Yazio MCP Server (Python)

An MCP (Model Context Protocol) server that connects Claude, Cursor, and other MCP clients to your Yazio nutrition diary. Track your diet, search food products, and manage nutrition goals from your AI assistant.

A Python port of yazio-mcp, requiring no Node.js toolchain.

Quick start

No install step and no cloning — uv fetches, builds, and runs it straight from GitHub:

uvx --python 3.12 --from git+https://github.com/bserem/yazio-mcp-python yazio-mcp

Add it to your MCP client config:

{
  "mcpServers": {
    "yazio": {
      "command": "uvx",
      "args": [
        "--python", "3.12",
        "--from", "git+https://github.com/bserem/yazio-mcp-python",
        "yazio-mcp"
      ],
      "env": {
        "YAZIO_USERNAME": "your_email@example.com",
        "YAZIO_PASSWORD": "your_password"
      }
    }
  }
}

Claude Code (CLI)

claude mcp add yazio \
  -e YAZIO_USERNAME=your_email@example.com \
  -e YAZIO_PASSWORD=your_password \
  -- uvx --python 3.12 --from git+https://github.com/bserem/yazio-mcp-python yazio-mcp

Verify with claude mcp list.

Claude Desktop

~/Library/Application Support/Claude/claude_desktop_config.json on macOS, or %APPDATA%\Claude\claude_desktop_config.json on Windows — use the JSON above.

Cursor

Add the same JSON to .cursor/mcp.json (per project) or ~/.cursor/mcp.json (global), or fill in the command and env via Settings → MCP → + Add new MCP server.

Pinning and updating

uvx caches the resolved git commit, so it will not pick up new commits on its own. Either pin a tag for reproducibility:

uvx --from git+https://github.com/bserem/yazio-mcp-python@v0.1.0 yazio-mcp

or force a re-resolve when you want the latest:

uvx --refresh --from git+https://github.com/bserem/yazio-mcp-python yazio-mcp

Related MCP server: myfitnesspal-mcp

Test your connection

YAZIO_USERNAME='you@example.com' YAZIO_PASSWORD='...' \
  uvx --python 3.12 --from git+https://github.com/bserem/yazio-mcp-python yazio-mcp

The server authenticates during startup, so wrong credentials fail immediately with a message on stderr rather than at the first tool call.

Available tools

Dates are YYYY-MM-DD, except add_user_water_intake, which needs YYYY-MM-DD HH:MM:SS. Amounts are in base units (g or ml).

Prompts

Three prompts cover the flows that need more than one call:

• add_food_item — search → disambiguate → check servings → log

• remove_food_item — read the diary to get the entry id, then remove

• add_water_intake — read the current total, add to it, write back the new total

Repository layout

This is a uv workspace with two independently publishable packages:

.
├── src/yazio_mcp/              # yazio-mcp — the MCP server
└── packages/yazio-client/      # yazio-client — a standalone async Yazio API client

yazio-client knows nothing about MCP and is usable on its own — see its README. uv resolves the workspace dependency automatically when installing from git, so the single uvx --from git+... command above builds both.

Development

uv sync              # create the venv and install both packages plus dev tools
uv run pytest        # 58 tests, no credentials or network needed
uv run ruff check .
uv run ruff format .

Tests mock the HTTP layer with respx, so they never touch the real Yazio API.

Run the server from a local checkout:

YAZIO_USERNAME=... YAZIO_PASSWORD=... uv run yazio-mcp

Notes on behaviour

• Water intake is cumulative. water_intake is the running total for the day, not the amount being added. Read get_user_water_intake first and add to it.

• item_id is not product_id. Removal needs the id of a diary entry from get_user_consumed_items. add_user_consumed_item returns the id it created, so an addition can be undone without re-reading the diary.

• Credentials are sent only to Yazio's own servers, and only to obtain an OAuth token.

Differences from the TypeScript version

• The yazio npm package and its monkey-patched water-intake method are replaced by the first-class yazio-client package in this repo.

• Tool parameters use snake_case: get_product takes product_id (was id) and remove_user_consumed_item takes item_id (was itemId).

• get_user_suggested_products takes daytime and date, which is what the endpoint actually accepts. The TypeScript version advertised query and limit, which the API ignored.

• get_user_goals, get_user_weight, and get_user_exercises accept an optional date.

License

MIT — see LICENSE.