a search syntax. The caller has to know which resources to query and how to join them. An MCP tool is a described task, such as "find this organization's endpoints" or "check whether this NPI is excluded". The tool states its inputs and outputs in plain terms, and it does the joins itself.
Can one tool use more than one source?
Yes. lookup_organization already combines NDH Endpoint records with EHR vendors' published endpoint lists and labels each endpoint by source. A tool could add NPPES, payer directories and the exclusion lists the same way. With FHIR APIs alone, each caller writes that combination again, and the copies drift apart.
Who can call these tools? Applications call them the same way AI agents do. An integration engine, a credentialing app and a chat assistant can call the same tool and get the same answer, so a fix lands in one place.
Does this replace FHIR? No. FHIR stays the data format underneath, and the NPD stays the source of record. Today's tools read data derived from the NPD's FHIR bulk files. The proposed tools would read FHIR APIs directly and return FHIR resources where a caller wants them.
How does a caller know whether to trust an answer?
Each tool should return its evidence with the answer: which source said what, and the date of that source. lookup_organization does part of this today: every endpoint carries its source, ndh or vendor_file, and the response carries the NDH release date. A caller can then decide whether a record from three months ago is good enough for the task at hand.
Tool goals
The step numbers refer to the numbered flow diagram (docs/NPD_Query_Flow.png) in the NPD Endpoint Feedback API proposal, version 0.2, shared in the NPD community: ftrotter-gov/CMSPayerEndpointSeed, docs/npd-how-did-it-go-api.md. The proposal calls itself an initial design, not an existing NPD API, and this page makes no claim about whether CMS will adopt it.
| Goal | NPD flow step | Status | Today |
|---|---|---|---|
| Discover: find an organization or practitioner and its endpoints | Steps 1 to 2 | Partly available | lookup_npi (a practitioner, by NPI), lookup_organization (an organization's endpoints, by NPI). No search by name yet |
| Validate: check status, such as Medicare enrollment, OIG and SAM exclusions, and NPI validity | Step 5 | Partly available | check_npi_cohort: OIG and SAM exclusions and the NPI check digit, for practitioner NPIs in the NDH, as of the last weekly run. No Medicare enrollment check yet |
| Report how an exchange went | Step 7 | Proposed | None yet. A tool would wrap the proposed POST /api/v1/interaction-reports (NPD Endpoint Feedback API proposal, v0.2), with an AINPI sandbox for testing. No NPD-hosted version exists |
| Update once, route to everyone | Not in the flow | Proposed | None yet. This is the hardest goal |
The tools available today run on the AINPI MCP server at https://ainpi.dev/api/mcp. The proposed tools do not exist yet. The group will draft and review them first.
What MCP does not solve
MCP does not settle authentication or trust between organizations. A tool can carry credentials, but the parties still have to agree on who may call what.
MCP does not fix data quality at the source. A tool that reads a wrong address returns that wrong address, with its source attached.
MCP does not decide who may update what. The update goal needs an agreement on which party may change which field, and on how a change is checked before others receive it.
How the group works
Sign up on this page. The first piece of work is to implement the tools above and test them against real directory data at the upcoming NPD community connectathon. AINPI takes part as a participant, not an organizer. The group reviews drafts in the open on GitHub in FHIR-IQ/AINPI. Anyone can propose a use case.
Independent open-source work by AINPI. Not produced, reviewed, or endorsed by CMS or the NPD team.