NVIDIA's BioNeMo Agent Toolkit is now wired into Anthropic's Claude Science, an AI workbench for life sciences research that entered public beta on June 30, 2026. The integration lets Claude-driven agents call NVIDIA-accelerated scientific workflows — genomics analysis, protein structure prediction, cheminformatics — directly from natural-language prompts, without scientists hand-configuring models or endpoints. 18 of the top 20 pharmaceutical companies already use NVIDIA BioNeMo, which gives the toolkit an immediate install base inside the labs Anthropic most wants Claude Science to land in.
The numbers behind the speedups are the point. NVIDIA Parabricks compresses genomic analysis from hours to minutes. RAPIDS-singlecell, developed with scverse, takes a 1.3-million-cell preprocessing and clustering workflow from 52 minutes to 25 seconds. nvMolKit delivers up to 3,000x acceleration on cheminformatics operations like similarity search and conformer generation. Those are the workflows an autonomous research agent has to call repeatedly during a single reasoning loop, and at CPU speeds they would simply not fit inside an agent's working horizon.
The toolkit packages accelerated capabilities as callable skills with structured metadata: purpose, required inputs, expected outputs. Claude Science reads those descriptors, selects the right tool for the step it's reasoning about, prepares valid inputs, and dispatches the job to NVIDIA compute — wherever it's deployed.
“Claude Science lets scientists use natural language to move their research from intent into action, without manually configuring models, endpoints, or software environments”— Anthony Costa, NVIDIA
Key facts
- 01NVIDIA's BioNeMo Agent Toolkit integrates with Anthropic's Claude Science as a callable resource for research agents.
- 0218 of the top 20 pharmaceutical companies already use NVIDIA BioNeMo across drug discovery, genomics, and protein engineering.
- 03RAPIDS-singlecell cuts a 1.3-million-cell preprocessing workflow from 52 minutes to 25 seconds.
- 04nvMolKit accelerates cheminformatics operations like similarity search by up to 3,000x.
- 05Claude Science entered public beta on June 30, 2026, with toolkit access via NVIDIA developer resources and GitHub.
Under the hood, the toolkit exposes BioNeMo's open models — Evo 2 for genomics, Boltz-2 and OpenFold3 for protein structure — alongside BioNeMo NIM microservices, which package those models as containerized inference endpoints with the accelerated software stack pre-tuned. An agent calls a single stable API; the production deployment details sit underneath. The same skills work across agent frameworks, not just Claude Science.
The worked example NVIDIA describes is cancer inhibitor design. A scientist names a cancer-causing antigen mutation and asks Claude to design potential inhibitors. The agent fingerprints a compound library, clusters promising hits, generates conformers for top candidates, layers in genomic context, and compares perturbation responses — each step a separate accelerated workflow — before returning a ranked list for the researcher to review and refine.
This is what agentic science actually requires. A reasoning loop that has to wait 52 minutes for one clustering job is not a loop; it's a batch pipeline with a chatbot bolted on top. Cutting that step to 25 seconds is what makes the difference between an agent that can iterate inside a single working session and one that can't.
For Anthropic, the integration is a credible play for a vertical that has historically been dominated by domain-specific platforms rather than general-purpose chat assistants. Claude Science gives pharma researchers a reason to standardize on Claude as the reasoning layer, with NVIDIA's stack handling the heavy compute. Anthropic is inviting beta researchers to flag which additional domain specialists and integrations they want next.
“NVIDIA BioNeMo Agent Toolkit is open and harness-agnostic, allowing the same scientific skills to work across agent frameworks and research platforms”— Anthony Costa, NVIDIA
The caveats matter. Public beta means uneven coverage across scientific domains, and an agent that picks the wrong tool or mis-specifies an input in cheminformatics can waste real GPU budget before anyone notices. The toolkit's skill descriptors are only as good as the metadata NVIDIA and partners write for each workflow, and scientists will still need to inspect outputs at each step — the loop is iterative, not autonomous. Claude Science is also one of several AI-for-science workbenches now competing for the same pharma R&D budget.
Strategically, this is NVIDIA extending its life sciences moat from infrastructure into the agent layer that sits above it. The full GPU-accelerated stack — Parabricks, RAPIDS, nvMolKit, BioNeMo models, NIM microservices — represents more than a decade of investment, and packaging it as agent-callable skills means every new AI workbench in pharma routes value back to NVIDIA compute. For Anthropic, it's a way to land Claude inside the 18 of 20 pharma giants already running BioNeMo without having to rebuild the scientific tooling from scratch. The losers are the standalone bioinformatics platforms that assumed their domain expertise would protect them from horizontal AI workbenches. NVIDIA will showcase more of the stack at GTC Berlin, October 20-22.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
