by NVIDIA
Agent-callable skills for protein folding, docking, and genomics on NVIDIA NIM
The BioNeMo Agent Toolkit gives AI agents open-source skills for NVIDIA life-science models, installed with one npx command. NVIDIA reports agent task completion rose from 57.1% to 100% in its own tests. Installing is free, and hosted endpoints need an NVIDIA API key.
Turning a decade of NVIDIA's own accelerated life-science libraries into drop-in, agent-callable skills with a single install command.
NVIDIA controls both the GPU hardware and a decade of proprietary life-science model IP (BioNeMo, Parabricks), making the underlying acceleration hard for a software-only competitor to match.
Main trade-off
Skills route agent calls correctly but do not improve the accuracy of the underlying scientific models, so outputs still require expert validation.
The toolkit costs nothing to install: scripts are Apache-2.0 and skills and documentation are CC-BY-4.0 per the repository's License section. Most NIM skills call NVIDIA-hosted endpoints with an API key from build.nvidia.com; the repository does not state a rate limit or a price for those endpoints, so check NVIDIA's current terms before relying on them. Local deployment of any model needs a GPU you supply. See NVIDIA and the skills directory for related listings.
npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit
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.
The permissions this connector asks for, as its own documentation states them.
6 actions an agent can call. Open one for its parameters and an example.
Example call
npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit --skill boltz2-nim --yes
Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| sequence | string | Required | Protein amino acid sequence (FASTA) to fold. |
| ligand_smiles | string | Optional | SMILES string of a ligand to co-fold and score binding affinity against. |
Returns
| Field | Type | Description |
|---|---|---|
| structure_pdb | string | Predicted 3D structure in PDB/mmCIF format. |
| confidence_score | number | Per-residue and complex confidence score. |
Example call
npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit --skill rfdiffusion-nim --yes
Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| target_pdb | string | Required | Target protein structure to design a binder against. |
| num_designs | numberdefault 10 | Optional | Number of backbone designs to generate. |
Returns
| Field | Type | Description |
|---|---|---|
| backbone_pdbs | array | Generated backbone structures as PDB files. |
Example call
npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit --skill evo2-nim --yes
Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| sequence | string | Required | Input DNA sequence, supported up to megabase-scale context. |
| task | stringdefault variant_effect | Optional | Analysis task to run: 'variant_effect' or 'generate'. |
Returns
| Field | Type | Description |
|---|---|---|
| predictions | object | Variant effect scores or generated sequence output. |
Example call
npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit --skill openfold3-nim --yes
Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| sequences | array | Required | One or more biomolecule sequences (protein/RNA/DNA) to co-fold. |
Returns
| Field | Type | Description |
|---|---|---|
| structure_files | array | Predicted complex structures with per-chain confidence scores. |
Example call
npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit --skill parabricks --yes
Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| fastq_input | string | Required | Path or URL to input FASTQ sequencing reads. |
| reference_genome | string | Required | Reference genome build to align against, e.g. GRCh38. |
Returns
| Field | Type | Description |
|---|---|---|
| vcf_output | string | Variant call file (VCF) with called variants. |
Example call
npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit --skill nvmolkit-usage --yes
Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| smiles_list | array | Required | List of candidate molecule SMILES strings to screen. |
| operation | stringdefault similarity_search | Optional | Operation to run: 'similarity_search', 'conformer_generation', or 'docking_prep'. |
Returns
| Field | Type | Description |
|---|---|---|
| ranked_results | array | Molecules ranked or filtered by the requested operation. |
NVIDIA BioNeMo Agent Toolkit is an open-source GitHub library of agent-callable skills for protein folding, docking, genomics and cheminformatics. It had 495 stars on 2026-10-23, installs with npx skills add, and works in Claude Code, Codex and other SKILL.md runtimes. Installing costs nothing, though compute does.
Broad model coverage
The catalog spans Boltz-2, DiffDock, Evo 2, GenMol, OpenFold3, RFdiffusion, Parabricks and nvMolKit, each as its own installable skill.
Three chained workflows
Protein binder design, generative virtual screening (GenMol, DiffDock, Boltz-2) and MSA-enabled structure prediction ship as ready-made multi-step skills.
One-command install
A single npx command installs the catalog, and flags let you choose one skill or target a specific agent such as Claude Code.
Agent-agnostic skill format
Every skill is a SKILL.md package, so any agent runtime that reads that format can use it, regardless of the backend LLM.
Open-model setup skills
Nine RFD4-Proteina skills plus KERMT, CodonFM and Proteina-Complexa set up models on your own hardware instead of calling a hosted NIM.
Vendor-reported skill lift
NVIDIA published its own with-and-without comparison at launch; the repository documents how to rerun a similar test against hosted endpoints.
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.
The nearest AI agent skills on Hokai for the same job.
| Skill | Works with | Connects by | Checked |
|---|---|---|---|
| Claude · Codex | Skill file | 23 Oct | |
| Claude · ChatGPT · Codex · Cursor +2 | MCP | 22 Nov | |
| -- | MCP | 17 Nov | |
| -- | Skill file | 27 Oct |
It is a GitHub library of agent-callable skills that lets an AI agent run NVIDIA life-science software: protein folding with Boltz-2 and OpenFold3, docking with DiffDock, DNA modelling with Evo 2, genomics with Parabricks and cheminformatics with nvMolKit. The repository also bundles three chained workflows and a newer group of open-model setup skills. NVIDIA presented it at its summer life-sciences launch and published its own with-and-without skill comparison.
Run npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit and pick skills interactively, or add --skill boltz2-nim --yes to install one. The --agent flag targets claude-code or codex, and --list browses the catalog without installing. Most NIM skills also need an API key from build.nvidia.com. The open-model skills instead need a separate model checkout and a local GPU.
The repository documents installs for Claude Code and Codex, and each skill is a SKILL.md package, so any runtime that reads that format can load it. The repository also ships plugin manifests so the toolkit installs through each agent's native plugin flow. NVIDIA says similar results hold across backend LLMs, but that is its statement and the repository publishes no per-model comparison.
Installing and reading the skills costs nothing, because the repository carries open licences for both code and skill text. What you pay for is compute. Hosted NIM skills call NVIDIA endpoints with an API key, and the repository states no price or rate limit for them, so read NVIDIA's current terms. Self-hosting a NIM or running the open-model skills needs your own GPU.
Accuracy comes from the underlying model, not the skill, so a specialist should still check predictions before anyone acts on them. Two OpenFold3-based entries are flagged by NVIDIA as awaiting validation, and deterministic evaluation tasks for local-GPU workflows are still a TODO. The headline lift figure is NVIDIA's own measurement.
NVIDIA BioNeMo Agent Toolkit is an open-source library of agent-callable skills from NVIDIA that packages its life sciences software, including BioNeMo, Parabricks, nvMolKit and RAPIDS-singlecell, so an AI agent can call it directly. It is not an app a scientist opens. Each skill is a directory with a SKILL.md file, optional reference material and optional scripts, and it tells the agent how to pick the right model, prepare valid inputs, run the job and explain the output across protein folding, docking, generative chemistry, genomics and protein design.
The repository's catalog on 2026-10-23 lists 21 product entries. NIM skills cover Boltz-2, DiffDock, Evo 2, GenMol, MolMIM, MSA-Search, OpenFold2, OpenFold3, ProteinMPNN, RFdiffusion and RFD4-Proteina. Three meta workflows chain them: protein binder design, generative virtual screening (GenMol, then DiffDock, then Boltz-2) and MSA-enabled structure prediction. Library skills wrap Parabricks, nvMolKit and cuEquivariance, and a newer open-models group (RFD4-Proteina setup, Proteina-Complexa, KERMT, CodonFM) sets up models that run on your own hardware. Install is one command, npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit, with --skill to pick one entry and --agent to target Claude Code or OpenAI Codex. The same skills are meant to work in any runtime that reads SKILL.md packages, the format Anthropic introduced for its agents.
NVIDIA announced the toolkit at BIO 2026 on June 23 and reported that agents without the skills completed 57.1% of required scientific tasks on average against 100% with them. That is NVIDIA's own figure, measured on its own task set, and the repository's evaluation section describes the method as running one prompt with and without a skill against the hosted NIM endpoint, graded by trajectory checks and an LLM judge. Native deterministic tasks for workflows that need a local GPU are listed as a TODO.
Who should look elsewhere: the skills raise how reliably an agent calls a model, not how accurate the model is, so low-confidence structures and generated molecules still need expert review. The repository itself says OpenFold3 and the MSA-Search to OpenFold3 workflow are pending validation after hosted runs timed out. The open-model skills need a separate model checkout and your own GPU. A team that wants a general connector library should compare Composio or build its own with the skill-creator skill, and anyone browsing alternatives can start from the skills directory.
Licensing is split. Scripts, tests and build tooling are Apache-2.0, skills and documentation are CC-BY-4.0, and the imported RFD4-Proteina Rosetta helpers are BSD-3-Clause. The GitHub repository showed 495 stars and 77 forks on 2026-10-23.
Protocol, authentication, permissions and the call contracts on this page are read from NVIDIA's own documentation. The verdict is Hokai's reading of them, not the maintainer's claim.