The NovoMCP compute core, on one desktop system.
GPU molecular dynamics, GPU docking, neural-network potentials, quantum chemistry and 31 ADMET models, running on NVIDIA DGX Spark. One compose file brings up the stack. Any MCP-compatible AI assistant drives it.
DGX Spark · GB10
chip
memory
file
vs an L40S
The engine's compute services, running locally.
Outside this release: structure prediction, generative design and free-energy methods. Those paths run on the standard engine deployment. Status as of October 2026.
Pull the images. Start the stack.
Prebuilt arm64 images for GB10. Nothing to compile. The engine comes up on localhost:8018 with every compute service wired to it on the local network.
docker compose -f compose.spark.yml up -d
docker compose -f compose.spark.yml ps # every service: runningRequires a DGX Spark or GB10-class system, Docker with the NVIDIA Container Toolkit, and about 40 GB of disk for the images. Full steps in the quickstart.
A MACE geometry relaxation on the GB10 GPU.
curl -s -X POST http://localhost:8018/mcp/tools/optimize_geometry_nnp \
-H 'X-API-Key: local' -H 'Content-Type: application/json' \
-d '{"arguments": {"smiles": "CCO", "method": "mace"}}'{
"result": {
"smiles": "CCO",
"energy_ev": -46.243916,
"energy_kcal_mol": -1066.4263,
"optimized_xyz": "9\nEnergy: -46.243916 eV\nC -0.89535478 0.19299133 -0.00831505\nC 0.46489329 -0.51024634 -0.04683582\nO 1.47175235 0.31191074 0.58511584\nH -0.86542882 1.14266840 -0.56420034\nH -1.64603085 -0.45831647 -0.47180070\nH -1.19028783 0.40331596 1.02890327\nH 0.75515675 -0.75208147 -1.08707740\nH 0.43229280 -1.46669087 0.49268060\nH 1.47300708 1.13644872 0.07152960\n",
"forces_max_ev_ang": 0.038468,
"converged": true,
"n_steps": 8,
"method": "MACE-MPA-0-BFGS",
"n_atoms": 9,
"wall_time_ms": 401.2,
"engine_used": "ase",
"note": null
},
"usage": {
"tool": "optimize_geometry_nnp",
"funnel_id": "fnl_u25bf8e1a2393_29872630"
}
}Ethanol, relaxed with MACE-MPA-0 on the GPU in 8 BFGS steps (−46.244 eV), the same relaxed energy reported in the field report. A GFN2-xTB single point (run_qm_calculation) and GPU docking (dock_molecules) run the same way.
1.2–1.5× the molecular dynamics throughput per watt of a datacenter L40S.
Raw MD speed on GB10 is about 0.4× an L40S. The workload is memory-bandwidth bound, and GB10 measures about 227 GB/s. The advantage is efficiency and a self-contained system. Double-precision math runs about 34× slower than single precision on GB10, so quantum chemistry stays on the CPU and neural-network potentials run in single precision on the GPU.
GROMACS SPC/E water MD benchmark · single DGX Spark (GB10) vs single L40S · CUDA 13.0, DGX OS, drivers 580.14x / 580.173.02 · reproduced on a second, independent GB10 · NovoMCP, 2026
An autonomous discovery loop on a desktop agent computer.
DGX Spark is built for always-on, local agent workloads. NovoMCP is the tool layer those agents call. Point any MCP-compatible assistant at localhost:8018, or pair it with a local model on the same machine. The agent plans the run, calls the GPU and quantum tools, ranks candidates, and writes every step to the local audit log.
The materials discovery funnel (mgm) runs fully on the Spark. The drug discovery funnel (agm) also calls omics, literature and compliance services that run outside the Spark.
Compute on hardware you own.
Pharma and biotech
Compute on proprietary structures stays in your building. Every run is written to a local audit log.
Academic labs
A one-time hardware cost. The engine and its models are open source.
Every team
The full compute stack in one desktop system, brought up from one compose file.
Three default tools query public databases (ChEMBL, RCSB and bioRxiv). Computation stays local.
Rent it, self-host it, or own the desk.
Rent GPUs
Wire the engine to on-demand GPU providers. Pay per run.
Your cloud
Self-host the full engine on your own cloud account.
Novo Station
Run the compute core on a DGX Spark you own.
Five porting failures on GB10, and the fix for each.
Porting the engine to DGX Spark surfaced five failures. All five resolved in packaging, threading, memory allocation and versions. The field report documents each one and the recipe that works.
The reproducible MACE and GROMACS workload is submitted to NVIDIA's DGX Spark playbooks (PR #126, under review).
What actually breaks when you port a chemistry stack to DGX Spark
Do I need a DGX Spark?
Novo Station targets GB10-class systems: DGX Spark and other GB10 machines. On other NVIDIA GPUs, the standard engine images run the same tools.
What does it cost?
The engine and Novo Station are free and open source under Apache-2.0. A DGX Spark is a one-time hardware purchase. Hands-on deployment is available as a paid service.
Does any data leave the machine?
Computation runs locally. Three default tools query public databases (ChEMBL, RCSB and bioRxiv). With a local model as the agent, the run itself stays on the Spark.
What isn't included yet?
Structure prediction, generative design and free-energy methods. They run on the standard engine deployment. The quickstart keeps a current per-service status.
How fast is it?
For molecular dynamics, about 0.4× a datacenter L40S, at about a third of the power: 1.2–1.5× the throughput per watt. The measurements and hardware details are above.
Can I try it without buying a Spark?
Yes. Rent a DGX Spark by the hour (see "Three ways to run the engine") and run the quickstart against test molecules.
Build it yourself, or have it deployed.
Build it yourself
Free and open source. Pull the images and follow the quickstart.
Have it deployed
NovoMCP sets up the engine on your Spark, end to end.
NVIDIA, DGX and DGX Spark are trademarks and/or registered trademarks of NVIDIA Corporation. NovoMCP is not affiliated with or endorsed by NVIDIA.