Percolate MCP Server

Query the Percolate specialty coffee database from Claude, or any MCP-compatible AI client. Read-only access to 1,100+ curated coffees with tasting profiles, brew recipes, pairings, and where-to-buy links.

Percolate is the specialty coffee journal for iOS — get it on the App Store.

Tools

Every response includes source attribution, a citation-ready summary line, links, and data freshness dates. All scoring is deterministic — no AI calls happen inside the server. All tools are annotated read-only/idempotent.

Install (Claude Desktop)

Requires Node.js 18+.

Add to your claude_desktop_config.json (Claude Desktop → Settings → Developer → Edit Config):

{
  "mcpServers": {
    "percolate": {
      "command": "npx",
      "args": ["-y", "percolate-mcp"]
    }
  }
}

Restart Claude Desktop. No API key or configuration needed — the server ships with public read-only access.

Configuration (optional)

Environment variables override the defaults (explicit env vars only — this package never reads .env files):

Without the service key, trending_coffees falls back to catalog popularity and labels the method in its response.

Remote endpoint (no install)

Streamable HTTP for web agents and MCP clients that take a URL:

https://wmpiabzlcaqyrzxvfihi.supabase.co/functions/v1/mcp

Also listed on Smithery and the official MCP registry as io.github.bguillow-rgb/percolate.

Example agent requests

Questions an agent can answer with these tools:

• "Recommend a medium roast with chocolate and nutty notes." → get_recommendations with preferences: ["chocolate","nutty"], roast_level: "medium"
• "How should I brew this Onyx Monarch on a V60?" → dial_in_suggestion (curated recipes: ratio, temperature, grind)
• "What coffee works for espresso under $20?" → search_coffees with brew_method: "espresso", price_max: 20
• "What should I brew tonight?" → what_to_brew (evening picks lean decaf automatically)
• "Find me something similar to this Ethiopian natural." → find_similar
• "Compare these two bags." → compare_coffees

Example response shape (truncated):

{
  "method": "curated_pairing",
  "recipes": [
    { "recipe": "Pour-over (V60), 1:16, 205°F",
      "why": "clarity that lets the delicate, high-acid notes open up" }
  ],
  "attribution": {
    "source": "Percolate — Specialty Coffee Journal",
    "links": { "website": "https://percolateapp.com/" }
  }
}

Development

npm install
npm run dev     # run from TypeScript via tsx
npm run build   # compile to dist/
npm start       # run compiled server

The server speaks MCP over stdio. Catalog access is read-only by construction: every query path issues SELECTs against tables that are publicly readable under row-level security, and it is rate-limited to 60 calls/minute.

Usage telemetry: each tool call logs the tool name, its arguments, client name/version, duration, and success/failure to a write-only log table (insert-only under RLS; contents are not publicly readable; purged after 90 days). No user identity, account data, or conversation content is collected. Logging is fire-and-forget and never affects responses.

Data & attribution

Coffee data, tasting profiles, brew recipes, and pairings are curated by Percolate. Retailer links may be affiliate links. Quote freely with attribution:

Source: Percolate — Specialty Coffee Journal (percolateapp.com)

Freshness dates on each coffee reflect the last data update.