Use the Assistant

The CogniChem Assistant is a chat in the app. Describe what you want to find out, for example "dock these five SMILES against COX-2 and rank them", and it works out which tools or workflow to use, looks up the data it needs, and proposes runs with their prices.

What it can do

• Find the right workflow or tool from the CogniChem catalog, and check a workflow before proposing it.
• Look up compounds and targets live in PubChem, ChEMBL, Open Targets, UniProt, and the RCSB PDB. Each result links to its source record.
• Read the method papers behind each tool.
• Answer quick calculations in the reply: identifiers and format conversions, standardization, drug-likeness rules and alerts, RMSD between two structures, fluid properties at one state, and predictions with public or your own MPNN models. These count against your plan's monthly utility and inference allowance, not your wallet, and the reply says which values were computed or predicted.
• Price runs at your plan's rates and check your wallet.
• Propose runs as plan cards for you to approve.
• Follow results of runs started from the conversation and read the files they produce.

It cannot start a job or workflow, or cancel or delete anything, on its own. Those tools are not available to it. Quick calculations are capped per message and per conversation, and every structure is screened first, like any run.

Plan cards: nothing runs until you approve

When the Assistant proposes a run, the conversation shows a plan card: the job or workflow, its inputs, and the server's estimate. Then:

• Approve starts the run. CogniChem checks the plan again and re-prices it at your current rates first. If the price went up, nothing starts and the card shows the new price to approve. If your wallet can't cover it, the card stays pending until you add funds.
• Dismiss discards it.
• Open in builder (workflow cards) opens the plan in the workflow builder as a draft you can edit and run yourself.

A card expires 30 minutes after it is proposed. Approved runs show their status on the card and appear in your job queue or workflow runs like any other. When a run finishes, Summarize result on its card shows the main numbers and top results without starting a message. Suggest next steps then asks the Assistant to propose a follow-up run, which you approve like any other.

Export notebook in the thread header downloads a Jupyter notebook that replays the thread's lookups and approved runs with the Python SDK (cognichem-client). It reads your API key from COGNICHEM_API_KEY and holds no credentials; running a submit cell starts a new, billed run.

Files

Attach protein structures (PDB), molecules (SDF, MOL), SMILES lists, or CSV tables of compounds and activities to a message. They are stored as artifacts in your library, and the Assistant passes them to tools by reference. File contents are not added to the conversation or shared with anyone.

To give the Assistant a structure without typing SMILES, click Draw and sketch it, or click Edit on a molecule in the viewer. The drawing goes into your message as SMILES. A SMILES list or CSV table opens as a grid you can sort and filter; select rows and click Ask for a plan to have the Assistant propose a scaffold, matched-pair, Free–Wilson, clustering, drug-likeness, or ADMET run on them. Protein models from AlphaFold DB have a Sequence tab with UniProt features: select residues to show them in 3D, or ask for a fold, an alanine scan, or Evo 2 scoring.

What it costs

Each turn is charged for the language-model tokens it uses, at your plan's rates; runs you approve are charged like any other job or workflow. Before a turn starts, CogniChem holds enough to cover it; afterwards you pay for the tokens used and the rest is released. Stopping a turn charges only for what was used so far.

Each conversation has a spend cap of $1.00. When the next turn would pass it, the Assistant stops and asks whether to continue; continuing raises that conversation's cap by $1.00.

The High reasoning level uses a stronger model for harder planning, at higher token rates.

Your data

Lookups send only names and IDs to the outside databases. Similarity and substructure search in ChEMBL needs a structure (a SMILES string) to leave CogniChem, so it is off until you turn on structure search for that conversation. The Assistant cannot turn it on itself.

The Assistant covers the details: limits, the turn time limit, and how plan cards are checked.