Loading the front page.
Membership · Practitioner · API and export
Catalyst is building the computational infrastructure for human biology: a provenance-aware knowledge graph that makes biomedical literature computable. The live product is its evidence graph and research interface; the next layers will assemble mechanisms and, over time, support dynamic physiological models.
The substrate, with its evidence attached
Not a citation pointing at a paper that may or may not say it. The verbatim span the claim was read from, stored beside it and returned with it.
Computed from study design, sample size, population and replication by a published function. No author, customer or model writes it.
A connection assembled through shared biology is labelled a hypothesis in the data itself. A weak claim cannot be drawn to look like a strong one.
One finding
Zn(2+) decreases Diarrhoea duration
“Twenty-four trials reported on the duration of diarrhoea, and the associated meta-analysis suggested that the duration of diarrhoea decreased significantly in the zinc group as compared to no zinc/placebo group (MD = −13.27 hours; 95% CI = −17.66, −8.89; 24 trials, 6249 participants, moderate certainty of evidence).”
A claim enters the graph only when a verbatim sentence supports it on its own, two independent extractions agree on it, and, where the evidence is limited or stronger, a refuter that has not seen them fails to break it. 16,970 papers read so far have produced 2,410 findings from 1,503 sources. That ratio is the product: a rule that costs findings rather than one that produces them.
Every rule, and the evidence strengths it has producedFrom the literature to a row, to scale
Enrolled from the literature against the questions the graph is asked.
Read by the extraction pipeline, claim by claim.
Behind the 2,410 findings that survived verification.
The same rows, evidence strengths and quotes whoever is asking. What changes is the question each door takes, and the shape the answer comes back in.
For people who want a straight answer about what they take, without a sales pitch or a scare.
What to stop, not only what to add
Zinc and copper, calcium and iron, a herb that shares a metabolising enzyme with common medicines. A conflict is shown with its citation only where a source documents it, and this door opens once documented interactions are curated. Until then, a connection like these is a labelled hypothesis.
No effect found is not the same as never studied
A trial that found no effect is a finding, and is shown as one. A question no study has asked is shown as unstudied, not as a negative.
What changed this month
Follow what you take and be told when a finding is replicated, contradicted or retracted, from what to what, and why.
Five ways in, one engine
“I am low in iron.”
An analyte and a direction: serum ferritin, decreased. What closes the gap, what made it, what blocks it at the gut, each with its own evidence strength.
“I am tired all the time.”
A phenotype. What the literature associates with it, plural and unranked, and the best-evidenced answer first even when it is not a compound.
“Here is a photograph of the label.”
A set of compounds with quantities. Which can be traced, which cannot be identified from a label, and what a proprietary blend cannot be dosed into.
“What does horny goat weed do?”
One compound. Every finding on it, the tissues it reaches, the proteins it touches, and what it depletes or blocks.
“Here are the eleven things I take.”
A selection. What competes for absorption, what should be hours apart, what shares a pathway, each as a cited interaction or a labelled hypothesis.
The engine is live at /explore today. The doors open in the order the graph can answer them.
For the dietitian, pharmacist, pediatrician or trainer who spends the appointment talking a client down from a headline. What that takes is the finding, the study, the population and the evidence strength, on one page.
The population is on the row
Who was studied, on what dose, for how long, and a finding demoted when that population does not match your client.
The disagreement is kept
A contradiction is shown beside the finding it contradicts, never averaged into one number.
The absence is inspectable
What was read and rejected for a compound, with the reason, so a headline can be traced to where it fell.
Three findings, as the tier prints them
Strong evidence
Zn(2+) decreases Diarrhoea duration
“Twenty-four trials reported on the duration of diarrhoea, and the associated meta-analysis suggested that the duration of diarrhoea decreased significantly in the zinc group as compared to no zinc/placebo group (MD = −13.27 hours; 95% CI = −17.66, −8.89; 24 trials, 6249 participants, moderate certainty of evidence).”
PMID:39641338n = 6,249replicated 5 times
Strong evidence
all-trans-retinol decreases Hospital length of stay
“Kiatchoosakun demonstrated a significant reduction in duration of hospital stay in the vitamin A treatment groups [mean difference (MD), −49.9; 95% CI −88.78 to −11.02; 20 infants].”
PMID:30496228n = 20replicated 3 times
Strong evidence
melatonin has no detected effect on All-cause mortality
“We found no significant difference in mortality between patients receiving melatonin and those receiving placebo (RR, 0.90; 95% CI: 0.77–1.06; I2 = 0%).”
PMID:36316858n = 2,350replicated 2 timescontradicted 2 times
The graph’s three strongest rows when this page was built. Each is a page a client can be sent to, with the same evidence strength and quote.
An API, an MCP server and a bulk export, for a product or a model that has to say something about a substance a person consumes. The public endpoint answers today with no key.
Five fields that make the data defensible
source.quote, evidence_strength, assertion_class, withheld, and replication and contradiction counts, on every row.
An MCP server
The graph as tools an assistant can call, so a model answers from assessed findings with citations instead of a confident guess.
Licences that travel
Database, record, release and licence on every reference row, and a bulk export whose manifest names what is missing.
Request
GET /api/v1/relations?node=CHEBI:16919&class=bothResponse
{ "node": { "id": "CHEBI:16919", "label": "creatine", "type": "compound", "href": "/compounds/creatine" }, "class": "both", "tier": "core", "findings": [ { "id": "01M102QACN9CTAD98A1CX76S62", "assertion_class": "documented", "predicate": "affects", "direction": "increases", "evidence_strength": "moderate", "subject": { "id": "CHEBI:16919", "label": "creatine", "type": "compound", "href": "/compounds/creatine" }, "object": { "id": "CAT:outcome/serum-creatinine", "label": "Serum creatinine concentration", "type": "outcome", "href": "/outcomes/serum-creatinine" }, "sentence": "creatine increases Serum creatinine concentration", "evidence_type": "crossover", "n": null, "population": "children aged 5-18 years with facioscapulohumeral muscular dystrophy; safety outcome, not the primary endpoint", "dose_regimen": "creatine monohydrate ~100 mg/kg/day (2.5-10 g daily by weight band) for 12 weeks per crossover period", "duration_days": 84, "effect_size": 12.63, "effect_metric": "mean_difference", "ci_low": 1.14, "ci_high": 24.12, "replication_count": 3, "contradiction_count": 0, "curated_at": "2026-08-26T22:23:18.405823+00:00", "permalink": "https://catalystproject.ai/e/01M102QACN9CTAD98A1CX76S62", "source": { "id": "PMID:40366059", "title": "Effect of creatine monohydrate on motor function in children with facioscapulohumeral muscular dystrophy: A multicenter, randomized, double-blind placebo-controlled crossover trial.", "url": "https://doi.org/10.1002/phar.70025", "year": 2025, "retracted": false, "quote": "Specifically, there was an increase in serum creatinine (mean difference 12.63 μmol/L, 95% CI 1.14 to 24.12) and decrease in eGFR (−25.79 mL/min/1.73 m2, 95% CI −44.3 to −7.25) in individuals on creatine supplementation compared to placebo." } } ], "relations": [ { "id": "01M0DBCYMYNNH2W88V0PB5XNSE", "assertion_class": "inferred", "predicate": "transported_by", "direction": "modulates", "subject": { "id": "CHEBI:16919", "label": "creatine", "type": "compound", "href": null }, "object": { "id": "UNIPROT:Q6ZSM3", "label": "SLC16A12", "type": "protein", "href": "/proteins/slc16a12" }, "sentence": "creatine is transported by SLC16A12", "source": { "database": "rhea", "name": "Rhea", "record": "RHEA:73043", "url": "https://www.rhea-db.org/rhea/73043", "release": "2026-08-19", "license": "CC BY 4.0" } } ], "page": { "limit": 50, "offset": 0, "total_relations": 34, "total_findings": 36 }, "withheld": { "relations": 0, "sources": [] }}Trimmed to 1 of 36 findings, 1 of 34 relations.
Read from the graph when this page was built, through the endpoint’s own mappers. Trimmed to one row of each kind; the real response carries 36 findings and 34 reference statements. The full reference.
A plan buys a door, throughput and delivery. It never buys different evidence or a different evidence strength.
A person deciding what to put in their body, and what to stop putting in it.
$9per month, or $79 a year
Early access, by email.
A dietitian, pharmacist, pediatrician or trainer who has to answer them, citably.
$29per seat, per month
Early access, by email.
Every reader, every crawler, every agent. No key, no account, no card.
Freeforever
One product, one team, one key. Early access, priced for what is built today.
$19per month, early access
A product with traffic, or a group that needs the graph in more than one system.
$399per month
Early access, by email.
The whole graph, inside your own infrastructure, under terms that permit redistribution.
Custompriced per engagement
US dollars, excluding tax. API plans are metered by call and counted on a 200; a 429 or a 503 is ours and is never billed.
More evidence at a higher price
Every plan serves the same rows at the same evidence strength.
An evidence strength you can influence
Evidence strength is computed by a published function. Nobody sets it.
Rows without provenance
There is no cheaper feed with the quotes stripped out.
Coverage promises
Nothing is called covered before somebody has read it.
A thing to swallow
We sell no substance. Referrals, when they open, never move an evidence strength.
Anything else: hello@catalystproject.ai
No. A membership is a list of the things you follow, held as node ids, and the context the evidence is matched against. Health profiles, symptom logs and anything a person reports about their own body are a class of data this platform does not collect, and introducing one would be a written decision rather than a feature.
No. The platform sells nothing you swallow and will never ship a product. What is decided and not yet open is a referral: a compound page may link to one lot-tested product from a named third-party seller (USP, NSF or a published certificate of analysis, and absent from the FDA tainted-products list), under a referral fee stated on the page. It sits below the evidence on that compound’s page only, never on a symptom, a search or an answer. It is never a stack or a blend. It shows the evidence strength the graph computed, very limited included, and no listing is gated or ranked on evidence strength.
No, and every answer says so. The graph holds what has been read, the response reports exact totals, and a node with nothing curated returns nothing curated instead of something adjacent.
Evidence strength is computed, so a wrong evidence strength means a wrong field, and a wrong field is a row we fix and re-derive rather than a position we defend. The row keeps its id so a product can reconcile it, and a retracted paper flags its rows rather than deleting them.
You can. Every entity is at a stable URL and every finding is server-rendered beside its quote. What you will not get is the assertion-class column, the withheld block, exact totals, or a licence position you can show a reviewer.
Within the terms each row carries, and each row carries them. The public tier is public-domain and attribution sources only. Redistribution beyond that is what the corpus licence is for, negotiated per upstream source.
Today the public endpoint counts requests per address for sixty seconds, in memory, to enforce its rate limit, and keeps nothing else about a query; our host logs the request like any web server does. Paid keys will be counted per call so a plan can be billed. Queries are not sold, shared or used to rank curation, and no plan will collect personal data beyond the billing relationship.
The graph covers compounds, proteins, tissues, pathways and measured outcomes: the shared biology under nutrition, supplementation and every substance sold as food. If your product has to say something defensible about a substance a person consumes, this is the layer under that sentence.
The public endpoint answers today with no key. Tell us which door you need and you will hear when it opens.
API referenceThe grading functionBulk exportSources and licences