Build1 distinct publisher3 min readPublished
NVIDIA's BioNeMo Agent Toolkit answers a real question for agents that know a task needs folding but not which model to call. The price of that answer is a database download and a Claude Science sandbox you have to open.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
The unit of work here is a skill, and a skill is packaging: which model to run, how to format the request, which parameters matter [3]. NVIDIA's premise is that this is where general agents fail on scientific tasks, because similar packages arrive with very different environment requirements and APIs [4]. That premise is sound and it is also narrow. A skill does nothing until a NIM microservice is answering on a port.
Which is why the load-bearing line in the walkthrough is a menu path. Claude Science runs in a sandbox by default, and calling BioNeMo NIM microservices requires compute endpoints that expose local or remote GPU resources [12]. The documented route is Customize > Compute > NVIDIA BioNeMo NIM > Connect, then import the skills from GitHub, store an NVIDIA API key, and connect to local endpoints that are Docker containers using the host GPU [13]. After that you prompt the agent to pull and launch msa-search, openfold3 and boltz2, and run a smoke test on each to confirm it is healthy [14]. Having the agent stand up its own dependencies is the right call. The failure you care about is an unhealthy endpoint noticed before a structure is reported rather than after.
Storage is where adoption gets priced. The itemisation does not close. NVIDIA names roughly 490 GB for the msa-search UniRef30 database under the UniRef30-only profile, and 30 to 40 GB for the Boltz-2 and OpenFold3 containers [10][11]. That is 520 to 530 GB against a stated budget of about 700 GB, leaving 170 to 180 GB attributable to nothing the post names [21]. Provision the larger figure. The profile choice is the bigger lever anyway: UniRef30 alone is about 35% of the full 1.4 TB set, a saving of roughly 910 GB [22]. The first long-running job in this pipeline is a download.
The benchmark is a claim about NVIDIA's task set. Internal benchmarks put task correctness at 60% without the skills and 100% with them, with token efficiency roughly doubled [5]. A 40-point absolute gain that terminates exactly at 100% tells you the set contained only tasks the skills can perform [23]. For it to transfer, your work has to sit inside biology, chemistry, genomics or drug discovery as the toolkit covers them [1], your baseline agent has to have had the same GPU and databases available and been failing on model choice or request shape rather than on missing compute, and your notion of correctness has to be gradeable the way NVIDIA graded its own. The token halving is the more portable half, since naming the endpoint and the parameters removes exploration turns in any domain.
The worked example is chosen with care. It takes the nuclear-pore protein Seh1 (C1GY11) from Paracoccidioides lutzii and an uncharacterised proposed partner (C1HCX1), motivated by Figure 4e of Han, Tsenkov, Venanzi et al. in bioRxiv [17][18]. Structures from other organisms suggest Mio can complete Seh1's otherwise open beta-propeller, which makes the pair a real test of MSA-supported complex prediction rather than a monomer that any folder gets right [19]. The session then runs multiple folding models on that alignment and compares the results [24]. The toolkit leaves that comparison to you. Where Boltz-2 and OpenFold3 disagree is where a researcher should be looking, since the skill layer only guarantees that both were called properly; agreement alone says nothing about correctness.
Ranked by verification strength, evidence, and original report placement.
NVIDIA BioNeMo Agent Toolkit packages more than a decade of NVIDIA BioNeMo life sciences models, libraries, and workflows into agent-callable skills for biology, chemistry, genomics, and drug discovery.
The toolkit is built to run with any agent framework and enables complex scientific workflows using specialised domain expertise.
A general-purpose agent might recognise that a task calls for protein folding or molecular docking, but not which model to run, how to format the request, or which parameters matter.
Similar scientific software packages may have vastly different environment requirements or APIs, making orchestration and operation challenging.
NVIDIA and Anthropic worked together to integrate the BioNeMo Agent Toolkit into Claude Science, enabling agents to discover, launch, and call BioNeMo NIM microservices directly.
Claude Science is Anthropic's AI workbench for scientific research.
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 31, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
leadership
Anthropic's "not a new model" launch cost Schrodinger 8.3%. The fight moved to plumbing1 distinct publisher
product
Anthropic's protein binders got tested by outside labs. The benchmark is still Anthropic's.1 distinct publisher
science
The real disclosure in Qwen3.8-Max is the rack: 2.4T open weights, 72 GPUs, 4K tokens/sec1 distinct publisher
invest
NVIDIA buys into a hosting market that 1,900 companies fund at $260m a quarter1 distinct publisher
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Checkable on procedure, bare on results
Almost everything operational in this story is falsifiable by a reader with the right GPU: the 490 GB UniRef30 slice against a 1.4 TB full set, 30–40 GB of folding containers, a named menu path, two UniProt accessions and a cited bioRxiv DOI. The exception is the claim that sells the toolkit — perfect task correctness — which NVIDIA gives as a bare pair of numbers, and there is no second account of any of it.
Shipped connector, no users on record
What can actually be observed is availability: a jointly built integration exists, skills can be imported from GitHub, and three microservices come up healthy on a host GPU. Nobody outside NVIDIA and Anthropic appears anywhere in this reporting — no lab, no pilot, no download or session count — so the honest reading is that the door is open and we have no idea who has walked through it.
One unverifiable headline over an honest invoice
The overstatement is narrow and lands in one sentence. NVIDIA is unusually frank about what this costs — hundreds of gigabytes, a sandbox you must open, an approval per endpoint — but the claim that would justify that cost is a perfect score on a benchmark nobody can see, and even the disclosed storage total does not add up: 490 GB plus 30–40 GB leaves 170–180 GB of the stated 700 GB unaccounted for.
The GPU vendor writing its own how-to
This runs on NVIDIA's developer blog, specifies L40S or H100 silicon, routes through an NVIDIA API key, and its thesis is that agents should call NVIDIA microservices; Anthropic gains a marquee scientific capability from the same page. None of that makes the setup steps wrong — first-party docs are usually the most accurate thing about a product — but the framing of the problem and the one flattering benchmark both come from the party that profits when you follow along.
Trust the recipe, hold the result
Confidence here splits cleanly by claim type. The storage figures, hardware floor, container list and biology example are safe to plan against; the 60%-to-100% correctness claim should sit in quotation marks until somebody outside NVIDIA reruns it, and with a single publisher there is nothing else to triangulate on.