Where can an AI agent get US healthcare facility data?
zReports Data is the official records of every US health facility, for AI agents: hospital, nursing home, outpatient surgery center, inpatient psychiatric facility, inpatient rehabilitation facility, hospice, dialysis center, home health agency, fertility and ivf clinic, addiction treatment center. One MCP endpoint (https://api.zreports.org/mcp) for Claude, ChatGPT and any MCP client, or one REST call with an OpenAPI document. Every figure is the government's own, quoted with its source, version and date. Where the file is silent, so are we. A free key is one email away on https://zreports.org/api: 500 calls a day, attribution to zReports required. Inside Claude or ChatGPT no key is needed.
The file is the publisher's own: CMS Care Compare and the Provider of Services file for hospitals, surgery centers, nursing homes, hospice, dialysis, home health, psychiatric and rehab facilities; the CDC's National ART Surveillance System for fertility centers; SAMHSA's locator for addiction treatment; state licensing files for child care. Each category is refreshed on the publisher's own cadence and every answer carries the dataset, the version and the as_of date.
Four tools cover the questions agents ask: find_facility (name words, city, state, zip, category), get_record (the publisher's row verbatim, with highlight lines and alternatives nearby), get_company (every facility under one chain or health system, with penalties and changes of ownership), compare (two to 5 facilities on the same published fields). Every answer is one envelope: answer, source, version, as_of, limitations, zreports_url. A missing value reads "not published", an unmatched name reads "no match".
The zReports grade and zScore for a facility are in its zReport and are never free; the publisher's own figures are. The data is sold as data: never placement, never advertising, and no facility can pay to change a line.
Try it in thirty seconds
In Claude
- Settings, Connectors, Add custom connector
- Name zReports, URL https://api.zreports.org/mcp, no authentication
- Ask: "Which nursing home in Roseville CA has the stronger record"
In ChatGPT or any MCP client
- Add an MCP server with the URL https://api.zreports.org/mcp
- Four tools: find_facility, get_record, get_company, compare
- No key needed inside an assistant
The four tools
| Tool | Arguments | Example |
|---|---|---|
find_facilityFind US health facilities in the zReports facility file by name words, city, state, zip or category (hospital, nursing, asc, psych, rehab, hospice, dialysis, homehealth, fertility, addiction, daycare). Returns matches with the facility_id to pass to get_record or compare. "no match" is an answer: the facility is not in the publisher's file. | name?, city?, state?, zip?, category? | open an example |
get_recordReturn the official records of one facility: the highlight lines, the publisher's row verbatim (official_record), alternatives nearby in the same category, a buy_url for the zReport (which holds the zReports grade and zScore), and the source, version and as_of date. Lines the publisher does not carry read "not published". | facility_id | open an example |
get_companyReturn every facility the publisher lists under one chain or organization name, with counts by category and state, the average score, penalties and fines on record, and changes of ownership. Use for health systems, nursing home chains, dialysis chains and surgery center operators. | name, state? | open an example |
compareCompare two to five facilities side by side on the same published fields: status, ownership, beds, ratings, penalties, deficiencies where the publisher reports them. The zReports grade and zScore are in each facility's zReport. | facility_ids | open an example |
Questions
Is it free?
A free key is one email away on https://zreports.org/api: 500 calls a day, attribution to zReports required. Inside Claude or ChatGPT no key is needed. Paid tiers on the /api page raise the limit and remove the attribution line.
How accurate is it?
Every figure is the government's own, quoted with its source, version and date. Where the file is silent, so are we. Three frontier models sat the US HealthCare Exam. OpenAI 42, Anthropic 44 and Google 34 confident wrong answers out of 50. zBench, answering from official data, 50 of 50. zReports' healthcare facility evaluation is now part of Stanford's MedHELM.
What does an answer look like?
Every answer is one envelope: answer, source, version, as_of, limitations, zreports_url. A missing value reads "not published", an unmatched name reads "no match".
Which assistants can use it?
Any MCP client: Claude (custom connector), ChatGPT (MCP server in developer mode or the app directory once listed), Gemini and any agent framework that speaks MCP or REST.