Agentic AI
Medical AI Skills
for Agentic Workflows
A verified catalog of agent-callable skills for medical imaging: DICOM utilities, CT and MR segmentation, synthetic data generation, and image reasoning, all built on MONAI workflows and NVIDIA MedTech models.
Overview
Trusted Building Blocks for Medical AI Agents
Each skill wraps one medical AI tool through its documented entry point, so agents and engineers can discover, invoke, chain, and reproduce it in their own environments. Skills are published only after passing verification and evaluation through a domain-aware evaluation engine.
Agent-Callable
Every skill ships a SKILL.md following the open Agent Skills specification, a machine-readable manifest, and scripts that emit structured JSON, ready for Claude Code and other agents.
Verified & Evidenced
A skill can exit successfully and still produce output you cannot trust. Manifests encode medtech invariants, every run yields a reproducible evidence pack, and paired verifiers audit domain quality.
Ecosystem-Native
Skills wrap MONAI-based workflows, MONAI bundles, DICOM utilities, and NVIDIA MedTech models, bringing the medical imaging ecosystem you already use into agentic pipelines.
Get Started
Add a Skill to Your Agent
Install any skill straight from the catalog with a single command; the CLI copies the skill into your agent's expected location. For the full trust harness with evidence packs, paired verifiers, and the gate ladder, clone the repository and use the Makefile targets.
- Pick a skill and target agent interactively, or pass flags for CI
- Your agent reads SKILL.md to learn when and how to run the tool
- Skills emit structured JSON, so outputs chain cleanly into the next step
# Interactive: pick a skill, pick an agent
npx skills add NVIDIA-Medtech/medical-AI-skills
# Non-interactive, e.g. for Claude Code
npx skills add NVIDIA-Medtech/medical-AI-skills \
--skill nv-segment-ct \
--agent claude-code --yes
# Full trust harness: clone and run with evidence
git clone https://github.com/NVIDIA-Medtech/medical-AI-skills
make run-skill SKILL=nv-segment-ctSkills follow the open agentskills.io specification, so they work with any compliant agent, not just one vendor's.
Catalog
Available Skills
From GPU-free DICOM preflight to whole-body segmentation and synthetic data generation. Each card links to the skill's folder with its SKILL.md, manifest, and fixtures.
dicom-series-preflight
GPU-free preflight of a DICOM series folder: corruption, orientation, PHI-tag presence, and series consistency checks before conversion or inference.
dicom-metadata-extract
Extract selected metadata from a DICOM file and flag standard-tag PHI presence.
dicom-series-to-volume
Convert a CT DICOM series folder to an HU-calibrated NIfTI volume with affine evidence.
nv-segment-ct
Run NV-Segment-CT (VISTA3D) whole-body segmentation on CT NIfTI volumes with label-map evidence.
nv-segment-ctmr
Run NV-Segment-CTMR segmentation on CT or MRI NIfTI volumes with label-map evidence.
nv-segment-ct-finetune
Auto-configuring VISTA3D continual-learning finetune on CT NIfTI labels via monai.bundle run.
nv-generate-ct-rflow
Generate synthetic CT volumes with paired 132-class masks using NV-Generate-CTMR rectified-flow synthesis.
nv-generate-mr
Generate synthetic body MRI volumes with NV-Generate-CTMR rflow-mr.
nv-generate-mr-brain
Generate synthetic brain MRI volumes with NV-Generate-CTMR rflow-mr-brain.
nv-generate-mr-brain-finetune
Finetune the NV-Generate-CTMR MR-brain diffusion UNet from a NIfTI datalist.
nv-generate-vae-finetune
Finetune the NV-Generate-CTMR MAISI VAE from CT or MRI NIfTI datalists.
nv-reason-cxr
Chest X-ray reasoning with NV-Reason-CXR-3B on a user-provided PNG or JPEG image.
Community
Get Involved
The catalog is Apache-2.0 and open to contributions from across the MONAI ecosystem: wrap a tool you rely on, improve the docs, or strengthen the evaluation harness. Because skills follow an open specification, you can also publish your own skill catalogs for the community.
Contribute a Skill
Wrap a new tool, add a paired verifier, or extend the harness. Contribution lanes, required proof, and DCO sign-off are all documented.
Documentation
Guides for using and authoring skills, trust and evidence, agent task maps, and skill-vs-readme comparison studies.
Eval Engine
The evidence-pack harness: a 10-step gate ladder, reproducible runs with environment locks, drift comparison, and replay.
Verifiers & Benchmarks
Skill-shaped auditors that check domain invariants like anatomy plausibility, geometry, HU ranges, and PHI scope, plus dataset benchmark manifests.
Medical AI Skills is maintained by the NVIDIA MedTech team under the Apache-2.0 license. It is a tools catalog for engineering use, not a clinical, diagnostic, or regulatory tool, and generated outputs are engineering artifacts, not clinical endorsements. Downloaded model weights carry their own licenses.