Cook (Cooklang recipes)

cook-mcp

An MCP server that gives your AI agent (Claude Code, Claude Desktop, Cursor, ChatGPT, or any other MCP client) access to your Cooklang recipe collection. Recipes stay as plain .cook and .menu files in a folder you own. The agent can read, search, validate and write them, build shopping lists, track a pantry and render reports, all locally and without an account. With a cook.md login it can also compute nutrition and import recipes from photos and social links.

Install

Claude Code:

claude mcp add cook -- npx -y @cookmd/mcp

Any client that takes an mcpServers config:

{  "mcpServers": {    "cook": {      "command": "npx",      "args": ["-y", "@cookmd/mcp"],      "env": { "COOK_RECIPES_DIR": "/path/to/recipes" }    }  }}

If COOK_RECIPES_DIR is not set, the server uses the workspace folder your client shares, or the folder the client was started in. See How the recipe folder is chosen.

Supported platforms: macOS (arm64, x64) and Linux (x64, arm64, glibc 2.35 or newer). There are no Windows builds yet.

Setup notes for specific clients: https://cook.md/help/mcp

Claude Code plugin

Claude Code users can install the cooklang plugin instead of adding the server by hand. It lives in cooklang/cooklang-skills and bundles this server plus the skills below:

/plugin marketplace add cooklang/cooklang-skills/plugin install cooklang@cooklang-skills

The plugin starts the server with your Claude Code project folder as the recipe root, and Claude Code picks the right skill from what you ask ("plan dinners for next week", "is this recipe valid?"). If you already added the server with claude mcp add cook, remove that entry so you don't run two copies.

The skills (also served by this server as cooklang://skills/<name> resources):

skills/ in this repo is the canonical copy; the plugin repo syncs from it. Each skill is skills/<name>/SKILL.md.

Tools

Free (local, no login)

Cook Basic / Pro (cook.md login)

import_recipe from a web page or pasted text works without a login. Photos and social-media links need a cook.md account and use your import allowance. It returns Cooklang text and does not save it; the agent validates it and calls write_recipe. Nutrition functions inside render_report also need Cook Basic or Pro.

Prompts and resources

Prompts: meal-planning, shopping-list, pantry, import-recipe, edit-recipe, nutrition-report, nutrition-goals, scale-recipe.

Resources:

• cooklang://spec: the Cooklang specification
• cooklang://syntax: a syntax reference
• cooklang://menu-format: the .menu meal plan format
• cooklang://skills/<name>: working guides for the agent, one per skill in the table above: cooklang-editing, cooklang-validation, export-recipe, meal-planning, metadata, nutrition-goals, nutrition-reports, organize-collection, pantry, recipe-import, recipe-search, report-authoring, scale-recipe, shopping-list

Safety

• Writes stay inside the recipe root. Paths outside it are refused.
• Every write is validated first; invalid Cooklang is not saved.
• The deprecated >> metadata syntax is refused. Use YAML frontmatter.
• There is no delete tool. The agent cannot remove your files.

Known limitations

• Pantry subtraction in shopping lists only works when units match. For example, 1 kg in the pantry does not cancel 200 g in a recipe. This is a limitation of cookcli-core.
• Symlinks inside the recipe folder are mostly not followed for reads and listing. A symlink requested by bare name without an extension, or reached through a recipe's @./ reference, may still be followed. This only matters if you put symlinks pointing outside the folder into your recipes.
• No Windows builds yet.
• No delete tool.

Environment variables

How the recipe folder is chosen

1. COOK_RECIPES_DIR, if set. Nothing else is consulted.
2. The client's MCP roots. If the client supports roots (it shares its open workspace folders with the server), the first file:// root that is an existing folder becomes the recipe root. The server asks on the first recipe tool call and again whenever the client reports that its roots changed. If the client doesn't answer within 5 seconds, the server uses the working directory and asks again later (at most once every 30 seconds).
3. The folder the client started the server in.

A root is a folder you opened on purpose, so it is used unless it is /, your home folder or inside an agent plugin install folder (for example ~/.codex/plugins/cache/..., ~/.gemini/extensions/... or the folder CLAUDE_PLUGIN_ROOT points to). The working directory is checked more strictly: it is also refused when it, or a folder up to four levels above it, has .claude-plugin/plugin.json, an Agent Plugins plugin.json or gemini-extension.json, because a plugin that starts the server in its own install folder would otherwise expose the wrong files and write recipes into a folder that is wiped on update. When nothing usable is found, recipe tools say that no recipe folder is set, why, and that COOK_RECIPES_DIR fixes it; auth_status shows recipe_root_source: "unset". Otherwise recipe_root_source is env, roots or cwd.

What each client does:

• Claude Code sends its project folder as a root (verified), so nothing to configure.

• Codex sends no roots (verified with 0.160.1) and may start the server in a plugin folder, so set COOK_RECIPES_DIR:

  sh

  codex mcp add cook --env COOK_RECIPES_DIR=/path/to/recipes -- npx -y @cookmd/mcp

• VS Code and Cursor document support for roots (not verified here). If recipe tools say no folder is set, set COOK_RECIPES_DIR in the server's config.

Things to ask

• "Plan dinners for next week from my recipes and make the shopping list."
• "What can I cook with what's in my pantry?"
• "Check my whole collection for broken references."
• "Import https://example.com/some-recipe as a recipe."
• "How much protein is in this week's plan?" (needs Cook Basic or Pro)

Building from source

cargo build --releasecargo test

The binary is target/release/cook-mcp. It speaks MCP over stdio. cook-mcp login and cook-mcp logout manage the cook.md login from a terminal.

Migrating from nutrition-mcp

@cookmd/nutrition-mcp keeps working: it is now a thin shim that runs @cookmd/mcp. To switch, change the package name in your MCP config to @cookmd/mcp. Your login carries over.

Two things changed. Recipe tools need COOK_RECIPES_DIR (old configs did not set it), or start the client in your recipe folder; without either, recipe tools tell the agent that no recipe folder is set. And render_report paths are now relative to the recipe folder.

License

MIT