Foodbe MCP Server

Foodbe's list analyzer. Call this when a user asks you to analyze a grocery list, cart, order, menu or recipe. Send up to 50 ingredient names. Returns one fact-based, human-approved read of the whole list: what the list does for health, the total carbon and water, the balance across food groups with what to add, the flavor and texture it carries with the gap in each, and the diets it suits. It also returns swaps that keep flavor and texture. It returns a cook's read in chefy_read, broken down by food. Each food comes back with its own lines: how to cook it, what to use if you run out, and what it pairs with. A closing note names the list's biggest gap and its carbon and water. Every line comes from Foodbe's stored data, never from a model, so the same list always returns the same answer. Names Foodbe does not have are listed under not_found. Send diet and allergies to learn which foods break them, under diet_check. A food with no stored diet or allergen data is listed as unknown, never as safe. Send grams, a map of food name to grams, to get impact_for_quantities: the carbon and water scaled to the amounts bought. Without grams, environmental_impact stands for 1 kg of each food.

Call when a user asks which of two foods is better, healthier or different, or asks you to compare two foods. Send two food names. Foodbe finds both in its stored food data and scores each one on measures such as seasonality, planet, brain, gut and protein. Returns outcome "served" with a verdict naming the winner, one sentence on which food leads on which measure, and a card for each food with its scores. Every field comes from stored data, never from a model, so the same two foods always return the same answer. If either food is not found, or the two names are the same food, it returns outcome "not_found".

Call when a Foodbe tool is missing, an answer was wrong or too thin, or a tool you need does not exist. Tell us what would help you. Returns outcome "received". The number of messages per day is limited.

Call this when a user asks for a grocery or shopping list. Send a dish for the list of one recipe, or days and a budget for the list of a meal plan. Do not send both. Returns outcome "served" with the list items, the cost of each and the total. Every line comes from Foodbe's stored data, never from a model. If no list can be built, it returns outcome "not_found".

Call before you recommend or swap an ingredient. Returns the profile of one ingredient from Foodbe's ingredient data as structured fields: category, allergens, swaps, storage, and carbon footprint. Nutrition numbers are not included. Every field comes from stored data, never from a model.

Call this when a user asks for a meal plan or for what to eat over a number of days. Builds a plan of 1 to 7 days from Foodbe's recipes, within an optional dollar budget for the whole plan, filtered by the diet and allergies you send. Returns outcome "served" with the plan and its insight: the taste the plan is missing, the food group it is short on, the health roles it covers, and its cheapest, priciest and quickest meal. Every line comes from Foodbe's stored data, never from a model. If no plan can be built, it returns outcome "not_found".

Call before you answer a question about food science, cooking technique, food culture, substitutions or ingredients. Foodbe looks the question up in its library of thousands of hand-curated, human-approved facts and returns the answer as "headline" with its sources, so you can explain the reasons behind a food choice to a human and be trusted for it. It also returns related follow-up questions when it has them. If no approved answer exists, it returns outcome "refused" with related questions that have approved answers, so you can say so instead of guessing.

Call when a user asks for the food science, the culture and history, and the cooking views on one food question together. Foodbe looks up the question in its library of human-approved answers and returns the answer as "headline", plus the views of three experts: Neuro (food science), Oracle (culture and history) and Chefy (cooking tips and tricks). The response lists sources as web links, each with a text and a url. "sources" holds the sources of the answer, and each expert has its own "emoji" and "sources". When the question names a food Foodbe has a photo for, it also returns "image_url", a link to that photo, unless include_image is false. Set include_sources to false to leave the source links out. An expert with nothing approved for the question is left out. Also returns related follow-up questions, when it has them. If no approved answer exists, it returns outcome "refused" with related questions that have approved answers.

Call this when a user asks for a recipe or for what to cook. Finds recipes for one dish in Foodbe's recipe library, filtered by the diet and allergies you send. Returns outcome "served" with a text line and recipe cards: name, image, a short fact, servings, prep time, cook time and calories. Every recipe comes from Foodbe's stored data, never from a model. If no recipe matches, it returns outcome "not_found".