Recall Check

Tell Alexa what you own. Hear about it when any of it is recalled, and what to do.

Recall Check is an Alexa+ add-on: a self-hosted MCP server (spec 2025-11-25, Streamable HTTP) that watches a household's belongings for United States safety recalls. It covers consumer products (CPSC), vehicles (NHTSA), and food and medicine (FDA).

Built for the Alexa+ track of Build, Ship, Shape: Amazon Developer Hackathon. MIT licensed.

The problem

The United States issues hundreds of safety recalls a year. Most owners never hear about the ones that affect them: a recall notice reaches the people who registered the product, and few people register an air fryer or a stroller. To find out, you would have to remember everything you own and search three government websites, repeatedly.

A voice assistant is already in the room with those products. Recall Check lets you say what you own once, in passing, and then does the checking for you.

Related MCP server: Custodian

A conversation

You: Alexa, add my Cosori air fryer, model CP158-AF. Alexa: Added your Cosori air fryer, model CP158-AF. Your Cosori air fryer is recalled. Stop using it now. A wire connection in the air fryers can overheat, posing fire and burn hazards. Ask me what to do about it.

You: What should I do about it? Alexa: Stop using your Cosori air fryer now. Consumers should immediately stop using the recalled air fryers and contact Cosori to receive their choice of a free replacement air fryer or another Cosori product. The remedy on offer is a replacement. The contact details and the official notice are in the Alexa app.

You: I got the replacement. Alexa: Done. I will not mention that recall on your Cosori air fryer again.

That exchange is real output from this repository against live recall data.

What it does

How sure it is

A false alarm costs trust and a missed recall costs safety, so Recall Check never reports a match on the brand alone. The recall has to name the maker and the kind of product. The model number then sets the confidence, and the wording follows it:

Vehicle recalls apply to "certain" vehicles of a make, model and year, so they are always reported as likely, with advice to confirm by VIN.

Meeting the Alexa+ requirements

Account linking (OAuth 2.1 with PKCE) is the next step for a real listing. See docs/ALEXA-READINESS.md.

The simulated Alexa+ experience

The Alexa+ developer tools are in preview with select partners and are not available to hackathon participants, so this repository includes its own host: a web page styled as an Echo Show, served at /.

It is a simulation of the assistant only. The add-on behind it is real. For every utterance the host opens an MCP client connection to /mcp over HTTP, exactly as Alexa+ would, calls the tools, and shows each call with its latency against the 500 ms limit.

The host understands utterances with built-in rules, so it runs with no API key. Set LLM_API_KEY and LLM_MODEL to have a language model choose the tools instead, through any OpenAI-compatible endpoint.

Run it

Requires Node.js 20 or newer.

npm install
npm run sync     # downloads recall data (about 7,000 recalls, 10 to 30 seconds)
npm start        # http://localhost:8787

Open http://localhost:8787 for the simulator. The MCP endpoint is POST http://localhost:8787/mcp.

To call the endpoint directly, set a token first:

ACCESS_TOKENS=secret:me npm start
npx @modelcontextprotocol/inspector   # connect to http://localhost:8787/mcp with header Authorization: Bearer secret

Configuration is in .env.example.

Tests

npm test

62 tests, no network and no keys needed. They cover matching (including that a brand alone, or a brand mentioned only in a description, never produces a match), the three agency adapters against real record shapes, the household service, what is spoken, the tools called through a real MCP client, and the simulator's understanding of utterances.

How it is built

src/
  sources/     one adapter per agency: CPSC, FDA, NHTSA
  text.ts      tokenising and model-number normalising, shared by index and matcher
  match.ts     the recall index and the matching rules
  service.ts   what Recall Check can do, independent of MCP
  speech.ts    every sentence Alexa says
  mcp.ts       the seven tools
  server.ts    HTTP: /mcp, /health, and the simulator
  auth.ts      who is calling
  store.ts     one file per household
  host/        the simulated assistant: utterance rules, optional model, MCP client
public/        the simulated device
addon-package/ the add-on manifest, in the format Amazon's CLI produces

Recall data is refreshed every six hours. If one agency is down, its rows are carried over from the previous snapshot, so an outage never empties the index.

Limits

• United States recalls only.

• Meat and poultry recalls (USDA FSIS) are not covered. That agency's API refuses automated requests.

• FDA coverage is ongoing recalls from the last ten years. Medical devices are not yet included.

• Matching is by words and model numbers. It does not read serial numbers, date codes or lot numbers, so "likely" and "possible" results always need a look at the label.

• Recall Check is not affiliated with Amazon or with any government agency. The official notice is the authority.

Documents

• docs/ALEXA-READINESS.md: what is done and what remains for a real Alexa+ listing.

• docs/FRICTION-LOG.md: what was hard while building against Amazon's developer documentation.

• docs/PRODUCT-FEEDBACK.md: answers to the hackathon's product feedback questions.