Nutrition intelligence API
Do you trust AI? Thought so. With Holobiont you get micronutrient details, sources you can inspect, and reference ranges grounded in verified sources. Turn meal photos into nutrition data and lab questions into cited answers, with the evidence behind each result.
Where the evidence is strongest
Reference-present nutrients detected. Amounts are still estimates.
Answers with source citations you can inspect.
Biomarker answers matching the library’s range in automatic checks.
Answers preferred in the blinded, two-order review.
Holobiont vs a plain GPT 6 prompt in our published benchmark. Explore the results
01 — What you can build
Each one is a single request. You send what your users have; the API returns structured, grounded data you can show, store or reason about.
Photo in, dish and items out: estimated grams per item, calories and macros, micronutrients, hidden oils and dressings, a confidence score and how much of the plate is backed by the nutrition database.
Items the database cannot match are marked as estimates, so your app always knows what it is showing.
Ask about a lab value in plain language. The answer is grounded in your biomarker library and knowledge base and comes back with numbered citations, follow-up questions and the tokens used.
Conversations persist, so a coach can continue where the user left off.
Point it at a product page or a whole Shopify or WooCommerce store. It reads the store data, decides what is a supplement, fills ingredients, dosage, safety and pharmacology, and attaches citations.
A full catalog of 500 products imports in one call.
Ask for references on a question, a supplement or a claim. The API searches academic papers, then judges each one for relevance (0–3) and for whether it supports or contradicts the claim.
Literature search itself is a plain lookup; judging the papers is the AI part.
02 — See it work
This is the real response for the photo on the left, replayed step by step. Nothing here is mocked: the items, grams, confidence, nutrients and the low-confidence oil are what the API returned.
A slice of toasted whole grain bread topped with mashed avocado and a sunny-side-up fried egg, seasoned with red pepper flakes and black pepper.
one large egg
a layer of mashed avocado
one thick slice of bread
oil used for frying the egg
Micronutrients
100% of the plate’s weight matched in the nutrition database · confidence 90 %
03 — Case study
HOLOHABITSHoloHabits is a home nutrition-testing platform: a finger-prick biomarker panel, an app and a coach that reads the results. When a user photographs a meal in the coach, the app sends it to the Holobiont API and shows what comes back.
The app posts the photo to /food/analyze with the user’s note and locale. The structured result maps straight onto its meal view.
Each item carries a confidence and whether its nutrients come from the database or an estimate. HoloHabits shows both, and lets the user correct portions.
The logged meal feeds the coach, which relates it to the user’s own biomarkers and daily targets.
04 — Accuracy
GPT 6 vs. Gemini 4 vs. Holobiont AI on the same meal photos and biomarker questions. Compare food recognition, nutrition estimates, reference ranges and cited answers. Choose a comparator to see the paired scores.
Every compared measure of both benchmarks, meal photos and biomarker questions, scaled to 0–100 and averaged with equal weight.
Holobiont scores 61.2 % higher in this comparison.
Dashed markers show the averages. Each connected pair is one measure. Better score.
Answers with source citations: 100 vs 0. Correct reference range stated: 70.7 vs 0.7. Review: cites sources: 68.4 vs 0. Answers the reviewer preferred: 75.3 vs 22. Micronutrients in the meal reported: 95.6 vs 50.3. Review: range interpretation: 74.4 vs 30.2. Reviewer score: 67.2 vs 25. Value placed correctly: 93.8 vs 74.8. Listed items that were really there: 74.5 vs 58.3. Review: actionable, safe advice: 59.2 vs 44.6. Ingredients on the plate recognised: 63.1 vs 48.6. Carbohydrate error: 55.2 vs 44.8. Micronutrient amount error: 55.4 vs 51.7. Fat error: 50.7 vs 49.3. Calorie error, relative: 64.2 vs 64.4. Portion weight error per item: 49.8 vs 50.2. Whole-plate weight error: 49.8 vs 50.2. Protein error: 49.1 vs 50.9. Plates within 20 % of weighed calories: 31.7 vs 35. Calorie error per plate: 47.7 vs 52.3.
Across 20 measures: ahead on 12, within noise on 7, behind on 1.
Every compared measure scaled to 0–100 (percentages as they are, errors as 100 minus the error, reviewer scores ×20, gram and kcal errors relative to the other arm) and averaged with equal weight. Ahead and behind count only differences whose 95 % interval excludes zero.
05 — Biomarker benchmark
GPT 6 vs. Gemini 4 vs. Holobiont AI on reference ranges, result interpretation and cited answers. The worked example below shows a plain GPT 6 prompt and Holobiont AI, unedited. The complete results compare the same questions across these systems.
The question
Is a Oleic acid C18:1 n-9 cis of 120 umol/L something to worry about?
Reviewer score 0.33 / 5
A value of oleic acid (C18:1 n‑9 cis) of 120 µmol/L isn’t enough on its own to say whether there’s a problem. Interpretation depends on the lab’s reference range, the type of test, and the rest of the fatty-acid panel and your health context. Oleic acid is a common fatty acid in foods and in the body; an isolated result is not usually a diagnosis. Check the reference range printed on your report and ask the clinician who ordered the test to interpret it alongside the other results. If you share the range and the reason the test was done (and any flagged values), I can help explain what the report says.
Reviewer score 4.67 / 5
Interpretation
120 µmol/L is below the organization’s reference range of 300–2,300 µmol/L for oleic acid C18:1 n-9 cis. That makes it a low result, but one fatty-acid measurement alone does not establish a health problem. Its meaning depends on the lab’s method, whether the sample was fasting, and the rest of your fatty-acid profile. Oleic acid is a monounsaturated fat found in olive oil, nuts, and seeds; it is also part of cell membranes. A low level can reflect lower intake or differences in absorption or metabolism. The result is not, by itself, a direct measure of energy, recovery, cognition, or cardiovascular risk.
Practical next steps
The supplied research does not establish a clinical cutoff at which an oleic acid value of 120 becomes dangerous. One observational study found associations between several fatty acids, including oleic acid, and diabetes risk; it does not show that a low oleic-acid result causes diabetes or predicts an individual’s risk.[4] Bottom line: it is below the listed reference range and worth reviewing in the context of the complete panel and your diet, but the number alone is not a reason to assume something is seriously wrong.
References
Every compared measure of this benchmark scaled to 0–100 and averaged with equal weight.
Holobiont scores 208.1 % higher in this comparison.
Dashed markers show the averages. Each connected pair is one measure. Better score.
Answers with source citations: 100 vs 0. Correct reference range stated: 70.7 vs 0.7. Review: cites sources: 68.4 vs 0. Answers the reviewer preferred: 75.3 vs 22. Review: range interpretation: 74.4 vs 30.2. Reviewer score: 67.2 vs 25. Value placed correctly: 93.8 vs 74.8. Review: actionable, safe advice: 59.2 vs 44.6.
Across 8 measures: ahead on 8, within noise on 0, behind on 0.
Every compared measure scaled to 0–100 (percentages as they are, errors as 100 minus the error, reviewer scores ×20, gram and kcal errors relative to the other arm) and averaged with equal weight. Ahead and behind count only differences whose 95 % interval excludes zero.
06 — Pricing
Move the sliders to the volume you expect. Prices are typical totals; the API charges exactly what each call uses.
Your 5 € welcome credit covers about 384 meal photos.
07 — Questions
One credit is 0.001 €. Every call is priced in credits: a meal photo analysis is about 13, a chat answer 15 to 30, a nutrition lookup 6. AI calls are charged for what they actually use plus a small platform share; refused requests cost nothing.
After setting up your company, describe in a few sentences what you plan to build (a project website is optional). We review it, usually within a working day, and email you when the credit is on your balance. If you do not want to wait, skip the step and buy credits right away.
No and no. There is one pay-as-you-go plan. Credits stay on your company’s balance until used, and you can turn on auto-recharge so work never pauses.
Recognition plus database grounding typically takes a few seconds. The response is complete JSON, so you can show the result as soon as it arrives.
The API answers 422 with a code: NOT_FOOD, MEDICATION or INVALID_IMAGE. You are not charged for refused requests, and photos are never stored.
Each analysis has an id. You can read it, list it and correct portions through the API for 180 days; after that it is deleted automatically. The photo itself is never kept.
Each recognised item is matched to a nutrition database and its nutrients per 100 g are scaled to the estimated grams. Items without a match use the model’s estimate and are labelled as such, so you can show the difference.
No. They are general wellness information grounded in your company’s biomarker library and knowledge base, with citations so users and practitioners can check the sources. Your app should say so too.
Your company’s records, keys and usage are isolated per company. Meal photos are processed and discarded; the resulting analyses are kept for 180 days and then purged nightly. See the privacy policy for details and contact us for a data processing agreement.
Sign up and you get the dashboard, API keys and the full reference with examples. The API reference is also linked in the footer.
The dashboard, the API keys and the usage view are ready in minutes; the welcome credit follows once we have approved the registration.
