A free MCP server that gives your AI agent USDA-grounded nutrition data. Stop letting it guess at your calories.
LLMs generate text; they don't retrieve facts. Even with USDA data in their training mix, the numbers they produce are interpolated guesses, not lookups.
An MDPI Nutrients study (2025) found GPT-4 underestimated calories by 36%, fat by 48%, and sodium by 53% in meal-photo nutrient estimation.
NutriBench reports 80% of LLM nutrition errors are carb predictions; the best model scored 51% accuracy.
Repeated queries returned different answers in a chatbot consistency study.
An LLM parses your ingredients at temperature 0. The nutrients are then looked up in USDA FoodData Central (2M foods) and totaled by deterministic code. Same input, same output, every time. The model never invents a number; it reads one.
Every ingredient comes back with match and conversion confidence between 0 and 1. Low scores flag fuzzy USDA matches or ambiguous portions. You know when to trust the number and when to double-check.
Plug NutrientAPI into your AI agent once, then use plain language.
analyze_recipe via NutrientAPIanalyze_recipe via NutrientAPIget_nutrition via NutrientAPIMCP traffic counts against the same plan quotas as direct API calls.
Works with Claude, ChatGPT, and any MCP-compatible client.