NVIDIA-BioNeMo/bionemo-agent-toolkit

NVIDIA BioNeMo Agent Toolkit – What It Is

Purpose – The repository bundles a collection of agent‑ready “skills” that let a language‑model‑based agent (e.g., Claude‑Code, Codex) call NVIDIA’s life‑science models and tools. Each skill contains the instructions, scripts, and reference data an agent needs to run a specific bio‑informatics or drug‑discovery task, from protein‑structure prediction to small‑molecule docking.

Key ideas

Skills = plug‑in style wrappers around NVIDIA NIM (NVIDIA Inference Model) services and open‑source libraries.
Agents can discover and install them via the skills CLI or through the native plugin marketplaces for Claude‑Code and Codex.
Workflows combine multiple skills (e.g., generate molecules → dock → score) so an agent can orchestrate multi‑step pipelines.

How to Get Started

Install the skills CLI (a Vercel‑maintained tool). Then either interactively browse the catalog or install a single skill directly, e.g.:
```
npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit --skill boltz2-nim --yes
```

Target a specific LLM agent if you want the skill to appear in that agent’s plugin store:

npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit --skill diffdock-nim --agent claude-code

Explore the catalog without installing:

npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit --list

The repository also ships JSON marketplace files (.agents/plugins/marketplace.json, .claude-plugin/marketplace.json) so the same skills can be added through the agents’ native UI.

Main Features / Skill Catalog

Area	Example Skill	What it does (high‑level)
Protein binder design	`protein-binder-design`, `complexa-binder-design`	End‑to‑end de‑novo design of proteins that bind a target, using NVIDIA’s generative models (NIM routes) and the Proteina‑Complexa flow‑based model.
Virtual screening	`drug-discovery-pipeline`	Generates candidate molecules (GenMol), docks them (DiffDock), and scores binding affinity (Boltz‑2).
Structure prediction	`msa-structure-prediction-pipeline`	Builds a multiple‑sequence alignment (MSA‑Search via ColabFold) then predicts a complex structure with OpenFold3.
Individual NIM models	`boltz2-nim`, `diffdock-nim`, `evo2-nim`, `genmol-nim`, `molmim-nim`, `openfold2-nim`, `openfold3-nim`, `proteinmpnn-nim`, `rfdiffusion-nim`	Thin wrappers that let an agent call the corresponding NVIDIA‑hosted inference endpoint (e.g., Boltz‑2 for binding‑affinity prediction).
Open‑source models	`complexa-`, `kermt-`, `codonfm-*`	Set‑up, embed, score, or fine‑tune open‑source foundation models for protein design, ADMET prediction, or codon‑level language modeling.
Genomics acceleration	`parabricks`, `genomics-workflow-acceleration`	Skills that invoke NVIDIA Parabricks for fast GPU‑accelerated genomic pipelines.
Cheminformatics utilities	`nvmolkit-usage`	Provides GPU‑accelerated fingerprinting, similarity, clustering, conformer generation, and substructure search via the `nvMolKit` library.
Equivariant NN primitives	`cuequivariance`	Exposes building blocks for equivariant neural networks (tensor products, CG coefficients).

Each skill lives in its own folder with a SKILL.md that describes the exact steps an agent should follow, plus optional scripts/ and references/.

Evaluation / “Skill Lift”

The repo includes a testing harness (astra-skill-eval) that measures how much a skill improves an agent’s ability to complete a task. The process:

Run a hosted evaluation case (evals/evals.json) against the real NVIDIA NIM endpoint.
Compare the agent’s performance with the skill versus without it.
The ACES grader combines deterministic checks (e.g., correct API call) with LLM‑based judgment.

Example command for the Boltz‑2 skill:

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

Some workflows (OpenFold3, MSA‑Search → OpenFold3) are still pending validation and currently time out.

Who Might Use This?

AI‑augmented scientists who want a conversational agent (Claude‑Code, Codex, etc.) to run sophisticated bio‑informatics pipelines without writing boiler‑plate code.
Tool developers building custom agents that need plug‑and‑play access to NVIDIA’s life‑science models.
Researchers evaluating LLM‑agent performance on real scientific tasks, using the provided evaluation harness.

License

The code (scripts, tests) is Apache‑2.0. The skill definitions and documentation are CC‑BY‑4.0. Individual skills may pull in third‑party data with their own terms.

Bottom line – The NVIDIA BioNeMo Agent Toolkit is a genuine, production‑grade collection of ready‑to‑use wrappers that let LLM agents call state‑of‑the‑art life‑science models and libraries, turning a generic coding agent into a domain‑expert for protein design, drug discovery, genomics, and more.