@cyanheads/cpsc-recalls-mcp-server

Public Hosted Server: https://cpsc-recalls.caseyjhand.com/mcp

Overview

Consumer product recalls from the CPSC saferproducts.gov database — toys, electronics, furniture, appliances, children's products, tools, and clothing. Search recalls by product, brand, retailer, or hazard, fetch full detail for a specific recall number, or pull a recent-recalls feed for a date window. Runs as a stdio process, a local Streamable HTTP server, or the public hosted endpoint above.

Tools

CPSC jurisdiction is consumer products only — food/drugs (FDA), motor vehicles/tires (NHTSA), boats (USCG), and pesticides (EPA) are not in this database; every response carries a jurisdiction note.

Related MCP server: mcp-cpsc

Capability reference

cpsc_search_recalls tool

• Nine text filters — product_name, manufacturer, retailer, importer, distributor, title_search, description_search, remedy (free-text instructions, not the remedy_options categories), and hazard_search (hazard text, product names, or remedy instructions; hazard is accepted as an alias) — all combine with AND; title_search is usually highest-signal

• A search needs at least one criterion — a non-blank text filter or a date bound; limit/offset alone fail with missing_criteria, whose hint points browsing to cpsc_get_recent. A blank or whitespace-only filter counts as omitted

• Word matching: every whitespace-separated word of a text filter must appear in that filter's field, in any order and case-insensitively; straight and curly apostrophes match each other, and %, _, and […] match literally. Words match the plain text the results show, so Stoopher & Boots matches a record CPSC stores as Stoopher & Boots, and tag markup such as valign matches nothing. No stemming or fuzzy matching. Each text filter accepts up to 500 characters

• Two independent date axes: date_start/date_end bound the recall issue date, updated_start/updated_end bound the date CPSC last published it; all four must be real calendar dates and a reversed range throws invalid_date_range

• limit (1–200, default 20) and offset (default 0) page through total_found, which counts every match before the window; a page returns fewer than limit recalls when it reaches the 64,000-byte response size budget, and its notice names the offset to continue from. has_more is the paging signal and truncated always equals it

• Returns hazard descriptions, remedy options and instructions, products, UPCs, manufacturer/importer/retailer/distributor names, images, cpsc_url, and per-recall data_quality_notes; manufacturer and importer render as separate roles — try importer, retailer, or distributor when manufacturer comes back empty

• Zero matches is an empty success, not an error: effectiveQuery echoes the criteria as applied, and a notice says which criterion to relax (the count before hazard_search emptied the set, fewer words, another company role, wider dates). An offset past the last match returns an empty page with its own notice. upstream_rejected (non-retryable) relays CPSC's own rejection, upstream_error (retryable) covers transient outages, including a temporary CPSC data-source failure the server has already retried once

cpsc_get_recall tool

• Accepts modern 5-digit recall numbers (e.g. "25043") and historical 1998–2001 records with letter suffixes (e.g. "99003a"); a malformed number is rejected as invalid input with both forms spelled out

• Returns the complete record — full description, all hazard and remedy detail, every product variant, UPCs, incident/injury narrative, manufacturer/importer/retailer/distributor names, country of manufacture, images, and coordinated-agency recall URLs

• description is nullable — a small number of genuine CPSC records carry no description text, and model numbers are usually embedded there rather than in a structured field

• Manufacturer and importer render under separate headings so role attribution survives into content[]

• data_quality_notes records gaps in the upstream record (absent description, hazard text, or product entries); empty when nothing is missing

• not_found when the recall number doesn't exist — resolve one via cpsc_search_recalls or cpsc_get_recent first; upstream_rejected (non-retryable) relays CPSC's own rejection, upstream_error (retryable) covers transient outages, including a temporary CPSC data-source failure the server has already retried once

cpsc_get_recent tool

• Look-back window of 1–365 days (default 30), anchored to today; older recalls are reachable through cpsc_search_recalls date bounds

• limit (1–100, default 20) and offset (default 0) page through total_found; a page returns fewer than limit recalls when it reaches the 64,000-byte response size budget. Narrowing days cannot page further back — the window is anchored to today, so shrinking it drops the oldest records rather than advancing past the newest

• has_more is the paging signal and truncated always equals it; an empty window, an offset past the last recall, and a budget-cut page each carry a notice saying what to do next

• Returns a lightweight record per recall — number, date, title, hazards, remedy types, product names, cpsc_url — plus data_quality_notes for gaps CPSC left empty

• Use cpsc_get_recall to retrieve full detail for any result

Features

Built on @cyanheads/mcp-ts-core: stdio and Streamable HTTP transports, pluggable auth (none / jwt / oauth), swappable storage (in-memory, filesystem, Supabase, Cloudflare KV/R2/D1), structured logging with optional OpenTelemetry tracing.

CPSC-specific:

• Full client for the CPSC saferproducts.gov public recalls API — search, single-recall detail, and a recent-recalls feed

• Every-word matching on all nine text filters, applied over the complete result set before paging so total_found and has_more stay accurate

• Handles both modern 5-digit recall numbers and historical 1998–2001 records with letter suffixes

• Jurisdiction boundary documented in every response — flags food, vehicle, and drug recalls as out of scope before an agent misattributes them

Agent-friendly output:

• Provenance on every response — source_note, cpsc_url, and (CPSC source text) blockquote labels distinguish relayed CPSC narrative from the server's own guidance

• Plain text on both surfaces — the HTML markup and character codes CPSC stores in some records (&, <br>, stray table tags) are converted to plain text before matching and output, and content[] escapes Markdown so model and serial numbers such as 1HFVE05**K4000003 keep every character; recall page, image, and coordinated-recall addresses that contain spaces are percent-encoded in content[] so each renders as one working link, while structuredContent carries them as CPSC stores them

• Pagination discriminators — total_found, offset, has_more, and truncated on every search/recent response so agents can tell when results are clipped and page with offset; each surface of a page stays within a 64,000-byte budget, and a notice gives the next offset when the budget ends a page early

• data_quality_notes on every response — gaps observed in the upstream record (missing hazard text, no product entries), derived from which fields CPSC left blank rather than any judgment call

• Jurisdiction note (cpsc_jurisdiction) on every response — lets agents route callers to the correct agency (FDA, NHTSA, USCG, EPA) when a product is out of scope

Getting started

Public Hosted Instance

A public instance is available at https://cpsc-recalls.caseyjhand.com/mcp — no installation required. Point any MCP client at it via Streamable HTTP:

{
  "mcpServers": {
    "cpsc-recalls-mcp-server": {
      "type": "streamable-http",
      "url": "https://cpsc-recalls.caseyjhand.com/mcp"
    }
  }
}

Self-Hosted / Local

Add the following to your MCP client configuration file.

{
  "mcpServers": {
    "cpsc-recalls-mcp-server": {
      "type": "stdio",
      "command": "bunx",
      "args": ["@cyanheads/cpsc-recalls-mcp-server@latest"],
      "env": {
        "MCP_TRANSPORT_TYPE": "stdio",
        "MCP_LOG_LEVEL": "info"
      }
    }
  }
}

Or with npx (no Bun required):

{
  "mcpServers": {
    "cpsc-recalls-mcp-server": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@cyanheads/cpsc-recalls-mcp-server@latest"],
      "env": {
        "MCP_TRANSPORT_TYPE": "stdio",
        "MCP_LOG_LEVEL": "info"
      }
    }
  }
}

Or with Docker:

{
  "mcpServers": {
    "cpsc-recalls-mcp-server": {
      "type": "stdio",
      "command": "docker",
      "args": ["run", "-i", "--rm", "-e", "MCP_TRANSPORT_TYPE=stdio", "ghcr.io/cyanheads/cpsc-recalls-mcp-server:latest"]
    }
  }
}

For Streamable HTTP, set the transport and start the server:

MCP_TRANSPORT_TYPE=http MCP_HTTP_PORT=3010 bun run start:http
# Server listens at http://localhost:3010/mcp

Prerequisites

• Bun v1.4.0 or higher (or Node.js v24+).

• No API key required — the CPSC public recalls API is freely accessible.

Installation

1. Clone the repository:

git clone https://github.com/cyanheads/cpsc-recalls-mcp-server.git

2. Navigate into the directory:

cd cpsc-recalls-mcp-server

3. Install dependencies:

bun install

4. Configure environment:

cp .env.example .env
# edit .env if needed — no required API keys

Configuration

All configuration is validated at startup via Zod schemas. Key environment variables:

No server-specific API keys are required. See .env.example for the full list of optional overrides.

Running the server

Local development

• Build and run the production version:

  # One-time build
  bun run rebuild
  
  # Run the built server
  bun run start:stdio
  # or
  bun run start:http

• Run checks and tests:

  bun run devcheck  # Lint, format, typecheck, security
  bun run test      # Vitest test suite

Docker

docker build -t cpsc-recalls-mcp-server .
docker run --rm -p 3010:3010 cpsc-recalls-mcp-server

The Dockerfile defaults to HTTP transport, stateless session mode, and logs to /var/log/cpsc-recalls-mcp-server. OpenTelemetry peer dependencies are installed by default — build with --build-arg OTEL_ENABLED=false to omit them.

Project structure

Development guide

See CLAUDE.md / AGENTS.md for development guidelines and architectural rules. The short version:

• Handlers throw, framework catches — no try/catch in tool logic

• Use ctx.log for request-scoped logging, ctx.state for tenant-scoped storage

• Register new tools in the createApp() arrays in src/index.ts

• Wrap external API calls: validate raw → normalize to domain type → return output schema; never fabricate missing fields

Contributing

Issues are welcome. Run checks and tests before submitting:

bun run devcheck
bun run test

License

Apache-2.0 — see LICENSE for details.