fda-mcp

Local MCP server exposing FDA medical-device intelligence tools. All data sources are public and keyless: openFDA (api.fda.gov), the fda.gov warning-letter endpoint (not in openFDA), and 510(k) summary PDFs from accessdata.fda.gov (parsed via pdftotext).

The server has no LLM inside — it is the toolbox. Your AI client (opencode, Claude Code, Claude Desktop, Cursor, VS Code Copilot) supplies the model and decides when to call the tools.

Requirements

Bun (tested with 1.4.x)
pdftotext (poppler-utils) on PATH — only needed for get_510k_summary

Run

bun src/index.ts    # speaks MCP over stdio; clients spawn this themselves

Tools

Tool	What it does
`lookup_classification`	Plain-English description or product code → product codes, device class, regulation
`search_510k`	510(k) clearances by code/applicant/device name/date range
`get_510k_summary`	One clearance + its summary PDF parsed, indications statement extracted
`search_maude`	MAUDE adverse events; list mode or `count_by` aggregation (e.g. top problem codes)
`search_recalls`	Device recalls by code/firm/product keyword
`search_warning_letters`	fda.gov warning letters, optionally CDRH-only (dataset absent from openFDA)
`check_postmarket_health`	Recalls + MAUDE volume + warning letters for a brand/company — predicate due diligence
`search_pma`	PMA approvals/supplements (Class III devices) by code/applicant/date
`search_registrations`	Establishment Registration & Listing — who makes/sterilizes what under a code (competitor/supplier discovery)
`lookup_udi`	GUDID/UDI records — resolve brand → legal company name, enumerate models/versions
`search_enforcement`	Recall enforcement reports with Class I/II/III severity (complements `search_recalls`)
`get_query_guide`	Full query guide (syntax quirks, endpoint cheat sheets, recipes) — read before `fda_fetch`
`fda_fetch`	Guarded raw GET against FDA hosts only — escape hatch for query shapes the typed tools miss

Most search tools accept count_by for aggregations (e.g. search_510k with count_by: "applicant.exact" = who files most under a product code).

Skill (for shell-capable agents — no server needed)

skills/fda-device-intel/SKILL.md teaches an agent with shell access (Claude Code, opencode, Codex CLI, Copilot agent mode) to query the same endpoints directly with curl — no process, no install beyond the markdown file:

One-liner install (Claude Code / any agent reading ~/.claude/skills/):

mkdir -p ~/.claude/skills/fda-device-intel && curl -fsSL -o ~/.claude/skills/fda-device-intel/SKILL.md https://raw.githubusercontent.com/happylemonprogramming/fda-mcp/main/skills/fda-device-intel/SKILL.md

Or paste the skill straight into any agent that can fetch URLs:

Fetch https://raw.githubusercontent.com/happylemonprogramming/fda-mcp/main/skills/fda-device-intel/SKILL.md and follow it to answer FDA medical-device questions.

Skill and MCP server share one knowledge source: get_query_guide serves the skill's body to MCP clients, so remote no-shell clients (Claude Desktop, ChatGPT, Teams/Copilot Studio) get the same guidance driving fda_fetch.

Connect a client

opencode

Add to opencode.json (project) or ~/.config/opencode/opencode.json:

{
  "mcp": {
    "fda": {
      "type": "local",
      "command": ["bun", "/home/lemon/Projects/fda-mcp/src/index.ts"],
      "enabled": true
    }
  }
}

Claude Code

claude mcp add fda -- bun /home/lemon/Projects/fda-mcp/src/index.ts

or in .mcp.json:

{
  "mcpServers": {
    "fda": {
      "command": "bun",
      "args": ["/home/lemon/Projects/fda-mcp/src/index.ts"]
    }
  }
}

Claude Desktop

~/.config/Claude/claude_desktop_config.json → same mcpServers block as Claude Code above.

Example prompts to try

"Look up the product code for a bipolar laparoscopic coagulation device, then show me the 5 newest clearances under it."
"Run a postmarket health check on POWERSEAL by Olympus."
"Get the indications statement from K213042's summary PDF."
"Any CDRH warning letters mentioning Medtronic?"

Caveats

MAUDE data lags months behind real time.
The warning-letter endpoint is an undocumented fda.gov internal API; it can change without notice. Keep request volume low.
Name matching is keyword-based; brand/company collisions happen (a "Voyant" query matches an unrelated contact lens). Every tool output says so.
Informational only, not regulatory advice; see https://open.fda.gov/terms/.