wolt-mcp

A thin Model Context Protocol server that exposes Wolt's public consumer endpoints to AI agents. Point it anywhere Wolt operates — the default coordinates are Tallinn, Estonia, but you can override per-call or via env vars.

Two tools, no magic:

• list_nearby — venues near a lat/lon, filtered by substring, rating, open/closed status.
• get_menu — full menu for a venue by slug, with prices, 30-day lows, and category structure.

It's read-only on purpose. Wolt's ordering API is gated behind merchant credentials; this server intentionally doesn't try to place orders.

Who this is for

You want an AI assistant (Claude Code, Cursor, Continue, Zed, or your own Agent SDK app) that can:

• Discover restaurants by cuisine, rating, or open status in your city.
• Read full menus with live prices into its context window.
• Watch favorite venues for deals — discount categories like ERIPAKKUMISED (Estonian), SPECIAL OFFERS, or KUUPAKKUMINE (monthly offer), plus lowest_price (the 30-day low used for EU compliance display).
• Draft weekly meal plans by composing items across a curated shortlist of venues.

Geographic coverage

Wolt's consumer-api.wolt.com is global; this server has no country hardcoding. Change WOLT_DEFAULT_LAT / WOLT_DEFAULT_LON (or pass lat/lon per call) to work anywhere Wolt delivers — Finland, Germany, Czechia, Estonia, Israel, Greece, Japan, and more. Verified live against Tallinn (59.4370, 24.7536); other regions should work with the same endpoints.

Install

git clone https://github.com/fogside/wolt-mcp
cd wolt-mcp
python3 -m venv .venv
.venv/bin/pip install -e .

Then register it with any MCP-capable client. For Claude Code, drop this at the project root:

{
  "mcpServers": {
    "wolt": {
      "command": "./.venv/bin/wolt-mcp",
      "args": [],
      "env": {
        "WOLT_DEFAULT_LAT": "59.4370",
        "WOLT_DEFAULT_LON": "24.7536",
        "WOLT_LANGUAGE": "en"
      }
    }
  }
}

For Claude Desktop, add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or the equivalent on Linux/Windows:

{
  "mcpServers": {
    "wolt": {
      "command": "/absolute/path/to/wolt-mcp/.venv/bin/wolt-mcp",
      "env": { "WOLT_DEFAULT_LAT": "59.4370", "WOLT_DEFAULT_LON": "24.7536" }
    }
  }
}

Any other MCP client that supports stdio transport works the same way — run wolt-mcp as the command.

Tools

list_nearby

Param	Type	Default	Notes
lat	float	$WOLT_DEFAULT_LAT
lon	float	$WOLT_DEFAULT_LON
radius	int (m)	3000	200–20000
query	str?	None	Substring match against name + tags
only_open	bool	False
min_rating	float?	None	0–10 scale
max_results	int	30
language	str	$WOLT_LANGUAGE or en

Returns a list of dicts: name, slug, id, online, rating, rating_volume, eta_minutes, price_range, tags, short_description, address.

get_menu

Param	Type	Default	Notes
slug	str	—	From list_nearby, e.g. vapiano-foorum
language	str	en
include_disabled	bool	False

Returns { slug, assortment_id, primary_language, selected_language, available_languages, categories: [{ id, name, slug, description, items: [...] }], uncategorised_items, item_count }. Each item has id, name, description, price, original_price, lowest_price, enabled, tags.

Prices are integers in minor units. 2390 = €23.90. Currency is not on items — infer from venue country.

Environment variables

Var	Default	Purpose
WOLT_DEFAULT_LAT	59.4370	Tallinn center
WOLT_DEFAULT_LON	24.7536
WOLT_LANGUAGE	en	Sent as Accept-Language
WOLT_MCP_LOG	WARNING	Python logging level

Example conversations

"Find the top five sushi places near me that are open right now." → list_nearby(query="sushi", only_open=True, min_rating=9, max_results=5)

"Pull Vapiano Foorum's menu and tell me what's vegetarian." → get_menu(slug="vapiano-foorum") then the LLM filters by description.

"Compare prices for pad thai between these three Thai venues." → one list_nearby + three get_menu calls.

"Summarise current deals at my favorite café." → get_menu — surface items in categories named like ERIPAKKUMISED / SPECIAL OFFERS / KUUPAKKUMINE, or with ⭐ in the name.

Detecting deals — an important gotcha

Wolt venues in practice use category membership as their deal signal much more often than original_price. A typical Estonian cafe will have a category named ERIPAKKUMISED ("special offers") or KUUPAKKUMINE ("monthly offer") containing the discounted items, while original_price on those items stays null.

When prompting your agent, don't rely on a price-diff — ask it to inspect category names and item names (⭐ prefixes are common) as the first-class signal. lowest_price is useful for longer-range price-tracking: it's the 30-day low that Wolt surfaces for EU price-transparency compliance.

Non-goals and limitations

• No ordering, cart, or checkout. Wolt's consumer cart/order flow requires authenticated user sessions + isn't in this MCP's scope. For commerce, use Wolt's merchant API (partner credentials required).
• No user-account actions. Can't read your order history or favorites.
• Language. The assortment endpoint often returns the venue's primary language (e.g. Estonian) even when en is requested — most venues don't publish auto-translated content. You'll see the actual selected language in the response's selected_language field.
• Rate limits. Wolt returns 429s under aggressive use. For personal use this doesn't matter; for anything heavier, add client-side pacing.

Terms of service

This project hits Wolt's public, unauthenticated consumer endpoints — the same ones wolt.com's web app calls. Automated access at scale is contrary to Wolt's Terms of Service. Use this for personal agent assistance, experimentation, or research. Don't build a scraper at scale with it.

Credits and sources of inspiration

• jonzarecki/wolt-sdk — the first published Wolt MCP. Inspired this project, though the published implementation was Israel-focused and had server↔client method mismatches at the time of this writing, which pushed us to rebuild rather than fork. MIT-licensed.
• OfiliPatrick/wolt-restaurant-api — early reverse-engineering of the consumer endpoints.
• Tomer Chaim's writeup and OzTamir's gist — how to discover undocumented web-app endpoints through DevTools.
• Built on FastMCP (v3.x) by the jlowin/fastmcp project.
• The broader Model Context Protocol spec from Anthropic.

Development

.venv/bin/pip install -e ".[dev]"  # once dev extras are defined
.venv/bin/wolt-mcp                 # runs the server on stdio — connect a client

Contributions welcome — especially venue-specific deal-detection heuristics for cities outside Estonia, or a venue_dynamic(slug) tool if you can find a working endpoint.

License

MIT.