Skip to content
Open source · Apache-2.0

The open computational chemistry engine AI agents call to do real discovery.

Cheminformatics, GPU simulation, and quantum chemistry through one MCP + REST surface. Any MCP-compatible AI assistant, any LLM. Install locally, self-host, or route regulatory work through the FAVES compliance API.

Install
$ git clone https://github.com/NovoMCP/novomcp \
    && cd novomcp/orchestrator \
    && pip install -r requirements.txt \
    && python main_https.py

Containers work too. docker compose up in the repo root does the same.

Scroll
The economics of deciding early

Bringing one drug to approval costs about $2.6 billion, and roughly one in twenty survives. NovoMCP moves the decision in silico, before the spend.

Apache-2.0
the
license
69
tools in
the catalog
14
work with
zero wiring
2
MCP servers
(Novo + Compute)
4
LLM providers
supported
8
FAVES
jurisdictions
Why an engine
Models are interchangeable. The engine they run on is not.

NovoMCP holds the 69-tool catalog, the NovoExpert ADMET and clinical-outcome models, the FAVES compliance layer, the immutable audit trail, and cross-run memory. The model is free to replace. The engine underneath carries the value.

The engine

One chemistry engine. One simulation engine.

Novo exposes cheminformatics + validated ADMET + compliance. Novo Compute runs the physics when a question needs it. Together they let an assistant think, test, and learn about molecules through a single MCP-compatible surface.

Novo

The chemistry engine

localhost:8018/mcp

Cheminformatics, ADMET, compliance, target evidence, and discovery orchestration. Install the open engine and reach all of it through one MCP server, from any MCP-compatible AI assistant.

  • RDKit properties, similarity, and structural alerts in-process
  • NovoExpert ADMET and clinical-outcome models
  • FAVES compliance at the point of decision
  • 11-stage autonomous discovery funnel
Novo Compute

The simulation engine

same endpoint, GPU wired

GPU and quantum chemistry for the questions that need real physics. Ships in the same open engine. The compute tools appear alongside Novo once the optional GPU services are wired.

  • AutoDock-GPU docking with strain correction
  • GROMACS molecular dynamics
  • xTB, CREST, and neural-network potentials
  • OpenFold3 structure prediction
any MCP-compatible assistant
you Profile this scaffold for ADMET and compliance, then dock the clean candidates against the target.
resolved 4 analogs · ran batch_profile across 84 properties
flagged 1 PAINS alert via check_compliance (FAVES)
docked 3 survivors with dock_with_strain · top pose −9.4 kcal/mol
every step written to the immutable audit log
AgentMode

The scientific method, run end to end.

One instruction starts an eleven-stage discovery funnel. Targets are found and validated, the literature read, actives pulled, candidates profiled, optimized, docked, gated on a calibrated clinical-clearance estimate, carried through molecular dynamics, and closed out with patient stratification against pharmacogenomic evidence.

Every stage writes to an immutable audit trail. Cross-run memory carries what prior runs learned into the next.

How AgentMode works
Governance

Every decision is on the record.

Every decision the AI makes is written to an immutable audit trail. Any funnel reconstructs completely: what was tried, what was rejected, why, and who approved it.

Observability is built into the engine from the first stage. It is the first question every enterprise buyer asks.

Immutable audit trail

Per-stage logging across the full discovery funnel. Reproducible, reconstructable, exportable.

FAVES V4 compliance

1,585 SMARTS across 8 jurisdictions. Runs inline at the point of decision, ahead of any terminal gate.

Cross-run memory

Funnel context persists across sessions. The engine remembers what prior runs learned.

Agent governance

RFC 8693 token exchange, per-agent key scoping, and agent-level audit on Enterprise.

Evidence

Numbers you can check.

Every figure here is measured and traceable to a public source: a benchmark run on a single GPU, a preprint, or the CI that runs on every fresh clone. Nothing rounded up, nothing unsourced.

214 ns/day
MD throughput
GROMACS · one A10G
27s
protein–ligand dock
AutoDock-GPU
3s
QM single-point
GFN2-xTB
362/362
FAVES V4 validation
five predetermined tests
31
ADMET models
benchmarked on public TDC
86s
NEB transition path
GFN2-xTB

Sources: speed figures measured on a single NVIDIA A10G (10 ns MD benchmark). FAVES validation: FAVES V4 preprint (chemRxiv). ADMET benchmarked against the Therapeutics Data Commons . NovoMCP’s own models on a public suite, no leaderboard claim.

Open source · Apache-2.0

The next discovery runs on an engine you can inspect.

Install the engine on your own hardware, or route regulated compliance work through the hosted FAVES API. Same engine underneath.