Openfda MCP Server

What openfda-mcp-server does

The openfda-mcp-server MCP server gives an MCP client access to FDA data through the openFDA public API. Its coverage includes human drugs, veterinary products, food, medical devices, tobacco products, adverse-event reports, enforcement records, recalls, drug shortages, labeling, approvals, clearances, and the National Drug Code Directory.

A consolidated drug profile accepts one brand or generic name and combines identity information with label, adverse-event, recall, approval, and shortage results. Separate search tools support more targeted work, such as finding device 510(k) clearances or PMA approvals, searching tobacco product problem reports, or locating veterinary adverse events. Results can also be aggregated by field, while field-description queries show searchable paths and verified expressions for counting.

This places openfda-mcp-server in the Search & Data Extraction category. It is suited to research assistants, medication-information workflows, FDA data exploration, and agents that need to chain identifiers such as NDC, RxCUI, or SPL set IDs into later lookups.

How it works

The server can run as a local stdio process or through Streamable HTTP. A public hosted endpoint is also listed at https://openfda.caseyjhand.com/mcp. Local installation examples use the npm package @cyanheads/openfda-mcp-server through npx.

Most search tools query live openFDA endpoints and return paginated records, with several tools allowing up to 1,000 records per page. Search inputs vary by endpoint. For example, adverse-event searches select drug, food, or device schemas; shortage searches can filter by status, therapeutic category, generic name, or company; and recall searches distinguish enforcement records from device recall records.

The field-description tool is useful before constructing a query. It groups field paths by category, provides types and descriptions, identifies whether a field has a valid aggregation expression, and includes query guidance for quoting, Boolean operators, exact matches, and date ranges. Count operations return ranked terms and indicate when results may be truncated.

Drug-label requests can return an outline when the response would exceed the inline text budget. A follow-up request can select specific sections. Structured output preserves raw SPL markup, while table sections are rendered as Markdown in text output.

Setup and configuration

Configure an MCP client with the package command and provide an OPENFDA_API_KEY value, as shown by the repository's installation links. The repository also provides a Docker package. The server can alternatively be reached through its hosted Streamable HTTP endpoint.

The DataCanvas tools are opt-in. They can list tables and schemas, run read-only SQL against a staged result set, and delete a staged table. The deletion tool is off by default. These tools are separate from the main live openFDA searches and should be enabled only when that staged-data workflow is needed.

Tools and capabilities

The openfda-mcp-server MCP server includes tools for:

• Building a multi-section drug profile.
• Searching human and veterinary adverse events.
• Finding drug shortages, tobacco reports, recalls, approvals, and device clearances.
• Looking up drug labels and NDC records.
• Describing endpoint fields and counting unique field values.
• Inspecting, querying, and optionally deleting DataCanvas tables.

Some tools expose endpoint-specific behavior. Veterinary records mask brand and manufacturer names as MSK, so active-ingredient searches are more appropriate. Device recall results use different status fields from enforcement records. A profile can return partial sections and identify upstream failures, while authentication, configuration, and cancellation failures stop the full profile request.

Limitations and notes

The service reflects the data and query behavior of openFDA. A missing profile section may mean no matching result or an upstream sub-query failure, so consumers should inspect the degradation metadata rather than treating every empty section as confirmed absence. Count completeness is not guaranteed when distinct terms exceed the requested limit, and openFDA may reject expressions that cannot be aggregated.

Endpoint schemas and sortable date fields differ by category. For example, adverse-event date fields vary across drug, food, and device records. Device recall searches support the recall endpoint, while non-device categories must use enforcement records. The DataCanvas query and deletion tools are optional and are not a replacement for the live API.