sakenowa-mcp 🍶

The first Model Context Protocol server for Japanese sake (日本酒) — a flavor-space engine, not an encyclopedia.

Give any MCP-capable LLM (Claude Desktop, Claude Code, …) a real sense of taste for sake. It wraps the open Sakenowa Data Project dataset — 1,700+ breweries, 3,200+ sake, 1,300+ six-axis flavor charts, monthly popularity rankings — and turns it into tools the model can reason with: search, profile, flavor-similarity recommendation, and side-by-side comparison.

As far as I can tell, there is no other sake MCP server in existence — existing "brewery" MCPs cover Western beer only. This one is built around what makes sake searchable: its flavor vector.

---

Why this exists

Ask an LLM "find me something like 八海山" and it guesses from training data. With sakenowa-mcp it does the real thing: it pulls 八海山's six-axis flavor vector and returns the nearest sake in flavor space — and can steer the search drier, lighter, richer, or to the deliberate opposite.

The positioning, in one line:

Sakenowa MCP = a flavor-space engine for sake. It does not store tasting notes or prices; it makes the shape of a sake's taste computable.

The flavor model

Every rated sake has six normalized axes (0–1):

axis	日本語	meaning
f1	華やか	floral & vibrant
f2	芳醇	mellow & full-bodied
f3	重厚	rich & heavy
f4	穏やか	calm & gentle
f5	軽快	light & smooth
f6	ドライ	dry

Tools

tool	what it does
sync_sakenowa_data(force=False)	Fetch/refresh the dataset into a local cache; report scale & attribution. Auto-refreshes weekly.
search_sake(query, limit=10, area="")	Find sake by brand or brewery name → IDs. Kanji queries match directly; area filters by prefecture.
get_sake_profile(brand_id)	Six-axis ASCII radar, dominant tags, an estimated four-type class (薫/爽/醇/熟), popularity rank, and the spec fields the data doesn't contain.
find_similar_sake(brand_id, mode, limit=5)	★ core ★ Nearest sake by flavor vector. mode: similar / drier / sweeter / lighter / richer / more_aromatic / calmer / contrast.
compare_sake([id, id, …])	Compare 2–5 sake across all six axes with per-axis spread.

Plus a recommend_sake MCP prompt as a friendly entry point.

Quickstart

Requires uv. Python is provisioned automatically (3.10+).

git clone https://github.com/mame0001/sakenowa-mcp.git
cd sakenowa-mcp
uv sync          # create venv + install
uv run pytest    # run offline tests

Add to Claude Code

claude mcp add sakenowa -- uv --directory /absolute/path/to/sakenowa-mcp run sakenowa-mcp

Add to Claude Desktop

In claude_desktop_config.json:

{
  "mcpServers": {
    "sakenowa": {
      "command": "uv",
      "args": ["--directory", "/absolute/path/to/sakenowa-mcp", "run", "sakenowa-mcp"]
    }
  }
}

Then ask: "Find me 3 sake similar to 久保田 but a bit drier," or "Compare 八海山 and 獺祭."

Try it without an LLM

uv run python -m sakenowa_mcp.demo   # full showcase: sync → search → profile → similar → compare
# or a one-liner:
uv run python -c "from sakenowa_mcp import data, search; ds=data.get_dataset(); print([b.name for b in search.search(ds,'八海山',5)])"

Configuration

env var	default	meaning
SAKENOWA_CACHE_DIR	~/.cache/sakenowa-mcp	where the JSON snapshot is cached
SAKENOWA_TTL_SECONDS	604800 (7 days)	how long before the cache auto-refreshes

Honest limitations

• ~41% flavor coverage (≈1,335 of 3,250 sake). Tools tell you when a bottle has no chart.
• No bottle specs. The dataset has no polishing ratio (精米歩合), rice variety, SMV/日本酒度, acidity, ABV, price, or junmai/ginjo grade. The tools say so explicitly so the model doesn't invent them.
• Estimated four-type class is a heuristic, self-calibrated against the dataset median — not the official SSI sensory category.
• Names are Japanese; romaji/kana search is best-effort substring matching.

Data & attribution

Sake data comes from the Sakenowa Data Project (さけのわデータプロジェクト). It is free and permits commercial use but requires attribution to "Sakenowa." Every tool response includes that attribution — please keep it intact. This project is not affiliated with or endorsed by Sakenowa.

License

Source code: MIT (see LICENSE). Sake data remains under the Sakenowa Data Project's terms.