openFDA MCP

An MCP server exposing FDA regulatory metadata — drugs, biologics, and medical devices — through the openFDA API.

Built by Black Swan Causal Labs as the identifier-resolution layer for a real-world-evidence (RWE) case roster: given an FDA application number or a product name, resolve it to authoritative regulatory metadata.

Why this exists

There are other openFDA MCP servers, and several are broader. This one is narrow on purpose: it is the instrument that resolved the application numbers in a specific published RWE dataset, and it exists so that dataset can name the tool that produced it.

That matters more than it might sound. Whether BLA 125123/2058 resolves to a particular product, or DEN160026 to a particular device class, is a decision made by a piece of software — and a different wrapper can yield a different roster. "We used openFDA" is not a sufficient methods statement; "openfda-mcp v0.1.0" is. If you use this in research, pin the version.

Its practical edge over a general openFDA client is the device half: resolving a CDRH submission number to a risk class takes a three-hop chain (number → product code → classification) with two non-obvious traps, both handled here.

Tools

Drugs and biologics (/drug/* — CDER, CBER)

Devices (/device/* — CDRH)

Install

pip install openfda-mcp

Add to your MCP client config:

{
  "mcpServers": {
    "openfda": {
      "command": "openfda-mcp",
      "env": { "OPENFDA_API_KEY": "${OPENFDA_API_KEY:-}" }
    }
  }
}

The API key is optional. Without one, openFDA allows 40 requests/min and 1,000/day, which is enough for interactive use. A free key raises it to 240/min and 120,000/day — worth having for bulk sweeps.

Two findings worth knowing

Both were established empirically and are not obvious from FDA's docs.

De Novo grants live in the 510(k) endpoint. DEN###### numbers are stored in the k_number field of /device/510k. There is no De Novo endpoint, and looking for one leads to the wrong conclusion that De Novo numbers can't be resolved. They can.

HDE numbers are not in openFDA at all. Neither the 510(k) nor the PMA endpoint carries H######, so no product code — and therefore no classification — is retrievable. This server still reports device_class: "III" for them, by regulatory inference: HDE is by definition the pathway for devices that would otherwise require a PMA. medical_specialty stays null, because that one really is unavailable, and device_class_source says which is which.

Transient failures are never silent

A genuine absence and a failed request are different things, and this package keeps them different:

• not found (HTTP 404, or 200 with no results) → returns None; safe to cache
• transient failure (timeout, connection error, 429, 5xx) → retried with backoff, then raises OpenFDATransientError; never cache this
• rejected request (other 4xx) → raises OpenFDARequestError

This is a direct response to a real defect: an earlier version swallowed every exception and returned None, so a single read timeout on one application number was cached as a real miss and silently blanked two fields on that record for weeks. Cached failures are indistinguishable from real absences, which makes them the worst kind of silent data loss.

On screen_for_rwe_signals

It is unvalidated. There is no ground-truth oracle for a discovery sweep, and below the strongest hits the results are dominated by applications whose labels use "registry" in an unrelated sense. Treat its output as candidates for human review — not as a finding, and not as a count to report. Establishing recall against a held-out set of known cases is open work.

Development

pip install -e ".[dev]"
pytest              # unit tests, offline
pytest -m live      # live checks against api.fda.gov

Live tests assert against known-good fixtures (K203571 → class II Ophthalmic, DEN160026 → class II Immunology, BLA761180 → LEO Pharma) so a change on FDA's side surfaces as a test failure rather than as quietly wrong data.

License

MIT