---
name: fda-device-intel
description: Query FDA medical-device databases directly with curl — no server or API key needed. Use for 510(k) clearances, predicate device research, product code classification, MAUDE adverse events, recalls (incl. Class I/II/III), warning letters, PMA approvals, establishment registrations, and UDI/GUDID brand-to-company lookup. Triggers - "find a predicate", "510k", "product code", "MAUDE", "FDA recall", "warning letter", "adverse events for <device>", "who makes <device type>", "postmarket health check", "device classification".
---

# FDA device intelligence via public APIs

All sources are public and keyless. Anonymous openFDA limits: ~240 req/min,
~1,000 req/day per IP — fine for interactive use; add small sleeps in loops.
Always cite record URLs in answers. Data is unvalidated and informational
only, not regulatory advice (https://open.fda.gov/terms/).

## openFDA query syntax (critical quirks)

Base: `https://api.fda.gov/device/<endpoint>.json`

- Join terms with **literal `+AND+`**; spaces inside a value become `+`.
  NEVER let a URL encoder turn `+` into `%2B` — openFDA needs it literal.
- Phrases: `field:%22two+words%22`. Date ranges: `field:%5B<lo>+TO+<hi>%5D`.
- `search=` filters; add `count=<field>` to aggregate instead of listing.
  Count-field names are inconsistent: text fields need `.exact`
  (`applicant.exact`, `registration.name.exact`), keyword fields must NOT
  have it (`registration.iso_country_code`). **Empty counts result = wrong
  field name, not zero data** — retry with/without `.exact`.
- `sort=<datefield>:desc`, `limit=` (max 100, aggregate max 1000), `skip=`.
- A 404 with `NOT_FOUND` means zero matches, not a bad URL.

## Endpoints and key fields

| Endpoint | Purpose | Key fields | Date field (format) |
|---|---|---|---|
| `510k` | Clearances | `k_number`, `product_code`, `applicant`, `device_name`, `decision_description`, `clearance_type` | `decision_date` (YYYY-MM-DD) |
| `classification` | Description/code → class | `product_code`, `device_name`, `device_class`, `regulation_number`, `medical_specialty_description` | — |
| `event` | MAUDE adverse events | `device.device_report_product_code`, `device.brand_name`, `device.manufacturer_d_name`, `product_problems`, `event_type`, `mdr_text[].text` | `date_received` (YYYYMMDD) |
| `recall` | Recalls w/ root cause | `product_code`, `recalling_firm`, `product_description`, `reason_for_recall`, `root_cause_description` | `event_date_initiated` (YYYY-MM-DD) |
| `enforcement` | Recalls w/ **severity** | `classification` ("Class I/II/III"), `recalling_firm`, `product_description`, `status` | `report_date` (YYYYMMDD) |
| `pma` | Class III approvals | `pma_number`, `supplement_number`, `trade_name`, `applicant`, `decision_code`, `product_code` | `decision_date` (YYYY-MM-DD) |
| `registrationlisting` | Who makes what | `products.product_code`, `registration.name`, `registration.iso_country_code`, `establishment_type`, `proprietary_name` | — |
| `udi` | GUDID; brand→company | `brand_name`, `company_name`, `version_or_model_number`, `device_description`, `product_codes.code` | — |

Examples:

```sh
# description -> product code (then class/regulation)
curl -s 'https://api.fda.gov/device/classification.json?search=device_name:bipolar+AND+device_name:laparoscopic&limit=5'
# newest clearances under a code
curl -s 'https://api.fda.gov/device/510k.json?search=product_code:GEI&sort=decision_date:desc&limit=10'
# who files most under a code since 2020 (clearance velocity)
curl -s 'https://api.fda.gov/device/510k.json?search=product_code:GEI+AND+decision_date:%5B2020-01-01+TO+2030-01-01%5D&count=applicant.exact&limit=10'
# MAUDE top problem codes for a product code since 2025 (note compact dates)
curl -s 'https://api.fda.gov/device/event.json?search=device.device_report_product_code:GEI+AND+date_received:%5B20250101+TO+20301231%5D&count=product_problems.exact&limit=10'
# serious recalls only
curl -s 'https://api.fda.gov/device/enforcement.json?search=recalling_firm:olympus+AND+classification:%22Class+I%22&sort=report_date:desc&limit=5'
# resolve a brand to its legal company (entity resolution)
curl -s 'https://api.fda.gov/device/udi.json?search=brand_name:ligasure&count=company_name.exact&limit=5'
```

## Warning letters (NOT in openFDA)

Undocumented fda.gov Drupal/Solr endpoint behind their search page; needs a
browser User-Agent. Returns JSON `{recordsFiltered, data: [[html cells]]}` —
cells are: posted date, issued date, company (with `href` to letter page),
issuing office, subject. Strip HTML tags from cells.

```sh
curl -s -A 'Mozilla/5.0 (X11; Linux x86_64)' \
  'https://www.fda.gov/datatables/views/ajax?view_name=warning_letter_solr_index&view_display_id=warning_letter_solr_block&view_path=/node/360089&search_api_fulltext=COMPANY&search_api_fulltext_issuing_office=Center+for+Devices+and+Radiological+Health&length=20&start=0&draw=1'
```

Omit `search_api_fulltext_issuing_office` for all centers. Matching is
full-text → false positives are common; verify the company cell of each hit.
Low volume only; endpoint may change without notice.

## 510(k) summary PDFs (indications for use)

The indications statement lives only in the summary PDF, not the API.
URL pattern from the K-number's first two digits (the year):

```
K213042 -> https://www.accessdata.fda.gov/cdrh_docs/pdf21/K213042.pdf
K025432 -> .../cdrh_docs/pdf2/K025432.pdf     (2002-2009: pdf2..pdf9)
K99xxxx -> .../cdrh_docs/pdf/K99xxxx.pdf      (pre-2002: often scanned, no text)
```

Download with `-A 'Mozilla/5.0'`, extract with `pdftotext -layout file.pdf -`
(or read the PDF directly if your Read tool supports PDFs). Then find the
"Indications for Use" / "Intended Use" section — **skip windows containing
"substantially equivalent (for the indications" or "premarket approval
application (PMA)"**: that's FDA decision-letter boilerplate preceding the
real summary. The true IFU is often the 2nd or 3rd occurrence. The PDF also
contains the predicate device table (predicate K-numbers) absent from the API.

## Recipe: postmarket health check (predicate/competitor due diligence)

For "is this device/company safe to cite as a predicate":
1. Recalls: `recall` by `product_description:<brand>`; get total + latest.
2. Severity: `enforcement` same query — any Class I?
3. MAUDE volume last ~2y: `event` by `device.brand_name:<brand>` (count only).
4. Warning letters: WL endpoint with company name, CDRH-filtered.
5. First resolve brand→company via `udi` to avoid name collisions.

Interpretation caveats to state in answers: keyword matches collide (a
"Voyant" query matches an unrelated contact-lens brand — verify every hit);
high MAUDE volume may just mean a big install base (interpret per-volume);
MAUDE data lags months behind real time; `recall` total ≠ `enforcement`
total (different views of the same actions).

## No shell? URL-fetch-only agents (browsing / web-fetch tools)

If you have no shell but can fetch URLs (ChatGPT browsing, Claude web fetch,
Copilot HTTP actions), most of this skill still works — every openFDA query is
a plain HTTPS GET with no headers or auth required.

- **Encoding (verified):** `%20` is equivalent to the literal `+` separator —
  `search=product_code:GEI%20AND%20decision_date:%5B2020-01-01%20TO%202030-01-01%5D`
  works identically. So it is safe to compose queries with spaces and let your
  fetch tool percent-encode them. The ONLY breaker is `%2B` (encoded plus) —
  if you get `SERVER_ERROR ... parse_exception`, your tool encoded `+` as
  `%2B`; rewrite the query using spaces instead of plus signs.
- **510(k) summary PDFs:** fetch the `accessdata.fda.gov/cdrh_docs/...` URL
  directly — many clients parse PDFs natively. If yours returns garbage or
  refuses binary content, this leg needs a shell or the fda-mcp server's
  `get_510k_summary` tool.
- **Warning letters:** the fda.gov endpoint is User-Agent-sensitive (curl's
  default UA passes; `python-requests` gets a 302 redirect). Try it; if you
  get a redirect, HTML challenge page, or non-JSON, fall back to fetching the
  human search page
  `https://www.fda.gov/inspections-compliance-enforcement-and-criminal-investigations/compliance-actions-and-activities/warning-letters`
  and reading its rendered table, or report that this leg needs the fda-mcp
  server.
- Everything else (all eight openFDA endpoints, count aggregations, sorting,
  pagination) is fully usable as plain URLs. Same rate limits and caveats
  apply.