Skip to main content
Connect Claude Code, Codex, Cursor or another MCP client to your records through /mcp. For offline terminology lookups without personal records, use the stdio server described below. The client controls where retrieved data is sent. A cloud assistant receives the readings it requests even if Mirobody uses local models. To keep the whole interaction local, the client must also use a local model.

Endpoints

All of these sit under HTTP_URI_PREFIX when one is set. On the Docker stack the server listens on port .

Personal MCP URL

A personal URL carries its own credential, so a desktop client needs no sign-in flow. The web client shows it under Settings → MCP link; your own client mints it with the access token it already holds:
One link reads one person’s record. Settings → MCP link has a row for your record and one for each family member whose record you can read. To use a family member’s record from an MCP client, make a link on their row and add it as a second server. Through the API, put that person’s user_id in the body of POST /personal/mcp. Generating a new URL revokes the previous link for the same creator and record. Links expire after MCP_URL_TTL_DAYS days ( by default); using a link does not extend it. Use Settings or the endpoints above to view and revoke links. A link stops working when its sharing permissions or creator authorization end. Sharing the record again does not reactivate a revoked link.
Treat a personal URL like a password: anyone who has it can read the linked record. Keep it out of screenshots and public configuration, revoke it if a device is lost, and use HTTPS for network access.
The URL starts with the origin the request arrived on. Behind a reverse proxy, that origin is only correct when the server trusts the proxy’s forwarded headers; see Deploy on a Server.

Tools an account sees

The endpoint serves these tools: tools/list is filtered per account: a data tool is hidden from an account that holds none of that data, so a client never sees a tool whose only answer would be “no data”. The filtered tools are . What each tool accepts and returns is in Health-Data Tools, and tools you add yourself are served the same way; see Adding Tools. The chat-only tools (the file tools, code_exec and ask_user) are not served over MCP.

Connect a client

Replace the URL below with your personal URL.
Cursor connects to a remote server directly. Add it to ~/.cursor/mcp.json, or to .cursor/mcp.json in a project:
ChatGPT and claude.ai connect from the cloud as well, so they need an HTTPS address they can reach: deploy the stack on a server first. An agent loop of your own calls the endpoint with any MCP client library. The server speaks the current stateless protocol revision and negotiates down for clients that still use the initialize handshake.

Local terminology over stdio

For terminology lookup without a server or personal record, run uvx --python 3.12 mirobody mcp. It serves MCP over stdio from the published package; --python 3.12 lets uv fetch the interpreter the package needs, since an older default picks a release from before 1.2. If Mirobody is already installed, mirobody mcp starts the same server. No database or model key is needed for reading, complaint, indicator or unit standardization. In a client that starts local MCP processes, use this configuration:
This server offers standardize_reading for FHIR Observations with LOINC and UCUM codes, standardize_complaint for ICPC-3, and resolve_indicator, normalize_unit and convert_unit for terminology lookup. Its standardize_report tool reads a PDF, image, spreadsheet or text report; that tool additionally needs the mirobody[parse] extra and a configured model key, and sends the document to the chosen model provider. To enable it with uvx, run uvx --python 3.12 --from 'mirobody[parse]' mirobody mcp. Use the personal URL above when the client needs a person’s stored health record. The protocol implementation and the per-account filter are in .