Dowser: find the money hiding in your life

# Dowser

**Finds the money hiding in your life.** A read-only MCP connector for AI assistants such as Meta Muse: open class action settlements, CPSC product recalls with refunds, unused credit card credits, expiring loyalty points and birthday freebies, each with an official link.

Built for the [Muse Connector Platform](https://muse.ai/platform). Works with any MCP client, and every tool doubles as a REST endpoint.

## Tools

| Tool | What it answers | REST |
|---|---|---|
| `search_settlements` | Which open U.S. class action settlements match the brands I use? Deadline, payout, proof needed, claim link. | `GET /api/v1/settlements` |
| `match_recalls` | Has anything I own been recalled in the last three years, and what is the remedy? | `GET /api/v1/recalls` |
| `card_benefits` | Which statement credits and perks on my cards reset soon and how do I use them? | `GET /api/v1/cards` |
| `points_expiry` | When do my miles and points expire, and what keeps them alive? | `GET /api/v1/points` |
| `birthday_freebies` | What do chains near me give away on birthdays, and by when must I join? | `GET /api/v1/freebies` |
| `check_for_new` | What is new since my last check? One call for a weekly sweep. | `GET /api/v1/updates` |

Plus one MCP prompt, `find_my_money`, that walks an assistant through a full sweep.

## How it works

```
data/settlements.json  ← scripts/build-data.ts  ← OpenClassActions.com + ClassAction.org (daily, GitHub Action)
data/recalls.json      ← scripts/build-data.ts  ← CPSC recalls API, 3-year window (daily) + live top-up per request
data/cards.json, programs.json, freebies.json     hand-maintained catalogues with last_reviewed + confidence per entry

src/app.ts     Express: POST /mcp (stateless streamable HTTP, official SDK), /api/v1/*, pages, /openapi.json
src/mcp.ts     registers the six tools (all readOnlyHint) and the prompt
src/tools/*    pure functions; the MCP and REST layers are thin wrappers
api/index.js   Vercel entry (imports dist/app.js)
```

No database, no accounts, no stored personal data. Matching is a small TF-IDF index over titles, product names and summaries, so "Bank of America" scores higher than "America" and generic words do not match everything.

## Run it locally

```bash
npm install
npm run build:data      # ~30 s: scrapes settlements, pulls 3 years of CPSC recalls
npm run dev             # http://localhost:3000  (MCP at