Skip to content

Use cases

New to agent-callable tools? This page answers the practical question — what would I actually do with this? — with concrete workflows. If you know cheminformatics but have never wired an LLM to a set of tools before, start here.

The shift: from library to callable engine

If you do computational chemistry, you already have RDKit, maybe a docking rig, maybe an ADMET model. You call them by writing scripts. NovoMCP doesn't replace any of that — it exposes those capabilities as tools an AI assistant can call for you, over one protocol (MCP) plus a REST API. What changes is who writes the glue:

  • Library workflow — you write a script that parses SMILES, calls RDKit, filters, calls the next thing, handles the errors, formats the output.
  • Engine workflow — you tell your assistant "profile these 12 compounds, drop anything that fails Lipinski, and show me the survivors sorted by QED" and it calls the tools in order, handling the plumbing.

Both are valid. You can call tools directly (curl, REST, or from a notebook) when you want determinism and reproducibility; the agent path is for exploration and for anyone who'd rather describe the goal than hand-write the pipeline. Same engine underneath either way.

Workflow 1 — Profile a molecule you're reading about

You hit a compound in a paper or a database and want its druglikeness picture without opening a notebook. Ask your assistant (or curl the engine directly):

"What's the molecular profile of aspirin?"

The get_molecule_profile tool returns molecular weight, LogP, TPSA, QED, hydrogen-bond donors/acceptors, rotatable bonds, aromatic rings, and Lipinski pass/fail — computed on the fly via RDKit, no data download required. This is one of the 11 tools that work fully local out of the box.

Workflow 2 — Triage a shortlist

You have a handful of candidate structures and want to narrow them before committing compute. In one request:

"Here are 15 SMILES. Calculate properties for all of them, flag any Lipinski violations, and rank the rest by QED."

The assistant fans this out across calculate_properties / get_molecule_profile, applies your filter, and hands back a ranked table. When you've wired the ADMET service (addie-models), the same shortlist can carry predicted solubility, permeability, and toxicity flags alongside the physicochemical properties.

Workflow 3 — Find neighbors of a hit

You have one active compound and want structurally similar molecules to explore around it:

"Find molecules similar to this scaffold and show me the closest 20."

search_similar returns near neighbors you can then profile or triage with the workflows above — the start of a lightweight hit-expansion loop, all from natural language.

Workflow 4 — Run the discovery funnel

For end-to-end target-to-candidate exploration, the engine ships a governed 11-stage discovery funnel — target discovery → validation → literature → known actives → ADMET → compliance → lead optimization → docking → clinical outcomes gate → MD → patient stratification. Trigger it from any MCP client with "Novo AG" or agm; the engine returns the staged protocol for the assistant to execute, pausing for your input at each gate (human-in-the-loop, not a black box). The heavier stages (docking, MD, structure prediction) run once you've deployed the corresponding compute services.

Workflow 5 — Call tools directly from a script

Not everything should go through an assistant. When you want a deterministic, scriptable call — in a notebook, a CI job, or a pipeline — hit the REST surface:

curl -s -X POST http://localhost:8018/mcp/tools/get_molecule_profile \
  -H 'Authorization: Bearer x' \
  -H 'Content-Type: application/json' \
  -d '{"arguments": {"smiles": "CC(=O)Oc1ccccc1C(=O)O"}}' \
  | python3 -m json.tool

Same tool, same result the agent would get — just called by you, reproducibly. The full catalog is in the API reference and the OpenAPI spec at /v1/openapi.json.

Who this is for

  • Bench and computational chemists who want a druglikeness/ADMET read without standing up a scripting environment for every question.
  • Teams building AI assistants for drug discovery who need a ready-made, standards-compliant (MCP) tool surface instead of wrapping a dozen libraries themselves.
  • Anyone with an MCP client (Claude Desktop, Cursor, Zed, Cline) who wants molecular intelligence available in the same place they already work.

If none of your work touches molecules, this isn't for you — and that's fine. But if it does, the value is not having to rebuild the plumbing between "I have a structure" and "I have an answer."

Next