NVIDIA BioNeMo Agent Toolkit: Agent Skills for Protein, Docking and Genomics Workflows
Turn any agent into a life science expert with NVIDIA BioNeMo skills.
At a glance
- What is it?
- The BioNeMo Agent Toolkit packages NVIDIA life science models and libraries as installable agent skills. It is a catalog of instructions and scripts rather than a Python package, and the README is explicit about what is and is not evaluated yet.
- Who is it for?
- Adopt it if you already drive a coding agent such as Claude Code or Codex and want protein, docking or genomics tooling callable from that agent without hand-writing tool wrappers. Do not adopt it if you need a supported Python library to import, or if your work depends on deterministic graders, since the README marks native Harbor tasks as a TODO.
- Can I use it commercially?
- Check first. The repository uses a licence we do not classify automatically, so read its LICENSE file before any commercial use.
- Is it still maintained?
- Yes. The repository last received commits 1 day ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 18, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The problem: life science models are callable but not agent-ready
A protein folding model or a docking service is only half the work. Someone still has to decide which tool fits the question, format the inputs, run the job, read the output files and explain the numbers. The BioNeMo Agent Toolkit addresses that gap by packaging each model or library as a skill: a directory of structured instructions, scripts and references that a coding or scientific agent can follow. The README describes the intent as giving an agent the ability to "select a tool, prepare inputs, run it, inspect outputs, and explain results", across single tasks and multi-step workflows. The audience is therefore not a bioinformatician writing a pipeline by hand. It is a team that already runs an agent and wants domain tooling inside it.
Skills, meta-skills and NIM endpoints: how the catalog is organised
The repository splits into four top-level skill groups. nim-skills holds one skill per hosted model, such as boltz2-nim, diffdock-nim, evo2-nim, genmol-nim, molmim-nim, msa-search-nim, openfold2-nim, openfold3-nim, proteinmpnn-nim and rfdiffusion-nim. open-models-skills covers models you run yourself: Proteina-Complexa is split across five skills (complexa-setup, complexa-target, complexa-design, complexa-sweep, complexa-evaluate-pdbs), and KERMT across eight, from kermt-setup through kermt-finetune, kermt-continue-pretrain, kermt-pretrain-scratch, kermt-add-cmim-pretrain and kermt-monitor. library-skills wraps non-model tooling: Parabricks for genomic analysis, nvMolKit for GPU-accelerated cheminformatics (fingerprinting, Tanimoto and cosine similarity, Butina clustering, ETKDGv3 conformer generation, MMFF optimisation, substructure search) and cuEquivariance for equivariant neural-network primitives. workflows sits above all of that with meta-skills that chain models: drug-discovery-pipeline runs GenMol, then DiffDock, then Boltz-2 to generate molecules, dock them and score affinity; msa-structure-prediction-pipeline pairs MSA-Search with OpenFold3.
The split matters because the runtime differs. A nim-skill expects a hosted NVIDIA NIM endpoint and an NGC_API_KEY in the shell. A complexa or kermt skill expects local execution. Chaining two NIM skills in a meta-skill means two network round trips and two sets of input preparation, which is where most of the failure surface lives.
Installing a skill with the skills CLI and running a first task
Installation goes through the skills CLI from vercel-labs, invoked with npx. The README gives an interactive form that prompts for a skill and a destination, and a non-interactive form that names the skill and skips prompts:
npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit --skill boltz2-nim --yesTo target a specific agent, the --agent flag is repeatable, so the same skill can land in more than one harness:
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 codexBefore installing anything, list what the repository offers:
npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit --listThe README also ships self-hosted plugin marketplaces at .agents/plugins/marketplace.json for Codex and .claude-plugin/marketplace.json for Claude Code, so the bionemo-agent-toolkit plugin can be installed through each agent's native plugin flow instead of the CLI. After either route, the skill's directory appears in the agent's skill location and the agent can read its instructions. Expect the agent to ask for inputs the skill cannot infer, such as a target sequence or a PDB identifier; the skill supplies the procedure, not the biology.
What the evaluation harness does and does not cover
The README is unusually direct here, and it is the most useful part of the document for anyone judging maturity. Each NIM skill ships one bounded case in evals/evals.json plus an evals/config.yml. The astra-skill-eval tool converts that case into a Harbor task and runs the same prompt twice, with and without the skill, so the reported number is skill lift rather than raw accuracy. The task requires a call to the real hosted NIM endpoint, and whether the agent finds and executes that call is itself part of the measurement. Grading combines deterministic trajectory checks with LLM judgments.
The second level, native Harbor tasks with deterministic RLVR-style graders, is marked TODO in the README and is not shipped as supported evaluation coverage. Those tasks would live under evals/harbor/ with their own verifier, and they are the ones that local GPU deployment, custom scientific scoring or extra infrastructure would need. One case per skill is a small sample. Treat the lift figure as a signal that the skill changed agent behaviour, not as a validated accuracy claim about the underlying model.
It is a catalog, not a Python package
The most common misreading of this repository is as an installable library. pyproject.toml says the opposite in its own comments: the repository is a catalog of agent skills, not an installable Python package, and the manifest exists so that the open-source dependency tree of the helper scripts is declared in machine-readable form for NVIDIA nSpect and Black Duck scanning. The declared dependencies are only biotite>=1.0, numpy>=1.26 and pyyaml>=6.0, mapped in comments to specific scripts under complexa-binder-design and protein-binder-design. Everything else the scripts import is Python standard library. If you were hoping to pip install a toolkit and call functions, this is the wrong project, and the README does not offer that path. The scripts are consumed by the agent as files, not imported by your code.
Where the toolkit is the wrong choice
Three cases stand out. First, reproducibility-critical work: if you need a frozen, versioned pipeline whose outputs you can regenerate bit for bit, an agent that selects tools and prepares inputs introduces variability the evaluation harness only partly measures. Second, offline or air-gapped environments: every NIM skill requires NGC_API_KEY in the invoking shell and a call to the hosted endpoint, and the README notes that no local NIM deployment is required for the hosted evaluations, which implies the hosted route is the expected one. Third, teams without a supported agent harness. The install paths assume the skills CLI or a Codex or Claude Code plugin marketplace; the README mentions partner harnesses discovering skills directly from the repo but does not enumerate them, so support outside those two flows is not documented.
A subtler limitation is scope. The catalog covers structure prediction, docking, generation and genomics acceleration. It does not cover wet-lab data management, assay analysis or regulatory documentation, and nothing in the README suggests it intends to.
Alternatives and how they differ
The nearest alternative is calling the same NVIDIA NIM endpoints directly from your own code. That gives you full control over retries, caching and input validation, and it removes the agent from the loop entirely. The difference is where the domain knowledge lives: with direct calls you write the input preparation and output interpretation yourself, while the toolkit encodes that as skill instructions the agent follows. For a fixed, well-understood pipeline, direct calls are simpler and more reproducible. For exploratory work where the right tool is not known in advance, the skill layer is the point.
A second alternative is building agent tools on a general framework such as the NeMo agent toolkit and wiring the BioNeMo models in yourself. That path gives you one integration surface for many tools but requires you to author the model-specific instructions that this repository already ships. The trade-off is maintenance: a general framework tracks its own release cadence, while these skills track the BioNeMo model endpoints.
Maintenance, upgrade cost and the licence split
The repository is not archived, and the last push was on 2026-09-17. The only release listed is v0.1.0 from 2026-06-23, described as the initial release. A single release in the project's history means upgrade guidance is thin: the README does not document a migration path between skill versions, and because skills are copied into an agent's skill directory rather than resolved as a dependency, upgrading means re-running the install command for each skill and each agent target. There is no lockfile for skills themselves; uv.lock covers only the helper-script dependencies.
On licensing, the repository is not straightforward. The declared project licence in pyproject.toml is "Apache-2.0 OR CC-BY-4.0", and the repository root carries LICENSE, LICENSE-APACHE-2.0, LICENSE-CC-BY-4.0 and NOTICE as separate files. The GitHub licence field reports NOASSERTION, which is consistent with a dual-licence layout that no single SPDX identifier captures. The OR means you choose one of the two, and the two are not equivalent: CC-BY-4.0 is a content licence suited to documentation and reference material, while Apache-2.0 carries an explicit patent grant and is the usual choice for code. Which one is appropriate depends on what you redistribute, and the README does not map skills to licences. Check the LICENSE files in the specific skill directory you install rather than assuming the root declaration settles it. This is not legal advice.
Editorial conclusion
Adopt it if you already drive a coding agent such as Claude Code or Codex and want protein, docking or genomics tooling callable from that agent without hand-writing tool wrappers. Do not adopt it if you need a supported Python library to import, or if your work depends on deterministic graders, since the README marks native Harbor tasks as a TODO. Before committing, verify three things: that the skill you need exists in the catalog table, that you have an NGC_API_KEY for any NIM-backed skill, and which licence applies to the specific skill directory you install, because the repository carries both Apache-2.0 and CC-BY-4.0 files and pyproject.toml declares them as alternatives.
Frequently asked questions
How are AI agents being used in drug discovery?
In this toolkit, agents are used to select a life science tool, prepare its inputs, run it, inspect the outputs and explain the results, both for single tasks and for multi-step workflows such as generating molecules with GenMol, docking them with DiffDock and scoring affinity with Boltz-2.
What is the NeMo agent toolkit?
The BioNeMo Agent Toolkit is a separate project from the NeMo agent toolkit. The README describes it as a catalog of agent skills that package NVIDIA life sciences libraries, tools and models, and it can be installed through the skills CLI or through Codex and Claude Code plugin marketplaces.
Is the BioNeMo Agent Toolkit a Python package I can pip install?
No. The pyproject.toml comments state that the repository is a catalog of agent skills, not an installable Python package, and the manifest exists so the dependency tree of the helper scripts can be scanned. The declared dependencies are biotite, numpy and pyyaml.
What do I need to run a NIM skill from the BioNeMo Agent Toolkit?
The hosted lift evaluations require NGC_API_KEY in the invoking shell, and each skill's evals/config.yml passes it into the Harbor tasks. The README states that no local NIM deployment is required for those evaluations.
How does the BioNeMo Agent Toolkit measure whether a skill helps?
Each NIM skill ships one bounded case in evals/evals.json plus evals/config.yml, and astra-skill-eval runs the same prompt with and without the skill against the real hosted NIM endpoint. The default ACES grader combines deterministic trajectory checks with LLM judgments to report skill lift.
Community notes