Skip to content

NovoMCP

The open computational chemistry engine for drug discovery and materials science.

NovoMCP is an agent-callable engine that exposes molecular intelligence, cheminformatics, ADMET prediction, molecular docking, molecular dynamics, quantum-mechanical calculations, protein structure prediction, and a governed discovery funnel through the Model Context Protocol and a curated REST API.

One engine. Two domains. Four surfaces.

Why NovoMCP?

Molecular intelligence is scattered — RDKit in one notebook, an ADMET model in another, a docking rig on a cluster, a compliance check in a spreadsheet, literature search in a browser tab. Wiring all of it into something an AI agent can actually call is the work nobody wants to redo per project. NovoMCP collapses it into one engine:

  • Agent-callable by default. 69 tools over MCP JSON-RPC plus a curated REST API — the same catalog whether you drive it from Claude Desktop, Cursor, a script, or the discovery funnel.
  • Runs on your laptop, day one. 14 tools work with no API keys and nothing to provision. (Four of the 14 default tools query public APIs — ChEMBL, ClinicalTrials.gov, bioRxiv, PubMed literature; the rest are fully offline.) The rest unlock as you deploy compute or bring data — nothing is gated behind a signup to get started.
  • An engine, not a wrapper. Intent recognition, orchestration planning, semantic tool search, and a governed 11-stage discovery funnel are built in — not left as an exercise for the caller.
  • Pluggable where it counts. Auth, metering, audit, and LLM provider each swap via env vars, so the same core runs unauthenticated on a laptop or metered-and-audited in production.

New here? Boot it in 2 minutes, then see how it fits together.

Get started

What it does

  • 14 tools work fully local out of a 69-tool catalog. The rest unlock as you deploy compute services, provide your own data, or subscribe to hosted APIs. See Tool availability for the full map.
  • Full 69-tool catalog spans cheminformatics (properties, similarity, filtering), ADMET prediction, docking, molecular dynamics, protein structure prediction, quantum-mechanical calculations, literature/patent search, regulatory compliance
  • REST API with a curated OpenAPI 3.1 spec at /v1/openapi.json
  • MCP JSON-RPC at /mcp/ — usable from any MCP-compatible client (Claude Desktop, Cursor, Codex, Zed, Cline, and others)
  • Autonomous discovery funnel — trigger with "Novo AG" or "agm" from any MCP client, returns an 11-stage protocol for the LLM to execute
  • Pluggable everything — auth (AuthGate), metering (UsageMeter), audit (AuditSink), LLM providers (OpenAI, Anthropic, Ollama, Azure) all swap via env vars

What lives elsewhere

By design, the following ship separately:

  • Curated 122M-molecule enriched corpus — publishing to AWS Open Data / Kaggle / Zenodo as a downloadable dataset. Cached lookups happen through the molecule-index service when configured.

None of these are required for the engine to run.

Licensing

  • Top-level tree (surfaces, wrappers, connectors, protocol docs): Apache-2.0 (LICENSE)
  • Orchestration core (orchestrator/mcp/): Business Source License 1.1 with a change date of 2029-07-12 to Apache-2.0 (LICENSE.core). Non-production use, self-hosting, and modification are permitted immediately.
  • Pre-trained model weights in companion repositories: MIT.

Contributing

  • Issues + PRs welcome on the public repo. See CONTRIBUTING.md for the workflow.
  • Discussions: use GitHub Discussions for questions and design conversations.
  • Support: best-effort. NovoMCP is a reference-quality open-source project; there is no support commitment.

Reading further