How it works
What happens after you press enter.
You do not have to know which model to use, how to install it, or what hardware it wants. You describe the problem. Everything below is what Corollary does with it — and how you check the answer afterwards.
End to end
Three steps, and you only do the first one.
Everything between the question and the answer is work you would otherwise be doing by hand.
You describe the problem
A sequence, a compound, a target, a question. No model names to memorise and no parameters to guess — say what you want to know the way you would say it to a colleague.
It picks the right tools
One model rarely answers a real question. Corollary works out which models to run, in what order, and what to hand each one — then shows you that plan before it spends anything.
You get a result you can check
Structures you can rotate, numbers you can trace to the model that produced them, papers linked to their source. When a tool fails, it says so instead of filling the gap.
Choosing the model
Describe the science, not the checkpoint.
Which model answers a question best is a real scientific judgement, and it changes as new ones land. Corollary makes that judgement in one place, against criteria you can read — so the reasoning behind a choice is something you can argue with rather than something you have to infer.
- What you gave it — a bare sequence or an alignment, a molecule as text or as a structure, one chain or several
- What you asked for — fold, design, binding, variant effect
- Whether you want the most accurate answer or the fastest one
- Whether the result has to be usable in commercial work
- How the model scores on published benchmarks for that exact task
- What is free to run right now, and what a paid run would cost
You can always overrule it. Pick a model yourself
You asked
“How tightly does this compound bind my protein?”
Best benchmark score for co-folding, and licensed for commercial work
The plan
- Fold the protein from its sequence (ESMFold-2)
- Co-fold it with the compound and score the pose (Boltz-2)
- Estimated cost
- $0.20
- Estimated time
- 49 s
Fallback: DiffDock, if the first pass looks unstable. Nothing runs until you say so, and you see the estimate first.
What it covers
Nine fields, one way in.
Each field brings its own models, its own databases and its own conventions. You reach all of them the same way, and a question that crosses two of them does not become your problem.
Protein Structure
14Fold a sequence, retrieve a predicted structure, or model a complex — from single-sequence folding to MSA-accelerated ensembles.
Protein Design
8Generate novel backbones, inverse-fold sequences onto them, and evolve variants toward a target property.
Docking & Binding
7Pose small molecules in a pocket, score protein–protein interfaces, and estimate binding free energies.
Antibody & Biologics
8Number, model, design and de-risk antibodies, nanobodies and TCRs — CDR design through developability.
Genomics & Single-Cell
8Genome-scale language models, regulatory element prediction, variant effect scoring, and single-cell atlases.
Chemistry & ADMET
20Molecular representation, property and toxicity prediction, generative chemistry, and retrosynthesis.
Multi-Modal Foundation
87Models that speak protein, molecule, genome and natural language in one representation space.
Simulation
4Molecular dynamics, end-point free energy, and conformational sampling for when prediction needs physics.
Literature
35Grounded retrieval and synthesis over the biomedical corpus, with every claim tied to a citation.
Time and cost
You see the price before you spend it.
Jobs range from instant and free to an overnight simulation. Every one is classified before it starts, and the estimate is shown while you can still change your mind.
Micro
Takes
< 1s
Typical cost
Free
For example
Fast
Takes
1–30s
Typical cost
$0.01–0.05
For example
Standard
Takes
30s–5min
Typical cost
$0.05–0.50
For example
Heavy
Takes
5–30min
Typical cost
$0.50–5.00
For example
Simulation
Takes
30min–24h
Typical cost
$5–100
For example
Every job carries a spending ceiling. One that reaches it stops and keeps whatever it produced.
Identical runs are cached. Asking the same question of the same input twice costs nothing the second time.
You are billed for the seconds a job runs and nothing while it sits idle.
Work you can defend
Two things we will not do.
Both of these are the difference between a tool you can cite and a tool you can only demo.
Guess at a number.
A confident answer with nothing behind it is not a rough edge in science — it is disqualifying. Every factual claim has to point at a database record, a model output with its confidence, or a paper that was actually retrieved. When a tool fails, you are told it failed.
It is also why Corollary declines to eyeball a partition coefficient from a structure it parsed itself. A wrong number that looks right is worse than no number.
Hand you a result you can't use.
Not every open model may be used in commercial work. Some are research-only, some are permissive, and the difference matters the moment a result goes into a filing. Corollary checks the licence before it runs anything and routes commercial work to models you are allowed to ship with.
- Cleared for commercial use
- 51/191
- In the library
- 542
Bring a question you already know the answer to.
It is the fastest way to find out whether this is any good.