An AI-readable cocktail database

cocktail.glass isn't only 500 recipes for people to read. The whole catalogue is built for machines too. Connect an AI agent over the Model Context Protocol, pull structured JSON, or read any page as Markdown. No API key, no sign-up, no rate limits.

Connect an AI agent over MCP

The fastest way in is the Model Context Protocol (MCP) server. It's a remote, stateless endpoint at https://cocktail.glass/mcp. Point any MCP client at it and five cocktail tools — plus a guided prompt — appear.

One click

Add to Cursor Add to VS Code

Claude Code

claude mcp add --transport http cocktail-glass https://cocktail.glass/mcp

Claude Desktop, or any MCP client

{
  "mcpServers": {
    "cocktail-glass": {
      "url": "https://cocktail.glass/mcp"
    }
  }
}

That's it. No key to paste. The server exposes five read-only tools — search_cocktails, get_cocktail_recipe, find_cocktails_by_ingredient, find_makeable_cocktails, and random_cocktail — plus a cocktails_from_my_bar prompt.

What you can ask it

Five tools and a guided prompt cover the questions that come up around a home bar:

• “What can I make tonight?” Share the bottles you own — or a photo of your cabinet — and get back the cocktails you can make right now, plus the ones you're a single ingredient short of.
• “How do I make a Last Word?” A hands-free recipe reference while you're mixing: exact measures, glassware, steps, and garnish.
• “I bought a bottle of mezcal — now what?” Every cocktail that uses a spirit or liqueur you already own, so nothing gathers dust.
• “Suggest a few drinks for six guests.” Build a varied menu across styles, or get one random pick when you can't decide.

See it work

Ask your assistant a question in plain language and it picks the right tool. Here is a real exchange, end to end:

You ask “What can I make with Campari?”

The agent calls a tool

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "find_cocktails_by_ingredient",
    "arguments": { "ingredient": "Campari" }
  }
}

cocktail.glass returns

{
  "count": 29,
  "cocktails": [
    { "name": "Negroni",      "category": "Classic", "url": "https://cocktail.glass/negroni/" },
    { "name": "Boulevardier", "category": "Classic", "url": "https://cocktail.glass/boulevardier/" },
    { "name": "Jungle Bird",  "category": "Tiki",    "url": "https://cocktail.glass/jungle-bird/" },
    { "name": "Americano",    "category": "Classic", "url": "https://cocktail.glass/americano/" }
  ]
}

Four of 29 matches shown. MCP delivers the list as the text content of a standard tool result.

Your assistant answers “You have 29 options with Campari. The classics are the Negroni and Boulevardier, or a Jungle Bird for something tiki.”

Every result carries a real page URL, so the agent can link straight to the full recipe, with measures, steps, glassware, and garnish.

More ways in

Not every client speaks MCP. The same catalogue is open over plain HTTP:

/cocktails.json

The entire catalogue as one JSON file. All 500 cocktails, with ingredients, measures, and preparation steps.

Markdown content negotiation

Request any page with an Accept: text/markdown header and you get a clean Markdown rendering instead of HTML. It's far cheaper for an LLM to read.

/llms.txt

A structured Markdown map of the whole site, following the llmstxt.org convention.

WebMCP

Agents running inside the browser get the same five tools in-page through navigator.modelContext, with no connection step.

/.well-known/api-catalog · MCP server card

Machine-readable discovery documents (RFC 9727 and SEP-2127) so agents and registries can find all of the above on their own.

Everything here is read-only and free to use. Browse the recipes →