NVIDIA BioNeMo Agent Toolkit

Turn any agent into a life science expert with NVIDIA BioNeMo skills.

Protein folding, molecular docking, generative chemistry, genomics analysis, protein design, and biomarker discovery — a decade of NVIDIA life sciences libraries, tools, and models, packaged as ready-to-call agent skills.

Each skill gives a coding or scientific agent structured instructions, scripts, and references to select a tool, prepare inputs, run it, inspect outputs, and explain results — across both single tasks and multi-step scientific workflows.

Install

Skills install with the skills CLI:

# interactive — pick a skill + install destination
npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit

# one skill, no prompts
npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit --skill boltz2-nim --yes

# target a specific agent (repeatable)
npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit --skill boltz2-nim --agent claude-code
npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit --skill boltz2-nim --agent codex

# browse the catalog without installing
npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit --list

The repo also ships self-hosted plugin marketplaces:

• Codex: .agents/plugins/marketplace.json
• Claude Code: .claude-plugin/marketplace.json so the bionemo-agent-toolkit plugin installs through each agent's native plugin flow as well. Skills are also discoverable by partner harnesses directly from the repo.

Skill Catalog

Product	Description	Skills
Protein Binder Design	End-to-end de novo binder design workflows — a NIM route and a Proteina-Complexa route.	protein-binder-design, complexa-binder-design
Generative Virtual Screening workflow	Generate candidate molecules, dock them to a target, and score binding affinity (GenMol → DiffDock → Boltz-2).	drug-discovery-pipeline
MSA-enabled protein structure prediction workflow	Build a multiple sequence alignment, then predict structure (MSA-Search → OpenFold3).	msa-structure-prediction-pipeline
Boltz-2	Biomolecular structure prediction + binding affinity (NIM).	boltz2-nim
DiffDock	Small-molecule docking and binding-pose prediction (NIM).	diffdock-nim
Evo 2	DNA sequence generation and variant scoring (NIM).	evo2-nim
GenMol	De novo molecule generation, scaffold decoration, lead optimization (NIM).	genmol-nim
MolMIM	Latent-space small-molecule generation and optimization (NIM).	molmim-nim
MSA-Search	Multiple sequence alignments via ColabFold (NIM).	msa-search-nim
OpenFold2	Monomer protein structure prediction (NIM).	openfold2-nim
OpenFold3	Biomolecular complex structure prediction (NIM).	openfold3-nim
ProteinMPNN	Inverse folding / sequence design for a target backbone (NIM).	proteinmpnn-nim
RFD4-Proteina	Joint protein structure and sequence design, conditioned generation, and reward-guided search through the NVIDIA hosted API or public NIM container.	rfd4-proteina-nim
RFdiffusion	De novo protein backbone and binder design (NIM).	rfdiffusion-nim
RFD4-Proteina (open model)	Native model setup, annotated-CIF authoring, generation, search, evaluation, campaign tracking, and delivery. Requires a separate RFProteina model checkout.	rfd4-proteina-setup, rfd4-proteina-annotated-cif, rfd4-proteina-generate, rfd4-proteina-search, rfd4-proteina-evaluate, rfd4-proteina-autoresearch, rfd4-proteina-commitment-curves, rfd4-proteina-campaign, rfd4-proteina-package
Proteina-Complexa	Protein binder design for protein and small molecule targets. Combines a pretrained flow-based generative model (built on La-Proteina) with inference-time optimization for high-quality binder generation.	complexa-setup, complexa-target, complexa-design, complexa-sweep, complexa-evaluate-pdbs
KERMT	Pretrained graph neural network for molecular property prediction (ADMET). Multi-task extension of GROVER with accelerated data loading via cuik-molmaker. SOTA on real-world ADMET data.	kermt-setup, kermt-infer, kermt-embed, kermt-finetune, kermt-continue-pretrain, kermt-pretrain-scratch, kermt-add-cmim-pretrain, kermt-monitor
CodonFM	Open codon foundation models: set up public Encodon checkpoints, extract coding-sequence embeddings, score masked-codon variants, and fine-tune on labeled data.	codonfm-setup, codonfm-embed, codonfm-score, codonfm-finetune
Parabricks	Agent-ready skills built on Parabricks for accelerated genomic analysis and workflows.	parabricks, genomics-workflow-acceleration
nvMolKit	GPU-accelerated cheminformatics library for molecular fingerprinting, Tanimoto/cosine similarity, Butina clustering, conformer generation (ETKDGv3), MMFF geometry optimization, and substructure search.	nvmolkit-usage
cuEquivariance	Build equivariant neural-network primitives (segmented tensor products, CG coefficients).	cuequivariance

Evaluating NIM Skill Lift

NIM skill evaluation has two levels:

1. Hosted skill lift with ACES/SkillEvaluator. The hosted skills listed below ship a single, bounded evals/evals.json case plus evals/config.yml. astra-skill-eval converts the case into Harbor tasks and runs the same prompt both with and without the skill. The task requires a call to the real hosted NVIDIA NIM endpoint; whether each agent successfully discovers and executes that call is part of the measurement. The ACES default grader combines deterministic trajectory checks with LLM judgments to report skill lift.
2. Native Harbor tasks (TODO). Deterministic RLVR-style graders are planned for workflows requiring local GPU deployment, custom scientific scoring, or additional infrastructure. They are not yet shipped as supported evaluation coverage. These tasks will live under evals/harbor/ and provide their own verifier.

The hosted lift commands below require NGC_API_KEY in the invoking shell; each skill's evals/config.yml passes it into the Harbor tasks. No local NIM deployment is required. The commands below disable the optional Astra viewer upload, allow up to twice the default agent runtime, and write reports under .nv-aces-verify/.

OpenFold3 and the MSA-Search to OpenFold3 workflow are pending validation. Their current hosted runs timed out and should not yet be used for pass/fail or model comparison.

The RFD4-Proteina NIM skill targets the released product, with a five-case dataset covering authenticated hosted inference, public container instructions, offline preparation, and result interpretation. Its live agent benchmark and a new service-account signature are pending for this revision; offline validation does not establish production inference success. See the RFD4 bundle notes for public provenance, access requirements, and client tests.

• Evo 2

  ASTRA_HARBOR_VIEWER_UPLOAD=0 astra-skill-eval evaluate nim-skills/evo2-nim --agent-eval -a claude-code --n-attempts 1 --n-concurrent 1 --timeout-multiplier 2.0 --results-dir "$PWD/.nv-aces-verify"

• OpenFold2

  ASTRA_HARBOR_VIEWER_UPLOAD=0 astra-skill-eval evaluate nim-skills/openfold2-nim --agent-eval -a claude-code --n-attempts 1 --n-concurrent 1 --timeout-multiplier 2.0 --results-dir "$PWD/.nv-aces-verify"

• OpenFold3 (pending validation)

  ASTRA_HARBOR_VIEWER_UPLOAD=0 astra-skill-eval evaluate nim-skills/openfold3-nim --agent-eval -a claude-code --n-attempts 1 --n-concurrent 1 --timeout-multiplier 2.0 --results-dir "$PWD/.nv-aces-verify"

• Boltz-2

  ASTRA_HARBOR_VIEWER_UPLOAD=0 astra-skill-eval evaluate nim-skills/boltz2-nim --agent-eval -a claude-code --n-attempts 1 --n-concurrent 1 --timeout-multiplier 2.0 --results-dir "$PWD/.nv-aces-verify"

• DiffDock

  ASTRA_HARBOR_VIEWER_UPLOAD=0 astra-skill-eval evaluate nim-skills/diffdock-nim --agent-eval -a claude-code --n-attempts 1 --n-concurrent 1 --timeout-multiplier 2.0 --results-dir "$PWD/.nv-aces-verify"

• GenMol

  ASTRA_HARBOR_VIEWER_UPLOAD=0 astra-skill-eval evaluate nim-skills/genmol-nim --agent-eval -a claude-code --n-attempts 1 --n-concurrent 1 --timeout-multiplier 2.0 --results-dir "$PWD/.nv-aces-verify"

• MolMIM

  ASTRA_HARBOR_VIEWER_UPLOAD=0 astra-skill-eval evaluate nim-skills/molmim-nim --agent-eval -a claude-code --n-attempts 1 --n-concurrent 1 --timeout-multiplier 2.0 --results-dir "$PWD/.nv-aces-verify"

• RFdiffusion

  ASTRA_HARBOR_VIEWER_UPLOAD=0 astra-skill-eval evaluate nim-skills/rfdiffusion-nim --agent-eval -a claude-code --n-attempts 1 --n-concurrent 1 --timeout-multiplier 2.0 --results-dir "$PWD/.nv-aces-verify"

• MSA-Search

  ASTRA_HARBOR_VIEWER_UPLOAD=0 astra-skill-eval evaluate nim-skills/msa-search-nim --agent-eval -a claude-code --n-attempts 1 --n-concurrent 1 --timeout-multiplier 2.0 --results-dir "$PWD/.nv-aces-verify"

• ProteinMPNN

  ASTRA_HARBOR_VIEWER_UPLOAD=0 astra-skill-eval evaluate nim-skills/proteinmpnn-nim --agent-eval -a claude-code --n-attempts 1 --n-concurrent 1 --timeout-multiplier 2.0 --results-dir "$PWD/.nv-aces-verify"

• MSA-Search to OpenFold3 workflow (pending validation)

  ASTRA_HARBOR_VIEWER_UPLOAD=0 astra-skill-eval evaluate nim-skills/meta-skills/msa-structure-prediction-pipeline --agent-eval -a claude-code --n-attempts 1 --n-concurrent 1 --timeout-multiplier 2.0 --results-dir "$PWD/.nv-aces-verify"

• Generative virtual screening workflow

  ASTRA_HARBOR_VIEWER_UPLOAD=0 astra-skill-eval evaluate nim-skills/meta-skills/drug-discovery-pipeline --agent-eval -a claude-code --n-attempts 1 --n-concurrent 1 --timeout-multiplier 2.0 --results-dir "$PWD/.nv-aces-verify"

Evaluating RFD4-Proteina Open-Model Skills

The nine RFD4-Proteina open-model skills carry their own evaluation datasets and are imported unsigned. Their native model tasks require the runtime and inputs described in each dataset. In particular, the setup smoke evaluation needs a writable model checkout at RFPROTEINA_ROOT inside each sandbox and GPU access. Installing these skills or running their offline helper tests does not provision that environment or establish a live Tier 3 pass.

Every skill is a directory with a SKILL.md (YAML frontmatter + instructions), optional references/, and optional scripts/. The generated aggregate lives in skills/bionemo-agent-toolkit/. For compatibility with the native plugin marketplaces, plugins temporarily remains a symlink to skills.

License

The toolkit uses these default licenses, with separately licensed bundles retaining their source terms:

• Source code (scripts, tests, build tooling): Apache-2.0
• Skills and documentation (SKILL.md, workflows, READMEs): CC-BY-4.0
• Imported RFD4-Proteina Rosetta helpers and regression tests: BSD-3-Clause, with CC-BY-4.0 for their skill text and documentation. See the bundle attribution.

See LICENSE for the license statement. Individual skills may reference third-party data sources with their own terms; consult each skill's references and the NOTICE file.