Download skills/bionemo-parabricks/SKILL.md from the publisher, or copy the path and open the repository. PlugMCP lists the skill and does not host the file.
Where does parabricks live?
The publisher keeps it at https://github.com/NVIDIA/skills.
Which agents can use parabricks?
Claude, Cursor, and Codex can read a SKILL.md you add from the publisher's repository.
>-
name: parabricks
description: >-
Plan NVIDIA Parabricks pbrun commands, translate CPU options and check runtime
readiness. Use for individual tools, including planning-only questions.
license: CC-BY-4.0 AND Apache-2.0
metadata:
author: "Angel Pizarro [email protected]"
version: "1.3.0"
tags:
- parabricks
- genomics
- nvidia
Parabricks
Purpose
Use this skill to discover the right NVIDIA Parabricks pbrun command, assess
runtime readiness, and generate version-aware command guidance for individual
tools.
Do not use this skill for whole-workflow inspection, acceleration planning,
or wiring optional GPU branches. For pipeline-level work, use
genomics-workflow-acceleration.
When to Use This Skill
Which pbrun tool fits the user's data and goal
GPU, driver, Docker, container, storage, or installation readiness
Command shape, flags, and validation for a specific Parabricks tool
Reference/index compatibility and CPU-to-Parabricks option translation
Troubleshooting a single Parabricks command or tool family
Planning-only or "do not run" questions about a named pbrun tool (for
example reference/index compatibility, output naming, or flag validity). Use
this skill's references to answer; do not skip it because nothing will be
executed.
Prerequisites
Command guidance: No Parabricks/NGC credential or local GPU/container
runtime is required to read references or generate commands.
Readiness helper: Requires Bash and GNU coreutils timeout (or gtimeout)
on the host being inspected. Missing runtime components are reported as
readiness gaps; without a timeout utility it exits before running probes.
STAR translation helper: Requires Python 3.7 or newer, using only the
standard library. It does not require a GPU or a Parabricks installation.
Execution: The documented Docker commands require a supported Linux host,
an NVIDIA GPU with sufficient memory, a compatible NVIDIA driver, Docker,
NVIDIA Container Toolkit, and access to the selected Parabricks image.
Pulling the image may require an NGC API key for authentication to nvcr.io.
See runtime-environment.md for
version-specific requirements and checks.
Tool scope: read this skill's references and user-supplied manifests; write only
requested command artifacts. Use the bundled helpers for translation or an
authorized host inspection. Documentation lookups use docs.nvidia.com; image
access uses the selected registry only for authorized container execution.
These instructions do not grant tool permissions; the host's permission policy
still applies to shell, file, network and optional documentation MCP access.
Inputs
Use the user's explicit arguments and supplied manifests as the source of
truth; use prior context only where it does not conflict. Read a named manifest
before translating it. For command guidance, identify the assay, input paths,
reference/index provenance, outputs and target release. Ask only for missing
facts that change the answer; label unresolved template values as placeholders.
Required inputs are the selected tool's read/reference and output arguments.
Optional inputs include intervals, resource bundles, logs and tuning parameters
when supported by that tool; do not invent them to fill a template.
Runtime-only questions do not need sample metadata. Supplied deployment paths
are not permission to probe local data or hardware.
If the user is unsure which tool applies, scan the Tool Reference Index
below, then load the matching references/pbrun-<tool>.md file.
Limitations
This skill routes and guides Parabricks commands. It does not install
Parabricks, infer missing sample metadata, guarantee output parity, provide
clinical interpretation, or promise exact runtime without benchmark data.
Instructions
Identify the requested release. The option references cover v4.7.1; examples
from another release are labeled. A software version and image tag can
differ: NVIDIA's 4.7.1 image is tagged 4.7.1-1. Preserve the user's verified
image selection; use command-conventions.md
for the shared launch and validation contract.
Specific command → matching references/pbrun-<tool>.md. Read it
before any web search or MCP lookup; use those only to verify
version-sensitive details the reference does not settle.
Combine the tool-specific mapping with shared-options.md
rows whose Tools column names the selected command. For STAR argument
arrays or multi-lane batches,
use the local translation helper; it preserves
tokens and reports unmapped semantics without executing a workload.
Collect missing biological and filesystem context before generating commands.
Return the requested command or artifact with unresolved assumptions. For
planning-only work, check command text against the references and supplied
facts. Filesystem, container and output checks apply only to an authorized
execution. Do not turn command preparation into a readiness inspection.
Available Scripts
Run these from the skill directory containing this file; paths are relative to
that directory. See the helper prerequisites above.
Script
Purpose
Arguments
scripts/check_parabricks_runtime.sh
Inspect the execution host; report gaps without installing anything
--path DIR (repeatable), --format text|json, --timeout SECONDS; container probes require --run-container-check or --run-parabricks-check with --parabricks-version TAG / --parabricks-image IMAGE
scripts/translate_star_to_parabricks.py
Convert explicit STAR argument arrays to reviewed command drafts
For explanations, give the decision, complete command, and remaining checks.
For file requests, write the specified path and schema, then read back and
validate the artifact against the input manifest. Report all job IDs, omitted
options and unresolved output contracts. Do not replace supplied values with
illustrative examples. The runtime helper's JSON has schema_version, checks
and readiness results; the translator emits status, draft_argv, command,
notes, issues and untranslated for each request.
Tool Reference Index
If the command is named, open references/pbrun-<command>.md directly.
For tool selection and official manual links, use tool-index.md.
Common routes:
Prepare the request using the translation schema.
Review issues and untranslated before using the draft. Tool references show
the pbrun portion; the shared container envelope is documented once in
command-conventions.md.
Troubleshooting
Error
Cause
Solution
Multiple plausible tools
Data type or goal underspecified
Ask for assay, inputs, caller preference, desired output; use Tool Reference Index
Exact flag requested
Options are version-sensitive
Check the selected tool reference, NVIDIA docs, or pbrun <tool> --help
Runtime question
GPU, Docker, drivers, or storage
Use runtime-environment reference and diagnostic script
Wrong tool family
Assay or input type unclear
Confirm DNA/RNA/methylation/long-read/pangenome before routing
CUDA or memory failure
Runtime not ready or GPU memory constrained
Assess runtime before tuning command flags
Documentation MCP unavailable
Connection, timeout or server error
Use the bundled release reference first; for an unresolved fact, fall back to the official versioned manual and mark any remaining uncertainty
Guardrails
Treat command availability and options as version-sensitive.
Do not infer exact flags from command names alone.
Do not collapse standalone tools and full pipelines when explaining tradeoffs.
Do not substitute DNA fq2bam for RNA, or germline for somatic callers.
Do not invent sample names, read groups, reference builds, known-sites files,
model files, graph resources, container tags, or output paths.
Do not install, upgrade, or modify packages. Label setup commands as user-run.
Do not claim CPU execution of Parabricks tools.
Do not claim biological or VCF parity without a comparison run.
Use the optional documentation MCP server only for details the bundled
references do not settle. If its connection fails, avoid repeated discovery
or lookup loops; use the official manual. Never claim a failed lookup verified
an option. Documentation access does not require genomic data or credentials.
Download skills/bionemo-parabricks/SKILL.md from the publisher, or copy the path and open the repository. PlugMCP lists the skill and does not host the file.
Where does parabricks live?
The publisher keeps it at https://github.com/NVIDIA/skills.
Which agents can use parabricks?
Claude, Cursor, and Codex can read a SKILL.md you add from the publisher's repository.