MCP App Store
by KeenEthics
ShopSaver icon

ShopSaver MCP Server for Claude

By ShopSaver

Overview

ShopSaver gives Claude live grocery prices from Australia's three big supermarkets: Coles, Woolworths and ALDI. There's no account and no setup. Ask in plain language. What you can ask: • "Where's Vegemite cheapest this week?" Finds a product across all three chains, ranked cheapest first by comparable unit price, so a 750 ml bottle and a 1 L bottle are compared fairly. • "Has this coffee gone up lately?" Shows how a product's price has moved over time. • "Would my weekly shop be cheaper at ALDI?" Prices a whole shopping list against another chain, line by line, and reports how much of the list it could actually compare. • "Is it worth waiting for this to go on special?" Gives a week-by-week probability that a product goes on special, based on its price history over Australian catalogue weeks (Wednesday to Tuesday). Built to be honest about what it doesn't know: • A price more than a day old is never presented as current. ShopSaver fetches it fresh, or says it couldn't. • Basket comparisons only name a cheaper chain when enough of the list could be compared, and products are compared only when they're shown to be the same quantity. • Forecasts are probabilities, not promises. With too little history, it says so instead of guessing. Free, read-only. Independent: not affiliated with Coles, Woolworths or ALDI. Prices can differ in store.

How to connect

  1. Claude: Open the ShopSaver listing in the Claude connectors directory and choose Connect. Claude listing ShopSaver exposes 4 tools here.

Tools

compare_basket

Claude
Price a whole basket against another chain, line by line. Built for exactly this: "would this shop have been cheaper elsewhere?" Two fields per line, and they mean different things: status did we find an equivalent? (matched / no_candidates / no_match_above_threshold / matched_but_no_price / matched_but_not_comparable / matched_but_stale / missing_source_price / unknown_product_id / invalid_input) verdict cheaper | dearer | same — PRESENT ONLY when status=="matched" (or, with `allow_stale, on a matched-but-aged line). There is deliberately no "unknown" verdict. An absent key cannot be misread as a price finding, and misreading "we couldn't match this" as a price finding is precisely the error this tool exists to prevent. NO VERDICT UNLESS THE BASKET SUPPORTS ONE. totals.conclusive says whether cheaper_chain is trustworthy. It is false — and cheaper_chain/delta_percent are null — when too few lines are comparable (coverage_percent below min_coverage_percent, default 50), when there are no comparable lines at all, when the live refresh stopped early (live_search_stopped:*), or when it ran and a chain failed (live_search_partial). conclusive_reason names the first blocker. A verdict read off 3 of 67 lines is a fabrication with a number attached. totals.source_total/target_total/delta and coverage_value_percent (by dollars, not line count) are always reported and always honest over the comparable lines. Totals cover comparable lines ONLY. unmatched_value reports what was left out, so a basket delta is never quoted as if unmatched lines were free. unmatched_by_status breaks the exclusion down by status. In unit_price mode a line total is the cost of the SAME QUANTITY of goods at each side's unit price — not the target pack price, which is a different amount of food by design in that mode. Loose produce / variable weight: per-kg against per-each is deliberately withheld as matched_but_not_comparable when the two shelf prices are not shown to be for the same quantity. Give such a line weight_kg with a per-kg price (or line_total with qty) so the unit basis is the weight rather than "each". Args: items: Basket lines. Each is either a description {name, brand?, size?, qty?, price?, line_total?, weight_kg?, line?} or a catalogue lookup {store_product_id, ...}: a store_product_id resolves name/brand/size from our stored catalogue at source_chain (filling only the fields you left blank) so a line can be identified exactly instead of by name. price is the PER-UNIT price; pass line_total instead when that is what the receipt shows, or weight_kg with a per-kg price for a variable-weight line. target_chain: Chain to price the basket against. source_chain: Chain the basket came from. Required to resolve a store_product_id. mode: "same_size" compares like-for-like pack sizes. "unit_price" allows different pack sizes and compares $/100g — results are marked approx where a size was a range. min_confidence: Defaults to 0.8, or 0.85 in unit_price mode where pack size no longer discriminates. explain_unmatched: Include near-miss reasons for unmatched lines. include_unmatched: Return unmatched lines too. Off by default: the returned lines are the verdict-bearing ones only (every exclusion is still counted in totals.unmatched*). Turn it on to get the full per-line detail. allow_stale: Quote a verdict for a matched line whose target price is older than the 24h freshness window, marking it stale: true with its age_days. Off by default: an aged price is not a current price, and quoting it as today's is the same fabrication-with-a-total this tool exists to avoid. Turn it on when an aged comparison beats a gap, and read age_days on every line you use. min_coverage_percent: Comparable-line coverage below which no cheaper_chain verdict is stated (clamped 0-100). Defaults to the server's basket_min_coverage_percent` (50). max_items: Guard against oversized baskets. live_search: Let a line the stored catalogue …

get_price_history

Claude
Get the price history (recorded price rows) for one product. Use this to answer "how has this product's price moved?" — past prices, recent specials, or whether a price is new. It returns the raw rows for a single `product_id (grab one from search_products) in newest-first order; each row is one recorded price with its recorded_at timestamp. Two ways to bound it that are easy to confuse: - days caps the NUMBER OF ROWS returned, not a date range. A product repriced twice a day yields ~15 days of history at days=30. - since_days is a real WINDOW: only rows recorded within the last N days. Clamped to 1-730 (the retention ceiling). Use since_days when you want "the last week of prices"; use days` only to cap how many rows you receive. Args: product_id: Product to fetch history for. days: MAX ROWS returned (clamped 1-365), not a time window. since_days: Only return rows recorded within this many days (clamped 1-730).

predict_price

Claude
Predict a product's price over the coming catalogue weeks. Use this to answer "is this worth waiting for?" — will it still be on special next week, and what will it cost when the promotion ends. It reads one product's stored price history (`product_id from search_products) and fits a two-state semi-Markov model over Australian grocery catalogue weeks, which run Wednesday to Tuesday. HOW IT WORKS. Every price change in the stored data lands on a Wednesday: a product sits at its regular price, drops to a promotional price for a catalogue week, returns, and repeats. So the model tracks two things — the state (on promotion or not) and how many consecutive weeks that state has lasted — and estimates, from this product's own completed runs, the chance the current run ends next week:: h_s(k) = P(this run flips after week k | it has already lasted k weeks) A SKU with too few completed cycles to measure that borrows its chain's rate. The forecast is a forward pass over that chain, so each week comes back as a PROBABILITY, not a verdict. WHY NOT "IT ALTERNATES WEEKLY". That rule is right about two thirds of the time and silently wrong the rest, because a chain sometimes skips a promo week — the product stays at the regular price for two weeks instead of one. A rule that commits to alternation inverts on that week and stays inverted for the rest of the horizon. Walked forward over 15 Wednesday origins on the tracked cyclic SKUs, forecasting only from rows before each origin, this model scored 0.78 day-level state accuracy (Brier 0.144) against 0.70 for the alternating rule and 0.49 for persistence. READING IT. regime says which question was answerable: - cycle — a high-low pattern exists and weeks carries a probability per catalogue week. p_special is the chance the product is on promotion that week; expected_price weights the two prices by it and likely_price is the more probable of the two. - flat — no promotion has ever been observed for this SKU, so the price is a constant. Every ALDI product is flat: ALDI runs no specials. So are home-brand basics at the other two chains. - insufficient_history — fewer catalogue weeks stored than the model needs. weeks is empty and basis.reason says why; do not read a cycle into it. basis.reliability is the disclosure that matters most: high means six or more of this SKU's own promo cycles were measured, low means the hazard table is mostly its chain's prior and the numbers are weak. The residual error is the chain's discretionary decision to skip a promo week, which no amount of history predicts. Read the probability, not just likely_price. current is the last OBSERVED state, not a live quote: check current.as_of and current.stale. The first block in weeks is the remainder of that same catalogue week (basis="observed") — transitions only land on Wednesdays, so that price holds until Tuesday. Args: product_id: Product to predict. Get one from search_products. horizon_days: Days ahead to cover (clamped 1-28 — two catalogue weeks is where the model was measured; accuracy decays past that). Returns a JSON object: api_version, product, regime, current, levels (the regular and promotional price the model switches between), weeks, product_state (this SKU's fitted hazard table, so a caller can audit it), basis and diagnostics`.

search_products

Claude
Search products by name, brand or chain; fetch live only when needed. PURPOSE. This is the primary lookup tool: answer "how much is X here?" with the current price per store, cheapest first. One call returns a ranked, cross-chain answer — you connect them, you do not need a second tool. WHEN IT GOES OUT TO INTERNET. Almost never. Queries the catalogue already covers are answered from storage in milliseconds with zero outbound traffic. Only two situations trigger a live fetch from all three supermarkets (≈11s, once per query per day): 1. the query matches NO stored product at all, or 2. a matched product's latest price is older than a day (the daily sweep has not touched it recently, so its price can't be trusted as current). The fetch persists fresh rows, and this call then answers from them. Stale rows (a price older than a day, e.g. a product the sweep no longer finds) are NOT served — you never receive a 24h+ price presented as current. BATCHED FETCHES. `queries runs the same read-through path, one wave at a time. A wave is one locked sweep scraping up to SWEEPER_CONCURRENCY queries at once, so a batch is cheaper in wall clock than the same queries fetched one by one. The phase is bounded by wall clock (batch_live_search_deadline_s, 20s — set below the MCP transport timeout so the call returns a partial answer instead of hanging up at the client), and each wave is given only what is left of it. A query the deadline runs out on is answered from storage and its diagnostics carry live_search.reason="deadline_exceeded"; so a batch may return some queries fetched fresh and later ones not looked for, each saying which. READING THE EMPTY RESULT. By default the response is an object (api_version + results + diagnostics), so an empty results is never bare: diagnostics always carries matched (products the predicates hit in the scanned pool), returned, limit_applied, and — when a live fetch ran or was declined — live_search.reason. An empty result therefore says *why*: nothing matched, a live fetch was declined (fetched recently / daily budget spent / scraper busy), or rows were withheld as stale. Before concluding "not stocked", read those fields; a term that matches no product name and a fetch that was declined are different claims from "we checked and it isn't there". diagnostics otherwise also reports which terms matched, when each chain was last scraped, and a verdict. ORDERING. Results are cheapest-first on comparable unit price, drawn fairly across chains — every chain contributes its cheapest before any chain contributes its second — so "where is X cheapest?" is answered by the first row per store. Every chain with a comparable row is guaranteed a seat (its cheapest, on any basis) before any chain's second row is drawn, so a chain whose matches all sit on a non-dominant basis still appears rather than reading as "not stocked". COMPARE on price_per_100g / price_per_100ml / price_per_each, NEVER unit_price, which each chain formats in its own units ("100G" vs "per 100 g"). Approximate/range rows (unit_price_approx, e.g. random-weight meat) sort after precise rows, because their midpoint can sit below a genuinely cheaper precise row. Rows you cannot compare (per-item next to per-weight, or an unparseable size) come after the ranked ones. diagnostics.scan_truncated means more products matched than were scanned — the top row is the cheapest of a slice, not of the whole aisle; narrow the query before quoting it. MATCHING. Per-term AND against name + brand together (the chains disagree about where the brand lives — Coles in a column, Woolworths in the name), so every term must appear in the combined name+brand. Args: query: Search terms, whitespace separated. Every term must match. queries: Run several searches at once. Returns a by_query` map instead of a bare list. Max 25; each query runs the read-through live-search path, bounded by a wall-clock deadline. chain: Restrict to 'coles', 'woolworths' or 'aldi'. is_on_special: Keep only rows whos…

Similar Apps

Listed on MCP App Store

Listed on MCP App Store

Add this badge to your site or README. It links to this page.

HTML

<a href="https://mcp-app-store.keenethics-labs.com/app/shopsaver"><img src="https://mcp-app-store.keenethics-labs.com/badges/listed-on-mcp-app-store.svg" alt="Listed on MCP App Store" width="220" height="44"></a>

Markdown

[![Listed on MCP App Store](https://mcp-app-store.keenethics-labs.com/badges/listed-on-mcp-app-store.svg)](https://mcp-app-store.keenethics-labs.com/app/shopsaver)