The engine. On your hardware. On your terms.
An MCP + REST server that any MCP-compatible AI assistant can call — Claude Desktop, Cursor, Codex, Zed, Cline, and others. 68 tools in the catalog; 11 work with a fresh clone and no wiring. Add optional compute services as you need them.
$ git clone https://github.com/NovoMCP/novomcp \
&& cd novomcp/orchestrator \
&& pip install -r requirements.txt \
&& python main_https.py$ git clone https://github.com/NovoMCP/novomcp \
&& cd novomcp \
&& docker compose upTwo install paths today (git clone + docker compose). The pip install novomcp and npx novo packages are name reservations, not installable builds; releases will land on docs.novomcp.com and in GitHub Releases.
$ curl -X POST http://localhost:8018/mcp/call \
-H 'Content-Type: application/json' \
-H 'Authorization: Bearer local' \
-d '{
"name": "get_molecule_profile",
"arguments": { "smiles": "CC(=O)Oc1ccccc1C(=O)O" }
}'Local calls use the placeholder bearer token local. Wire NOVOMCP_API_KEY when you deploy behind an auth gate.
Eleven tools work the second the server boots.
- Molecular properties (LogP, TPSA, HBA/HBD, MW, and 20 more) via RDKit in-process
- Similarity search against public collections
- Literature + ClinicalTrials + bioRxiv search
- Autonomous discovery-mode prompts (Novo AG)
- Engine health, tool catalog, credit meta
Point env vars at your own services. Each capability is provider-agnostic.
The engine speaks HTTP to each service. Ship the reference containers from the same repo, or swap in your own — nothing about the interfaces below locks you to a specific provider. Compliance (check_compliance) is the exception: it calls the hosted FAVES API directly, which is a specific service with its own contract.
FEP tools (run_fep_check, run_fep_screen) are not in the open engine — they run on the paid Novo Compute surface.
Observability built in, not bolted on.
The engine writes an immutable audit trail; compliance runs at the point of decision; memory persists across runs; agents acting on your pipeline are scoped and logged.
Immutable audit trail
Every decision the AI makes is written to an immutable audit trail. Per-stage logging across the full discovery funnel — reproducible, reconstructable, exportable.
FAVES V4 compliance
1,585 SMARTS across 8 jurisdictions, run inline at the point of decision rather than as a terminal gate. Compliance is a property the engine checks while it reasons.
Cross-run memory
Funnel context persists across sessions. The engine remembers what prior runs learned and carries it into the next — a program compounds knowledge instead of restarting it.
Agent governance
RFC 8693 token exchange, per-agent key scoping, and agent-level audit on cloud deployments. Each agent acting on your pipeline is identified, scoped, and logged.
Where we are not yet
SOC 2 not started. Formal validation not started — published benchmarks stand in for it openly. NovoExpert-3 is competent on cardiovascular and GI and not yet validated on oncology, CNS, or infectious indications. Stating what's unfinished is part of the audit.
Try it locally in five minutes.
Clone, install, ask an assistant about aspirin. If the engine breaks, open an issue.