What it does

“Has anything in my house been recalled?” is a hard question to answer from memory. Recall Radar gives a household assistant a watchlist and seven MCP tools to check public US product, food, device and drug recall records, plus European medicine shortages. A person can add a crib mattress, ask about their household, and see the matching notice with its date, hazard, remedy and official link.

The entry runs a self-hosted MCP server over Streamable HTTP using MCP protocol version 2025-11-25. Its TypeScript server imports and calls the MCP SDK at runtime. A companion web app simulates a smart display: it shows the answer, source cards, transcript and each MCP tool call. The app also has browser speech controls and typed input. The watchlist is stored locally and checked on request; it does not send automatic alerts.

We built this during the hackathon window. CPSC, openFDA and EMA need no API key. A deterministic, visibly labelled scripted mode lets a judge run the real MCP tools and live public data without a model account. With an OpenAI key, a model (gpt-6-luna) drives the same tools through an agent loop; we ran that live, and the demo video shows it. The video's questions are entered from the page address rather than spoken: microphone and speaker were not tested, and its narration is a synthetic voice. This is a simulated Alexa+ experience; no Alexa+ device connection is claimed.

The most useful design choice is provenance: a spoken headline stays short, while each card carries the source URL and date. If an upstream fails, the response reports a warning. Medicine text tells users to consult a pharmacist or doctor before changing treatment. This is an on-demand information aid, not medical or safety advice.

How we built it and what we learned

@modelcontextprotocol/sdk 1.30.1 serves seven tools over stateless Streamable HTTP; the same SDK client drives the app's tool-call loop and bundled command-line client.
CPSC recall search, openFDA food/drug/device enforcement and EMA's shortage JSON are normalized into one source-linked result shape. The watchlist persists to a local JSON file.
The simulated display shows MCP traces and recall cards. Browser speech controls exist, with typed input available when speech is unsupported or permission is unavailable.
Real integration work exposed openFDA's 404 response for zero matches, Node proxy behavior, and the SDK's SSE initialize response. FRICTION.md in the repository records reproducible steps and workarounds.
The first live model call failed: gpt-6-luna refuses function tools on Chat Completions unless reasoning_effort is none. Our fixture tests had passed. The same day the CPSC API answered 503 for most requests; the server reports a missing source in its answer and retries with a date-only listing.
We caught and repaired two demo blockers during exercise: generic stop-using wording on FDA drug and medical-device records, and a responsive layout that covered watchlist removal in an 800×519 viewport. EVIDENCE.md in the repository has the checks and limits.