NVIDIA BioNeMo Agent Toolkit: 100% Task Completion (2026)
NVIDIA BioNeMo Agent Toolkit gives AI agents 6+ NIM skills for protein folding, docking, and genomics, lifting task completion from 57% to 100%. Free to start.
NVIDIA BioNeMo Agent Toolkit turns any AI agent into a life-science expert with free, open-source skills for protein folding, molecular docking, genomics, and drug discovery. It plugs into Claude, Codex, or any agent runtime via NVIDIA NIM microservices, needs only an NVIDIA API key to start, and raised task completion from 57% to 100% in NVIDIA's own tests.
50+ companies already use NVIDIA BioNeMo Agent Toolkit, a free, open-source library of agent-callable skills for protein folding, molecular docking, genomic variant calling, and cheminformatics screening. Built on NIM microservices like Boltz-2, RFdiffusion, Evo 2, and Parabricks, it raised agent task completion from 57.1% to 100% in NVIDIA's benchmark, and works with Claude, Codex, or any agent runtime.
Maker: NVIDIA · Protocol: CUSTOM · Auth: api key
Compatible agents: Claude (Claude Code, Claude Science), Codex, Any agent runtime supporting Anthropic Skills or MCP tool-calling
Required runtime: Node.js (for the npx skills installer), NVIDIA API key from build.nvidia.com for hosted NIM endpoints, Optional local GPU node for self-hosted NIM deployment
About NVIDIA BioNeMo Agent Toolkit
NVIDIA BioNeMo Agent Toolkit is an open-source library of agent-callable skills that packages more than a decade of NVIDIA life sciences software, including BioNeMo, Parabricks, nvMolKit, and RAPIDS-singlecell, into tools any AI agent can call directly. Rather than a standalone app a scientist opens, it is a set of instructions, scripts, and references that let an agent select the right model, prepare valid inputs, run the job, inspect the output, and explain results across protein folding, molecular docking, generative chemistry, genomics, protein design, and biomarker discovery. NVIDIA reported that agents without these skills completed only 57.1% of required scientific tasks on average, while agents with the skills attached reached 100% completion and produced twice as many passing assertions per 1,000 tokens. Each skill ships as a directory containing a SKILL.md file with YAML frontmatter and instructions, plus optional reference material and scripts, and is installed with a single command: npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit. Under the hood, most skills call NVIDIA NIM microservices over a REST API secured with an NVIDIA API key from build.nvidia.com, while a smaller set of open models not yet packaged as NIM are exposed through MCP server wrappers. The toolkit is explicitly agent-agnostic: it was designed to work the same way whether the calling agent is Claude, Codex, or another runtime, and NVIDIA states similar performance holds across different backend LLMs. At launch, more than 50 life sciences companies were already using the toolkit, and Anthropic wired it directly into Claude Science, its public-beta workbench for scientists, so that domain-specialized agents in genomics, proteomics, single-cell analysis, cheminformatics, and clinical research can call BioNeMo skills as a built-in resource. Typical workflows include chaining RFdiffusion to generate candidate protein backbones with Boltz-2 to fold and score them, running Parabricks to call genomic variants in minutes instead of hours, and using nvMolKit to screen thousands of candidate molecules before committing to a full docking run. The skills and their code are free and open source under a CC BY 4.0 license, and the hosted NIM endpoints on build.nvidia.com are free to use for prototyping, capped around 40 requests per minute. Production deployment of self-hosted NIM microservices requires an NVIDIA AI Enterprise license, available with a 90-day free trial before paid tiers apply, and a GPU node is optional unless running NIMs locally. NVIDIA announced the toolkit on June 23, 2026 at BIO 2026, with the GitHub repository seeing active development shortly after, including the July 2026 Claude Science integration. NVIDIA continues to expand the skill catalog and has flagged validation as an open concern: the skills improve how reliably an agent invokes the right model, not the underlying scientific accuracy of that model, so low-confidence structures and generated molecules still need manual review before use.
Key Features
- Task completion jumps from 57% to 100%: NVIDIA's internal benchmark found agents without BioNeMo skills completed only 57.1% of required scientific tasks, versus 100% with skills attached.
- GPU-accelerated genomics: Parabricks cuts whole-genome variant calling from hours on CPU down to minutes on GPU.
- 52-minute to 25-second single-cell preprocessing: RAPIDS-singlecell compresses a 1.3-million-cell preprocessing workflow from 52 minutes down to 25 seconds.
- Up to 3,000x faster cheminformatics: nvMolKit accelerates molecular similarity search and conformer generation operations by up to 3,000x over CPU RDKit.
- Agent- and model-agnostic skills: Every skill ships as a SKILL.md package usable by Claude, Codex, or any agent runtime, and works with any backend LLM, not just NVIDIA's own models.
- Adopted by 50+ life sciences companies: More than 50 companies were already using the toolkit's callable skills for protein design, docking, and biomarker discovery at launch.
Use Cases
- Protein binder design: An agent chains RFdiffusion to generate candidate backbones and Boltz-2 to fold and score each candidate against a disease target in one automated loop.
- Genomic variant triage: An agent runs Parabricks on a patient's FASTQ file to call variants in minutes, then uses Evo 2 to predict which variants likely disrupt protein function.
- Virtual compound screening: An agent uses nvMolKit to screen thousands of candidate molecules by similarity before requesting full docking runs, cutting screening time from hours to seconds.
- Literature-grounded hypothesis generation: A science agent like Claude Science combines BioNeMo structure predictions with literature search to propose and justify a next experiment.
Install
npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkitRequirements
- Node.js runtime for the npx skills installer
- NVIDIA API key from build.nvidia.com for hosted NIM endpoints (starts with nvapi-)
- Optional local GPU node with an NVIDIA AI Enterprise license for self-hosted, production NIM deployment
Actions
Boltz-2 Structure Prediction and Scoring
Predicts protein-ligand complex structures and binding affinity scores for co-folding and validation in protein binder design.
npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit --skill boltz2-nim --yessequence(string) — required: Protein amino acid sequence (FASTA) to fold.ligand_smiles(string): SMILES string of a ligand to co-fold and score binding affinity against.
RFdiffusion Protein Backbone Design
Generates novel protein backbone scaffolds for binder design workflows using diffusion-based generative modeling.
npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit --skill rfdiffusion-nim --yestarget_pdb(string) — required: Target protein structure to design a binder against.num_designs(number): Number of backbone designs to generate.
Evo 2 Genomic Sequence Analysis
Runs the Evo 2 foundation model over DNA sequences to predict variant effects and generate genomic sequences.
npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit --skill evo2-nim --yessequence(string) — required: Input DNA sequence, supported up to megabase-scale context.task(string): Analysis task to run: 'variant_effect' or 'generate'.
OpenFold3 Complex Structure Prediction
Predicts protein, nucleic acid, and small-molecule complex structures as an open alternative to closed folding models.
npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit --skill openfold3-nim --yessequences(array) — required: One or more biomolecule sequences (protein/RNA/DNA) to co-fold.
Parabricks Genomic Variant Calling
Accelerates whole-genome and exome variant calling pipelines, cutting hours-long CPU pipelines down to minutes on GPU.
npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit --skill parabricks-nim --yesfastq_input(string) — required: Path or URL to input FASTQ sequencing reads.reference_genome(string) — required: Reference genome build to align against, e.g. GRCh38.
nvMolKit Cheminformatics Screening
Runs GPU-accelerated molecular similarity search, conformer generation, and generative chemistry screening up to 3,000x faster than CPU RDKit pipelines.
npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit --skill nvmolkit --yessmiles_list(array) — required: List of candidate molecule SMILES strings to screen.operation(string): Operation to run: 'similarity_search', 'conformer_generation', or 'docking_prep'.
How to Invoke
Installed as SKILL.md packages via 'npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit', then invoked by name by any compatible agent (Claude, Codex, or another runtime); each skill calls an underlying NVIDIA NIM microservice over REST, with MCP server wrappers used for models not yet packaged as NIM.
Pricing
Skills and code are free and open source (CC BY 4.0). Hosted NIM endpoints on build.nvidia.com are free for prototyping, capped around 40 requests/minute. Production self-hosted NIM deployment requires an NVIDIA AI Enterprise license, with a 90-day free trial before paid tiers apply.
Strengths
- Raises agent task completion on scientific workflows from 57.1% to 100% in NVIDIA's own benchmark.
- Packages a decade of NVIDIA life sciences tooling (Parabricks, BioNeMo, nvMolKit) as ready-to-call skills instead of requiring custom integration code.
- Agent-agnostic: works with Claude, Codex, or any agent runtime, and any underlying LLM.
- Free hosted NIM endpoints on build.nvidia.com for prototyping, with no infrastructure to manage.
Weaknesses
- Skills improve task completion and efficiency but do not make the underlying scientific models more accurate; low-confidence structures and generated molecules still need manual validation.
- Hosted build.nvidia.com endpoints are capped around 40 requests/minute and are explicitly for small-scale development, not production inference.
- Production deployment of self-hosted NIMs requires a paid NVIDIA AI Enterprise license once the 90-day trial ends.
Frequently Asked Questions
What is NVIDIA BioNeMo Agent Toolkit and what does it do?
NVIDIA BioNeMo Agent Toolkit is a free, open-source library of agent-callable skills that gives any AI agent life-science capabilities: protein folding, molecular docking, generative chemistry, genomics analysis, protein design, and biomarker discovery. It packages more than a decade of NVIDIA life sciences software, including BioNeMo, Parabricks, and nvMolKit, as skills an agent can select, run, and interpret on its own. NVIDIA announced it on June 23, 2026 at the BIO conference. In NVIDIA's own benchmark, agents without these skills completed only 57.1% of required scientific tasks, while agents with the skills reached 100% completion. The toolkit is agent-agnostic, meaning it works the same whether the calling agent is Claude, Codex, or another runtime.
How do I install and set up NVIDIA BioNeMo Agent Toolkit?
Install it with a single command: npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit, or target one specific skill with npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit --skill boltz2-nim --yes. Each skill is a directory containing a SKILL.md file with YAML frontmatter plus optional reference material and scripts that the agent reads at run time. Most skills call NVIDIA NIM microservices over REST, so you need an NVIDIA API key from build.nvidia.com, which typically starts with nvapi-. A GPU node is optional and only required if you choose to self-host a NIM locally instead of using the hosted endpoints. No separate database or account setup is needed beyond the API key.
Which agents and LLMs support NVIDIA BioNeMo Agent Toolkit?
The toolkit is built to be agent-agnostic, so it works with Claude (including Claude Code), Codex, and any other agent runtime that can install and read SKILL.md-format skills. Anthropic has integrated it directly into Claude Science, its public-beta workbench for scientists, where domain-specialized agents in genomics, proteomics, single-cell analysis, cheminformatics, and clinical research call BioNeMo skills as a built-in resource. A smaller number of open models not yet packaged as NIM microservices are exposed through MCP server wrappers instead. NVIDIA states similar performance gains can be expected regardless of which backend LLM is driving the calling agent.
How much does NVIDIA BioNeMo Agent Toolkit cost?
The skills and their code are free and open source under a CC BY 4.0 license, so there is no charge to install or read them. Hosted NIM endpoints on build.nvidia.com are also free to call for prototyping, with a community-observed cap of roughly 40 requests per minute rather than a published SLA. Production use that requires reliable throughput calls for self-hosted NIM deployment, which needs an NVIDIA AI Enterprise license; NVIDIA offers a 90-day free trial before paid enterprise pricing applies. A GPU node is only required for that self-hosted path, not for using the hosted prototyping endpoints.
Is NVIDIA BioNeMo Agent Toolkit open source?
Yes. The toolkit and its skills are openly published on GitHub at NVIDIA-BioNeMo/bionemo-agent-toolkit under a CC BY 4.0 license, and NVIDIA has stated the underlying BioNeMo models, training code, and weights are also fully open. NVIDIA maintains the repository and continues to add new skills and integrations, including the July 2026 Claude Science wiring. Individual underlying components, such as specific NIM microservices, may carry their own license terms, so teams should check each resource before production use. The open license is what lets any agent vendor or research lab install and modify the skills freely.
What are the best alternatives to NVIDIA BioNeMo Agent Toolkit?
The closest comparisons are general-purpose agent frameworks like NVIDIA's own NeMo Agent Toolkit, which connects and optimizes teams of AI agents but has no life-science domain skills baked in, so it needs custom tool-building to reach parity. Standalone scientific tools such as AlphaFold3 or RFdiffusion cover a single capability (folding or backbone design) rather than a full callable skills suite an agent can chain together. Custom in-house MCP servers wrapping individual bio models are the other alternative, but they require building and maintaining the input/output handling BioNeMo Agent Toolkit already ships. Pick BioNeMo Agent Toolkit when you want a ready-made, multi-skill suite; pick a single-model wrapper only if you need just one capability.
Who is NVIDIA BioNeMo Agent Toolkit best for?
It is best for computational biologists, bioinformatics engineers, and pharma R&D teams who are building agent-driven drug discovery or genomics pipelines and want protein folding, docking, and variant calling available as callable skills instead of custom integration code. AI and agent developers who want to add life-science expertise to an existing general-purpose agent are also a strong fit. A concrete use case is a team chaining RFdiffusion and Boltz-2 to design and score protein binders in one automated loop. It is not a good fit for casual users with no scientific workflow, or teams needing production-grade inference guarantees without first securing an NVIDIA AI Enterprise license.
How does NVIDIA BioNeMo Agent Toolkit compare to NVIDIA NeMo Agent Toolkit in 2026?
NVIDIA NeMo Agent Toolkit is a general-purpose, open-source library for connecting and optimizing teams of AI agents across any domain, while BioNeMo Agent Toolkit is a domain-specific skills package built on top of that broader ecosystem for life sciences. BioNeMo ships ready-made skills like Boltz-2 folding, RFdiffusion backbone design, and Parabricks variant calling with real parameter schemas, whereas NeMo Agent Toolkit expects you to define your own tools and workflows. If the goal is scientific discovery work such as protein design or genomics, BioNeMo Agent Toolkit gets an agent productive immediately; if the goal is orchestrating a custom multi-agent system with no life-science focus, NeMo Agent Toolkit is the correct base layer.