yandex-lavka-mcp

An MCP server that lets an AI assistant order groceries from Yandex Lavka — search products, build a cart, and place a real order — with an explicit human confirmation before any money is charged.

What it does

Money safety. Placing an order is a deliberate two-step flow: checkout_preview returns the full summary and charges nothing; confirm_order(confirmed_total) refuses unless a preview was just run and you pass back the exact total it showed. Change the cart and the preview is invalidated — you must preview again.

3-D Secure. confirm_order submits the order and charges the on-file card, then polls payment status. If your bank requires 3-D Secure, payment_status comes back wait_user_action and a redirect_url is returned — open it to finish paying (a headless charge cannot complete 3DS). cancel_order(order_id) cancels.

Multiple locations / cities. Catalog, prices and cart are location-scoped. use_address("Дача") switches to a saved address; set_delivery_address("Казань, улица Баумана, 1", flat="12") works for any address in any city (it geocodes via Lavka's own address search).

How it's built

• Python 3.12+ · FastMCP · httpx.
• client.py — the async API client (session auth, CSRF, request building, trims huge payloads).
• endpoints.py — every API path in one place (overridable from config, no code change).
• server.py — the MCP tools the assistant sees.

The API sits under https://lavka.yandex.ru/api/v1/providers/* (plus /api/v1/orders/submit for placing orders). Requests need the CSRF token from the homepage HTML plus X-Lavka-Web-* headers — the client handles this.

Setup

1. Install

uv venv && uv pip install -e .

2. Provide your Yandex session (one time)

Log into Lavka in your browser first, then get the session cookies into ~/.config/yandex-lavka-mcp/config.json.

macOS — pull cookies straight from Chrome (one Keychain prompt → Allow):

uv pip install -e '.[browser]'
python scripts/extract_chrome_cookies.py          # auto-detects your profile

Any OS — paste the Cookie header from DevTools (Network → any lavka.yandex.ru request → Request Headers → Cookie):

python scripts/import_cookies.py --header "Session_id=...; yandexuid=...; L=..."

Session cookies expire — re-run when calls start failing with "Lavka session is not authorized".

3. Set a delivery location

Copy config.example.json to ~/.config/yandex-lavka-mcp/config.json and edit, or set it from the assistant with use_address / set_delivery_address. The catalog only works once a location is set. Smoke-test:

python scripts/smoke.py "молоко"

4. Register with your assistant

Claude Code:

claude mcp add yandex-lavka -- uv run --directory /path/to/yandex-lavka-mcp yandex-lavka-mcp

Claude Desktop (mcpServers):

{
  "yandex-lavka": {
    "command": "uv",
    "args": ["run", "--directory", "/path/to/yandex-lavka-mcp", "yandex-lavka-mcp"]
  }
}

Remote deploy (order from your phone)

By default the server speaks stdio (local clients). Set YANDEX_LAVKA_MCP_TRANSPORT=streamable-http to expose it over HTTP so a hosted instance can back a claude.ai custom connector (phone / web).

A prebuilt Dockerfile is included. Secrets are injected at runtime — never baked into the image:

docker build -t yandex-lavka-mcp .
docker run -p 8000:8000 \
  -e YANDEX_LAVKA_MCP_TRANSPORT=streamable-http \
  -e YANDEX_LAVKA_MCP_CONFIG_JSON="$(cat ~/.config/yandex-lavka-mcp/config.json)" \
  yandex-lavka-mcp

(The image defaults to stdio; the TRANSPORT env above switches it to HTTP.)

Environment variables

Authentication (any OIDC provider)

A public endpoint spends real money, so protect it. claude.ai's custom connector UI only supports OAuth (no static bearer / custom header — that works only in Claude Code/Desktop). This server is a provider-agnostic OAuth 2.1 resource server: point it at any OpenID-Connect provider (Zitadel, Keycloak, Auth0, Google, …) and it validates JWT access tokens against that provider's JWKS and advertises it via OAuth protected-resource metadata.

Enable it by installing the server extra (pip install '.[server]', already in the Docker image) and setting:

A network-exposed HTTP transport refuses to start unless OAuth is configured (it spends real money). Set YANDEX_LAVKA_MCP_ALLOW_INSECURE=1 only if you front it with your own auth. Leaving OAuth unset is allowed for loopback/local use.

Session cookies expire; when calls start failing, re-capture them and update the YANDEX_LAVKA_MCP_CONFIG_JSON secret. There is no headless Yandex login. The captcha pass (below) is a separate secret, so it survives this.

Captcha from a server

Yandex may start answering a hosted instance with a captcha — decided by the server's IP, not the session (the same cookies keep working from home). Tools then fail with "Yandex anti-bot returned a captcha". Solve it once through that IP:

ssh -f -N -D 1080 <your-server>     # SOCKS proxy out of the server's IP
uv run --with playwright python scripts/solve_captcha.py --proxy socks5://127.0.0.1:1080

A throwaway Chrome window opens on the captcha; solve it. The script prints the spravka cookie to stdout — set it as the YANDEX_LAVKA_MCP_SPRAVKA secret (pipe it straight in) and restart. It lasts about 30 days; repeat when the captcha error comes back.

Develop

uv pip install -e ".[dev]"
pytest

One account = one cart

Lavka keeps a single server-side cart per account, guarded by an optimistic cartVersion. This server serializes its own cart writes and retries on version conflicts, so parallel tool calls in one session are safe. But don't drive the same Yandex account from two places at once (e.g. this server and a second MCP session, and the Lavka app): they all write the one shared cart, and you'll see items from the other writer appear in yours. Use a single client at a time.

Security & privacy

• Cookies and address live only in ~/.config/yandex-lavka-mcp/config.json (chmod 600), git-ignored. Never commit them.
• The server never adds payment methods or changes account settings.
• Ordering always requires an explicit confirmed total.

License

MIT. Unofficial project, not affiliated with Yandex.