Live · 2,022,437 indexed papers · updated daily

Your AI will invent
a citation. Ours won't.

A Model Context Protocol server for dietitians, nutritionists and health coaches. It connects Claude or ChatGPT to real regulator positions, real reference intakes and two million real papers — and it says found: false when it doesn't know, instead of filling the gap.

Connect it free Watch it refuse

Free while the register grows. No card, no metering.

nutrition-evidence · streamable-http 10 tools connected
3 tool calls 0 invented values 1 honest refusal every field carries its register + version
The failure mode

A wrong PMID costs more than a blank one.

Ask a general model for the evidence on an additive and it will answer fluently. It will also mix up which body said what — attributing an IARC hazard class to the FDA, quoting a reference intake for a compound that has none, or citing a PMID that resolves to an unrelated paper. None of that reads as an error. It reads as an answer.

For a practitioner, that's the expensive kind of wrong: you repeat it to a client, or you put it in a handout with your name on it. This server is built on one rule — never fill a gap. Every tool returns a row from a versioned register, or it returns nothing and says which register it looked in.

2,022,437papers indexed, full-text searchable
~500regulator positions, each with a source URL
247additives with a written evidence read
10tools, zero stubbed responses
Query “ultra-processed” + cardiovascular matches 293 of 2,022,437 the dev slice returns 2 — which is why the corpus is gated
The toolset

Ten tools. Each one names its source.

Your assistant reads each tool's description and decides when to call it. These are written so it reaches for the register instead of its own memory — and so the answer arrives with the register's name and version attached, ready to quote.

ingredient_dossierstart here

What the FDA, EFSA, JECFA, IARC, NTP and California OEHHA have each actually said about an additive — with the year and a link — plus a written read of the human evidence. Separates a retailer banned it from a study found something, because they are not the same claim.

247 substances · 247/247 written reads · 218 with a source URL
nutrient_reference

Reference intake, function, food sources and deficiency signs. Returns has_intake_reference: false for the 125 compounds that have no RDA, AI or UL at all — the ones supplement labels invent numbers for.

169 nutrients · 44 with an intake reference
find_studies

Full-text search across the corpus, returning real PMIDs with journal, year, DOI and PMCID. Filter by study design or year. Tells your assistant to cite these and nothing else.

2,022,437 papers · GIN full-text index
disease_causes

Attributable fractions with their confidence interval and the cohort they came from — so a population number never gets quoted as one person's risk.

19 diet-related conditions
claim_evidence · list_claims

Graded verdicts on diet–health claims, each with a field stating what the evidence does not establish. Returns unclassified rather than guessing a grade.

123 graded claims · 13 with verified citations — see below
evidence_card

Builds a client handout or a post graphic from the register — every cell carrying its grade, its gap and its PMIDs, your name in the footer, and an honest count of what isn't cited yet.

self-contained HTML · print or screenshot

Also: upf_availability, diet_patterns, registers.

What it won't do

The gaps, printed on the box.

A tool that claims total coverage is the thing this one is a corrective to. So here is every register and its real state — the same numbers a gate check prints before any of it ships.

RegisterCoverageState
ingredient additives247 / 247 Complete. Every substance has a written read; 218 carry a regulator URL.
nutrient references44 / 169 By design — the other 125 compounds genuinely have no reference intake, and the tool says so instead of inventing one.
graded claims123 / 123 Complete. Every verdict carries machine-verified citations: established claims their landmark evidence, contested claims both sides of the split, and the lower grades papers at their own tier — an emerging verdict cites the hedged review, not a triumphant one.
retail UPF availability3 / 75 Mostly a map of what nobody has measured. Not a feature — included so you know the shape of the hole.
disease attribution19 / 19 Each with interval and source cohort named.
effect estimatesnone The corpus holds titles, abstracts and MeSH terms — no extracted hazard ratios. find_studies states this in its own response so nothing downstream quotes a number that was never there.
Why publish this

Because you're going to check. The first practitioner who opens a card, finds an uncited grade and posts about it does more damage than the gap itself — so the gap ships labelled. Lead with the additive dossiers: that register is finished.

Get access

Free, while the register grows.

This is built and running, by one person, and the register is young — 247 ingredients fully sourced, the graded-claims citations still being verified one paper at a time. Charging for it before that work is further along would be selling the part that isn't finished. So it's free: use it, lean on it, and tell me where it's wrong.

Early access

Everything, free

$0no card, no metering
Early users keep preferential terms if a paid tier ever lands.
  • All ten tools, connected to Claude, ChatGPT or any MCP client
  • Unlimited lookups — no per-query metering
  • Client handouts and post graphics with your name and credential in the footer
  • Say which register gets built next, and it goes in the queue
  • Every new register as it lands
Get connected Reply comes with the server URL and a two-minute setup, usually the same day.
Being straight about it

Free is not a trick and not forever-guaranteed. If this earns a paid tier one day, it will be because the claims register is fully cited and worth paying for — and the people who used it while it was young will be treated like it. What you'll never see is a bill you didn't agree to.