Morrisons杂货数据MCP服务

Morrisons MCP Server

A self-hosted MCP (Model Context Protocol) server that scrapes Morrisons grocery data and exposes tools for product search, per-product label nutrition, and recipe costing. Answers questions like "how much will this recipe cost at Morrisons?" or "what does this product's label say per 100 g?"

MCP Tools

get_product_detail response

Morrisons' product payload carries no GTIN/EAN, so none is returned.

pick_products and the picker UI

pick_products(ingredients, max_results=8) takes food names without quantities (at most 15; 1–12 candidates each) and returns { ingredients: [{ ingredient, query, results: [ProductResult] }] }. Every ProductResult carries url (https://groceries.morrisons.com/products/<retailerProductId>, which Morrisons redirects to the slug URL). The tool declares _meta.ui.resourceUri = ui://morrisons/picker; that resource is the picker page (picker_html.py, MIME text/html;profile=mcp-app). It loads the MCP Apps browser client from unpkg and product images from groceries.morrisons.com, both declared in its CSP. Submit calls sendMessage with one line per ingredient: - <ingredient> → <name> (id <retailerProductId>, <url>), or none of these / not chosen. A failed search gives that ingredient an empty results list rather than failing the call.

shop_products and the shopping list UI

shop_products(items, max_results=6) takes up to 40 shopping-list lines. Each line is { name, links, quantity, amount }: name is searched for alternatives; links are the line's Usual products as Morrisons product URLs (slug or not) or bare retailerProductIds, Main product first; other shops' links are ignored. A line needs a name or a link; with links only, the Main product's name drives the search. quantity (1–99, default 1) is the packs to buy and amount (e.g. "1.2 kg") is shown as a hint.

It returns { items: [{ name, query, quantity, amount, usual_ids, gone_usual_ids, results }] }. results holds the Usual products first, in link order, then up to max_results (0–12) alternatives. A Usual product found in the search results is taken from there; otherwise it is looked up by id (MorrisonClient.get_product). gone_usual_ids lists the ones Morrisons no longer sells. Lines are searched concurrently, still one Morrisons request per 0.5 s. A failed search or lookup thins that line instead of failing the call.

The view (shop_html.py, ui://morrisons/shop):

• pre-ticks each line's first Usual product Morrisons can supply; a line without one starts skipped;
• sorts the alternatives by best match, price, or unit price (per kg, litre or each; Morrisons gives any unit price under £1 in pence, e.g. 90.0p/kg), with the Usual products always first;
• has a pack stepper per line and a running total (shelf prices, so ≈; Morrisons applies offers in the trolley).

Add to trolley calls openLink with:

https://groceries.morrisons.com/#add-to-trolley=<total>;<productId>:<packs>,<productId>:<packs>,...

productId is the product's UUID (what Morrisons' trolley takes), not the numeric retailerProductId. The fragment never reaches Morrisons' server. The view also shows the link to copy, for when the host can't open it or opens it somewhere the userscript doesn't run.

Trolley userscript

userscript/morrisons-trolley.user.js runs on groceries.morrisons.com. When a page opens with #add-to-trolley=…, it:

1. clears the fragment, so a reload can't add twice;
2. reads the page's session state (CSRF token, logged in or not) from a fresh copy of the home page;
3. asks "Add N products (≈ £T) to your Morrisons trolley?";
4. POSTs /api/cart/v1/carts/active/apply-quantity with [{productId, quantity}], first-party with the user's own cookies;
5. opens /basket.

How it behaves:

• Packs are added, not set. Morrisons' own remove button sends negative quantities to the same endpoint. So the script remembers each link it has added and flags a second use in the confirm.
• Logged out: the list is kept for 30 minutes and added after the user logs in, with one prompt rather than one per login page.
• Rejected: a damaged link adds nothing, and a refused update (any HTTP error) adds nothing and isn't remembered.

Install it once:

• iPhone/iPad: the free Userscripts app. Turn it on in Settings → Safari → Extensions and allow it on groceries.morrisons.com. Then open the raw script URL below in Safari and install it from the extension's menu. Log into Morrisons in Safari. iOS Chrome can't run extensions.
• Desktop: Tampermonkey or Violentmonkey; open the raw URL and install.

Raw URL (managers use it for updates too): https://raw.githubusercontent.com/cometto2007/morrisons-mcp/main/userscript/morrisons-trolley.user.js

The Morrisons iPhone app claims groceries.morrisons.com links (/, /basket, /products/*…), so with the app installed, links open there and the userscript never runs. This flow assumes the app isn't installed.

Setup

Endpoint

The server speaks MCP over Streamable HTTP at POST /mcp on port 8000 (host 0.0.0.0), in stateless mode: each request stands alone, so idle clients and server restarts don't break sessions. There is no SSE endpoint.

Environment Variables

All optional. Copy .env.example to .env to set them.

Docker (recommended)

docker compose up -d

The image (ghcr.io/cometto2007/morrisons-mcp, linux/amd64 and linux/arm64) listens on port 8000 inside the container. Compose publishes no ports; Traefik routes to it, so MCP clients use https://<traefik-host>/mcp.

Running Locally

# Install dependencies (Python 3.12+)
pip install -e ".[dev]"

# Start the server on http://0.0.0.0:8000/mcp
python -m morrisons_mcp.server

Example Usage

Search for products:

search_products("chicken breast", max_results=5)
→ [
    { name: "Morrisons Chicken Breast Fillets 600g", price: 4.50, unit_price: "£7.50/kg", ... },
    ...
  ]

Get nutrition data:

get_product_detail("108444543")
→ { found: true, name: "Morrisons British Whole Chicken Medium 1.45kg", pack_size: "1.45kg",
    net_quantity: { value: 1450, unit: "g" },
    nutrition_per_100g: { basis: "100g", basis_note: "as consumed", energy_kcal: 177, protein_g: 27.3, fat_g: 7.5,
                          saturates_g: 2.1, carbohydrate_g: 0, sugars_g: 0, fibre_g: 0, salt_g: 0.2, ... }, ... }

get_product_detail("999999999999")
→ { retailer_product_id: "999999999999", found: false, name: null, ... }

Cost a recipe:

cost_recipe(
    ingredients=["500g chicken breast", "1 tin chopped tomatoes", "2 cloves garlic"],
    servings=4,
    recipe_name="Chicken Tomato"
)
→ { total_cost: 6.25, cost_per_serving: 1.56, unmatched_count: 0, ingredients: [...] }

Pick products (MCP App):

pick_products(["chopped tomatoes", "garlic"], max_results=8)
→ { ingredients: [ { ingredient: "chopped tomatoes", query: "chopped tomatoes", results: [ { name, pack_size, price, image_url, url, ... }, ... ] }, ... ] }
# In the Claude app the cards render in the chat; Submit posts, as the user:
#   Morrisons picks:
#   - chopped tomatoes → Morrisons Chopped Tomatoes (400g) (id 105415501, https://groceries.morrisons.com/products/105415501)
#   - garlic: none of these

Architecture

MCP client / gateway
    │  Streamable HTTP POST /mcp (MCP protocol)
    ▼
morrisons-mcp (FastMCP server :8000)
    ├── SessionManager      ← anonymous Morrisons cookies
    ├── MorrisonClient      ← search + BOP endpoints
    │   └── ProductCache    ← SQLite TTL cache
    ├── IngredientParser    ← "500g chicken breast" → structured data
    ├── FuzzyMatcher        ← rapidfuzz token matching
    └── NutritionParser     ← BeautifulSoup HTML table parsing

Caching

The server caches API responses in SQLite to reduce load on the Morrisons website:

The SQLite database is stored at /data/cache.db, backed by a Docker named volume (morrisons_data) for persistence across container restarts.

Development

# Install with dev dependencies
pip install -e ".[dev]"

# Run tests
pytest tests/ -v

# Optional live check against the real Morrisons API (3 requests)
MORRISONS_LIVE=1 pytest tests/test_live_morrisons.py

# Browser tests of the shopping list view and the trolley userscript, in headless
# Chromium against stand-ins for Morrisons and unpkg (skipped without Playwright)
pip install -e ".[dev,browser]" && playwright install --only-shell chromium
pytest tests/test_shop_browser.py

Project Structure

src/morrisons_mcp/
├── server.py            # FastMCP app + 5 tool definitions
├── morrison_client.py   # Morrisons search + BOP API client
├── session_manager.py   # Cookie/session acquisition + refresh
├── cache.py             # SQLite async cache (aiosqlite)
├── ingredient_parser.py # "500g chicken breast" → ParsedIngredient
├── fuzzy_matcher.py     # Match ingredients to products (rapidfuzz)
├── picker_html.py       # The pick_products MCP App page (ui://morrisons/picker)
├── shop_html.py         # The shop_products MCP App page (ui://morrisons/shop)
├── mealie_client.py     # Optional Mealie pantry-staple check (cost_recipe)
├── nutrition_parser.py  # Per-100 g/ml label parsing + net quantity
└── models.py            # All Pydantic data models
userscript/
└── morrisons-trolley.user.js  # Adds shop_products picks to the user's Morrisons trolley