Product · CogniChem Assistant

Describe the problem. Approve the plan.

The Assistant turns a chemistry question into a ready-to-run job or workflow, priced for your plan. It proposes; you decide. Nothing runs until you click Approve.

Describe your problem Or browse workflows

Describe the problem
In plain language, with your files attached if you have them: a target, a compound list, a question about a series.
Review the plan and its cost
The Assistant picks a ready-made workflow or a single tool, or chains the steps itself when no template fits. It fills in the inputs, checks them, and shows a card with the API's estimate for your plan.
Approve, and follow the run
Your click starts it. The Assistant can then read the run's status and results back to you in the same conversation.

Examples

Ask the way you would ask a colleague.

Pick a ready-made workflow, explain one, run a single tool, or build a workflow when none fits. Each answer cites the papers behind the methods and ends in a plan card you approve or dismiss.

[example conversations]Fig. 01

I've attached a SMILES library and a PDB of my target. Which compounds are most likely to bind?

“Screen a compound library against a target” fits. It converts your SMILES to 3D, finds the binding pocket with the Binding-Site Detector, and docks every compound with AutoDock Vina on CPUs, about $0.13 a typical run. Your two files are already filled in.

Trott & Olson (2009), Journal of Computational Chemistry
Krivák & Hoksza (2018), Journal of Cheminformatics

Proposed workflow · waiting for you

Screen a compound library against a target

SMILES → 3D embed → dock

Steps: 3
Runs on: CPU
Estimate: Priced for your plan

Why does this workflow dock everything with Vina, and only then run Boltz-2?

Vina is fast and runs on CPUs (about $0.13 a typical run), so it can screen the whole library. Boltz-2 predicts the full protein–ligand complex and a binding affinity, which tells you more but needs a GPU (about $1.93 a run on an Nvidia A10). So the workflow spends Boltz-2 only on the compounds Vina ranks highest.

Trott & Olson (2009), Journal of Computational Chemistry
Passaro et al. (2025), openRxiv

Proposed workflow · waiting for you

Re-predict docking hits with Boltz-2 affinities

Vina triage → Boltz-2 re-predict (batch)

Steps: 5
Runs on: CPU · A10 GPU
Estimate: Priced for your plan

Predict solubility and hERG risk for the compounds in this file.

That takes one job, not a workflow. ADMET Predictor scores solubility, hERG, and the other ADMET endpoints from SMILES with pretrained Chemprop models. It runs on CPU, about $0.042 a typical run.

Swanson et al. (2024), Bioinformatics

Proposed job · waiting for you

ADMET Predictor

Steps: 1
Runs on: CPU
Estimate: Priced for your plan

Dock my library against this target, but only the compounds with low predicted hERG risk.

No ready-made workflow filters before docking, so I chained one. ADMET Predictor keeps the compounds whose predicted hERG risk is below 0.3; they are embedded in 3D and docked with AutoDock Vina into the pocket the Binding-Site Detector finds. Every connection checks out, and you can open it in the builder to change any step.

Swanson et al. (2024), Bioinformatics
Trott & Olson (2009), Journal of Computational Chemistry

Proposed workflow · built for this request

Dock the compounds that pass an hERG filter

ADMET filter → 3D embed → dock

Steps: 4
Runs on: CPU
Estimate: Priced for your plan

Illustrations of exchanges. Prices, papers, and workflows are the catalog's. In the app, the plan card shows the estimate the API computed for your plan, and approving re-checks it.

What it can do

Read, look up, estimate, propose.

These are the Assistant's tools, grouped by what they are for. Everything outside this list, including starting a run, is closed to it.

Find the right workflow or tool
Searches the CogniChem product graph, reads job schemas and workflow ports, and checks a workflow before proposing it.
kg_product_searchkg_product_getkg_product_neighborhoodkg_product_how_to_runlist_job_typesget_job_schemalist_workflow_templatesget_workflow_node_portsvalidate_workflow
Look up compounds and targets
Live lookups in PubChem, ChEMBL, Open Targets, UniProt, and the RCSB PDB. Every result carries a link to its source record.
resolve_compoundslookup_chembllookup_targetslookup_uniprotlookup_pdblookup_properties
Read the method papers
Searches the reference graph of the papers behind each method, so answers point to the source.
kg_reference_searchkg_reference_getkg_reference_neighborhood
Price it before it runs
Estimates jobs and workflows at your plan's rates, checks your wallet, and quotes its own next turn.
estimate_jobestimate_workflowget_wallet_balanceestimate_llm_turnget_session_spend
Propose runs for you to approve
Turns the plan into a job or workflow card with its estimate. Only your Approve click starts it.
propose_job_runpropose_workflow_run
Follow the results
Checks the status of runs started from the conversation and reads the files they produce.
get_job_statusget_workflow_runlist_session_artifactsget_artifact_record

Safety model

You stay in control of every run and every dollar.

The rules below are enforced by the API, not by the Assistant's instructions.

It cannot start anything itself

The Assistant's tools can search, look up, estimate, and propose. Submitting, cancelling, resuming, approving, and deleting are not available to it at all; the only path to a run is your Approve click in the app.

What you approve is what runs

A proposal freezes the exact job or workflow, its inputs, and its estimate. To change anything, the Assistant makes a new proposal, or you open the workflow in the builder and launch it yourself.

Re-priced when you approve

Approving re-checks the inputs and the estimate at your current plan. If the price went up since the card was shown, nothing runs until you approve the new amount; if it went down, it runs at the lower price.

Plans stay current

A plan card stays open for 30 minutes, so an old plan can never be approved by accident later. Need it after that? Ask again and the Assistant makes a fresh card, priced at today's rates.

Every conversation has a spend cap

Assistant tokens in one conversation stop at $1.00 by default. Only your Continue click raises the cap, by $1.00 at a time. Approving a run costs no tokens; compute is billed like any other run.

Your files stay files

Attached PDB, SDF, and SMILES files are stored in your library and passed to jobs by reference. Their contents never enter the conversation, and uploads are never added to any shared knowledge graph.

Pricing

Token rates by plan.

Low and High reasoning each have their own rates. Higher plans pay less per token.

Low reasoning: USD per million tokens
Plan	Input	Cached input	Output
Basic	$0.70	$0.07	$4.20
Starter	$0.50	$0.05	$3.00
Pro	$0.40	$0.04	$2.40
Enterprise	$0.30	$0.03	$1.80

High reasoning: USD per million tokens
Plan	Input	Cached input	Output
Basic	$7.00	$0.70	$35.00
Starter	$5.00	$0.50	$25.00
Pro	$4.00	$0.40	$20.00
Enterprise	$3.00	$0.30	$15.00

Compare plans

FAQ

Questions about the Assistant

Can I use the Assistant without it running anything?

Yes. Ask questions, compare methods, or have it price a plan. A run only starts when you click Approve on a card.

What if no ready-made workflow fits?

The Assistant builds one. It chains catalog tools by their declared inputs and outputs, checks every connection, and prices the result. Workflow cards have an Open in builder button, so you can change the plan yourself before you launch it.

What does it cost?

Assistant turns are billed per token at your plan's rate (below), within the $1.00 conversation cap. Jobs and workflows you approve are billed per second of compute, exactly as if you had launched them yourself.

Can I use my own AI client instead?

Yes. Point Claude Code, Cursor, or any MCP client at the hosted CogniChem MCP server (https://mcp.cognichem.com/mcp) with your API key to get the same jobs, workflows, and lookups. You pay your own model costs; we bill only the compute.

Where do its answers come from?

From the CogniChem product graph (what each tool does and how to run it), a graph of the method papers behind the tools, and live lookups in public databases. Lookup results link back to their source records.