AI Agents Are Starting to Query Healthcare Data Directly

Two forces are converging on US healthcare data at once, and neither is slowing down. First, federal regulators are forcing payers to expose patient, provider and prior-authorization data through a standard healthcare data API — not as a nice-to-have, but as a compliance deadline. Second, the way AI agents actually reach that data is being standardised too, as the Model Context Protocol moves from an Anthropic-only project to a vendor-neutral open standard. Put together, US healthcare information is becoming more machine-readable, and more machine-queryable, than at any point before. For developers building on top of it, that raises a practical question: whose data should an agent trust?

CMS-0057-F: the healthcare data API mandate

The regulatory push has a name: CMS-0057-F, the Interoperability and Prior Authorization Final Rule, published in the Federal Register on 8 February 2024. It requires impacted payers — Medicare Advantage plans, Medicaid and CHIP programmes, and Qualified Health Plan issuers on the federal exchanges — to stand up four production FHIR APIs by 1 January 2027: a Patient Access API, a Provider Access API, a Payer-to-Payer API, and a Prior Authorization API. These build on established standards, including HL7 FHIR US Core, Da Vinci PDex, CARIN Blue Button, and the Da Vinci PAS, DTR and CRD implementation guides for prior-authorization workflows.

The operational side of the rule already has teeth. Since 1 January 2026, covered payers must give specific reasons for prior-authorization denials and meet fixed decision timeframes — 72 hours for urgent requests, seven days for standard ones. From 1 January 2027, payers must also fold prior-authorization data (excluding drug-related requests) into the Patient Access API. This is not a roadmap item sitting in a slide deck somewhere. It is a live, dated legal obligation to expose structured healthcare data through a documented healthcare data API.

MCP's road from Anthropic project to shared standard

The second force is about how software actually reaches that data. Anthropic introduced the Model Context Protocol in November 2024 as an open standard for connecting AI models to external tools and data sources — a common language so an agent doesn't need a bespoke integration for every API it touches. On 9 December 2025, Anthropic donated MCP to a newly formed Agentic AI Foundation, a directed fund under the Linux Foundation, co-founded with Block and OpenAI, with Google, Microsoft, AWS, Cloudflare and Bloomberg joining as supporting members. At the time of that announcement, Anthropic cited more than 10,000 active public MCP servers, over 75 connectors in Claude's own directory, and combined Python and TypeScript SDK downloads exceeding 97 million a month, with adoption already spanning ChatGPT, Cursor, Gemini, Microsoft Copilot and VS Code.

Handing MCP to a vendor-neutral foundation matters more in healthcare than in most sectors. Medical data is sensitive, heavily regulated, and touched by dozens of systems — EHRs, payers, labs, pharmacy platforms — each of which might connect a different AI assistant. If every vendor invents its own way for an agent to fetch a patient record or look up a code, every one of those integrations has to be built, secured and maintained separately. A single, openly governed protocol means a provider, payer or reference-data platform can expose one integration and have it work with any MCP-compatible agent, rather than negotiating a bespoke connector with each AI vendor in turn — which is the whole point of standardising the healthcare data API layer in the first place.

Claude for Healthcare: what's reported so far

Where this is heading in practice is still coming into focus, and it's worth being precise about what is and isn't confirmed. Trade press — including Fierce Healthcare and Becker's Hospital Review — reported that Anthropic launched 'Claude for Healthcare' at the J.P. Morgan Healthcare Conference in January 2026 ('JPM26'), describing it as a HIPAA-compliant toolkit connecting Claude to US medical data infrastructure for health systems, payers and patients. Reported integrations include HealthEx for consumer health-record access, beta rollouts covering Apple Health and Android Health Connect, and connectors for Medidata, ClinicalTrials.gov and ToolUniverse. This reportedly followed an earlier 'Claude for Life Sciences' launch covered in October 2025.

None of this comes from a verified primary Anthropic page at the time of writing — it's press-sourced context, not a confirmed specification. But the direction it points in is consistent with everything above: AI agents connecting directly into clinical and administrative data, via a standard protocol, rather than working from what a model happened to memorise during training.

Why a live reference layer beats a model's memory

This is exactly where the practical risk sits for anyone building on top of healthcare data today. Large language models are trained on a fixed snapshot of the internet. NPIs are issued and deactivated continuously; ICD-10-CM codes are revised annually; provider addresses, specialties and organisational affiliations change constantly. Ask a model to recall an NPI or a diagnosis code from memory and it will often produce something plausible-looking rather than something correct — a classic hallucination risk, except the output looks like a clean ten-digit number or a valid-format code rather than obviously wrong prose. In administrative and billing contexts, a wrong-but-plausible NPI or ICD-10 code doesn't fail loudly. It fails quietly, downstream, as a rejected claim or a misrouted referral.

Consider a developer building an AI scheduling and eligibility assistant for a multi-provider clinic group. Early prototypes let the model answer 'what's Dr. Patel's NPI?' straight from its own training data — it was fast, and it was right often enough to pass a demo. In production, it started returning NPIs for providers who had moved practices, or numbers that were subtly wrong, and eligibility checks failed against the real payer system on the other end. The fix wasn't a better prompt. It was routing that lookup to a live, MCP-connected NPI lookup instead of the model's memory — the same principle covered in how AI agents should handle medical billing data. Once the agent queried current NPPES data through an MCP server rather than guessing, the wrong-number failures disappeared.

This is the case for treating structured, live, verifiable healthcare reference data as infrastructure an agent calls, not a fact a model recalls. Patientary exists for exactly that gap: live NPI lookup sourced from CMS NPPES data, the full ICD-10-CM code set, and — for builders wiring agents up directly — an MCP server that lets Claude and other MCP-compatible agents query that data the same way a person would use the site. If you're comparing options, how NPI lookup tools stack up and validating an NPI number are good starting points; for the wider landscape, see the best healthcare APIs for developers.