trundler

A local MCP server for grocery shopping. Runs entirely on your own machine and residential connection — no cloud services, no datacenter IPs, no bot-detection logistics.

Supported providers:

• Countdown / Woolworths NZ — full account access (login + cart + order history)

• New World (Foodstuffs) — anonymous price/product browsing

• Pak'nSave (Foodstuffs) — anonymous price/product browsing

Providers at a glance

Different chains expose different things, so trundler's capabilities vary by provider.

In short:

• Countdown / Woolworths — you log in once in a real browser; after that you can search, view specials, and add items to a real cart and read your order history.

• New World & Pak'nSave — no login needed to search and compare prices, but they are read-only: you cannot add to a cart or see order history (yet). Because Foodstuffs pricing is per-store, you must choose a store first with set_store.

Why the difference? Countdown authenticates a real user session, which unlocks the cart. The Foodstuffs (New World / Pak'nSave) read APIs serve anonymous guests, so browsing needs no login — but the cart requires an authenticated session that isn't wired up yet (see Roadmap).

Related MCP server: Swissgroceries MCP

How it works

Everything lives in one local process over stdio — the agent launches trundler as a subprocess. No ports, no CORS, no session server. There are two auth models depending on the provider:

Countdown / Woolworths — browser-assisted login

• login opens a real browser window; you sign in yourself (handling any MFA/captcha), and trundler captures the session. No password is stored. Because the login happens in a real browser on your home connection, it's the most bot-resistant approach.

• Everything after that is authenticated fetch() using the captured cookies + XSRF token — no browser needed.

• Silent refresh renews short-lived tokens by briefly relaunching a headless browser with the saved session. If it has fully expired, tools tell the agent to run login again.

New World / Pak'nSave — anonymous guest token

• No login. trundler mints an anonymous guest token by loading the store homepage and reading the session cookie it hands out, then calls the read API with plain fetch(). The token is cached and refreshed automatically (~30-minute life).

• The homepage sits behind Cloudflare bot-management that rejects Node's fetch, so the token mint shells out to curl (which passes). curl ships with Windows 10 1803+, macOS, and Linux — no extra install needed on a normal machine.

• Pricing is per-store: use list_stores to find one and set_store to select it. Your choice is persisted per provider.

Session/config data is stored per provider outside the repo:

• Windows: %LOCALAPPDATA%\trundler\<provider>\

• macOS/Linux: ~/.config/trundler/<provider>/

Install

npm install          # also downloads the Chromium browser (via postinstall)
npm run build

If the browser didn't download automatically, run npx playwright install chromium. (Chromium is used only for the Countdown login; New World / Pak'nSave don't need it.)

Register with your agent

Add to your MCP config (e.g. .mcp.json):

{
  "mcpServers": {
    "trundler": {
      "command": "node",
      "args": ["D:/Projects/MCP/trundler/dist/index.js"]
    }
  }
}

During development you can point it at the TypeScript source instead:

{
  "mcpServers": {
    "trundler": {
      "command": "npx",
      "args": ["tsx", "D:/Projects/MCP/trundler/src/index.ts"]
    }
  }
}

After changing server-level code (including the presentation instructions), rebuild (npm run build) and reconnect the MCP — instructions and tool lists are sent once at connection time.

Setup per provider

Countdown / Woolworths — log in once:

npm run cli login          # or: node dist/cli.js login   (after build)

A browser window opens — sign in to Woolworths; it closes once the session is captured. Verify any time with npm run cli check.

New World / Pak'nSave — no login; just pick a store. Via your agent:

list_stores  { "provider": "newworld", "query": "auckland" }
set_store    { "provider": "newworld", "storeId": "<id from list_stores>" }

Do the same with "provider": "paknsave" for Pak'nSave. Until a store is set, the read tools return a "no store selected" error.

For local testing you can pre-seed a store via env var instead of set_store: TRUNDLER_NEWWORLD_STORE_ID / TRUNDLER_PAKNSAVE_STORE_ID. These are test overrides only — never a shipped default.

Tools

Every tool takes an optional provider argument (default: countdown).

Calling a login/cart/order tool on New World or Pak'nSave returns a clear "requires login — not yet supported" error rather than failing silently.

Product listings

trundler ships server-level instructions telling the agent how to present product lists: label each item with a capital letter (A, B, C…) for easy quantity-picking, show price per unit and sort cheapest-per-unit first, and include a link to open each product's photo/detail page. This makes cross-product and cross-chain price comparison straightforward.

Adding a provider

Implement ShoppingProvider (see src/core/provider.ts) in a new src/providers/<name>/ folder and register it in src/providers/index.ts. The MCP tools are provider-agnostic and dispatch automatically. Foodstuffs banners (New World, Pak'nSave) share one FoodstuffsProvider parameterised by a banner config — adding another Foodstuffs banner is a few lines in src/providers/foodstuffs/banners.ts.

Capabilities a provider doesn't support (e.g. cart on a read-only provider) simply throw an error, which surfaces to the agent as a tool error.

Roadmap

• Foodstuffs cart & orders (Tier 2). New World / Pak'nSave cart and order history need an authenticated session. Their login exchange is Cloudflare-protected against automated browsers, so it must be captured from a genuine browser (connect to your real Chrome), not Playwright's bundled Chromium. Not yet implemented.

Trade-off

trundler refreshes on demand, not on a schedule — for Countdown it can't renew tokens while your machine is off, and an expired session needs a quick manual login. That's the cost of staying local and on a residential IP. If you later need unattended scheduled runs, host it on an always-on machine at home (not a datacenter, or you reintroduce the bot-detection problem).