NVIDIA/skills: 386K installs, signed GPU workflows, and one safe first test (2026)

NVIDIA/skills is worth a one-skill sandbox trial when your agent work touches CUDA, cuOpt, Jetson, NeMo, or another NVIDIA stack. The catalog page currently lists 368 skills and 386.1K total installs. NVIDIA's public repository shows about 3.1K stars and 582 commits. 1 2

Start with cuopt-numerical-optimization-api. It gives an agent product-specific rules for linear programming (LP), mixed-integer linear programming (MILP), and beta quadratic programming (QP), then lets you test one small optimization locally. NVIDIA's current weekly GitHub Trending snapshot did not verify a seven-day star delta for this repository, so the adoption signal here is directory installs plus current repository activity and official verification. 2 3

What the catalog changes

A normal coding-agent session starts with a broad request and asks the model to find the right product documentation, API shape, variable types, and failure modes. An NVIDIA skill packages those instructions in SKILL.md with supporting references, so the agent has a product-specific starting point when the task matches. NVIDIA describes the skills as portable instruction sets built on the open Agent Skills specification. 4

The catalog is a distribution mirror. NVIDIA says the skills are maintained in their product repositories and synchronized into NVIDIA/skills through an automated daily process. The catalog spans CUDA-X libraries, cuOpt, cuDF, cuPyNumeric, DeepStream, Holoscan, Jetson, NeMo, Omniverse, RAG, robotics, and other NVIDIA workflows. 1

NVIDIA's README gives installation examples for Claude Code, Codex, Snowflake CoCo, Cursor, and Kiro. The generic skills CLI also lists Cline with the agent identifier cline. That makes Cline a CLI target to verify in a disposable project; NVIDIA's README does not provide a Cline-specific installation example. 1 5

The verification story adds an integrity check to the distribution step. NVIDIA's process includes review, scanning, evaluation, a skill card, a detached skill.oms.sig signature, catalog publication, and synchronization. A valid signature shows that the installed directory matches the signed release. The signature itself is one part of a review process, so read the skill card and inspect the requested actions before granting a skill access to a real project. 4 6

Install one skill first

Use the current skills CLI. NVIDIA's README requires version 1.5.16 or newer and warns that version 1.5.15 and earlier can install a skill without making it visible in Claude Code. The @latest selector keeps the first command on the current CLI release. 1

Browse the catalog, then install only the cuOpt skill into one host:

npx skills@latest add nvidia/skills --list

npx skills@latest add nvidia/skills \
  --skill cuopt-numerical-optimization-api \
  --agent claude-code \
  --yes

Change claude-code to codex or cursor when that is your host. NVIDIA documents those agent-specific forms and allows repeated --agent flags for a multi-host install. 1

For a newly installed skill, start a new agent session. Claude Code can reload the skill in the current session with:

/reload-skills

Keep these management commands nearby:

npx skills update
npx skills list
npx skills check

update refreshes installed skills and flags upstream removals, renames, or merges. list shows the installed set, and check previews available updates. 1

Optional signature check

Use this check when the skill will enter a shared repository, a CI environment, or a machine with production credentials. Run it against the final installed directory:

pip install model-signing

curl -L -o nv-agent-root-cert.pem \
  https://raw.githubusercontent.com/NVIDIA/skills/main/nv-agent-root-cert.pem

model_signing verify certificate SKILL_DIR \
  --signature SKILL_DIR/skill.oms.sig \
  --certificate-chain nv-agent-root-cert.pem

The NVIDIA signing guide says the signature covers the skill directory's SKILL.md, scripts, references, assets, and supporting files. Strict verification catches files added after signing. 6

The repository's top-level license field says Apache-2.0, while the README separates CC BY 4.0 for documentation and skills from Apache 2.0 for source code. The selected cuOpt skill declares Apache-2.0 in its own front matter. Check the individual skill's metadata when redistribution or internal packaging matters. 1 7

First test: a small cuOpt LP

The cuOpt skill tells the agent to choose Python, C/C++, or CLI from the user's working interface. It also maps a linear objective with continuous variables to LP, integer or binary variables to MILP, and squared or cross terms with continuous variables to beta QP. The prompt below keeps the first run local, bounded, and inspectable. 7

Paste this into the agent after installing the skill:

I am testing cuOpt in a temporary project folder.

Solve this small shipping LP:
- x_a and x_b are continuous tonnes of two products.
- Minimize 4*x_a + 7*x_b.
- Demand: x_a + x_b >= 100.
- Capacity: 2*x_a + x_b <= 160.
- Bounds: 0 <= x_a <= 80 and 0 <= x_b <= 100.

Before writing code:
1. Explain why this is an LP rather than a MILP or QP.
2. Choose Python, C/C++, or CLI from the tools available in this project.
3. Check whether the cuOpt package and runtime are already available.

If the cuOpt runtime is unavailable or the interface is unclear, stop and report the exact blocker. Do not install system packages and do not silently switch to another solver.

If cuOpt is available, write the smallest local runnable example. Show:
- the model and variable bounds;
- solver status and objective value;
- x_a and x_b;
- the output file path and one rerun command.

Keep the test local. Do not modify credentials, remote machines, or unrelated project files.

A useful first result has four visible parts: the agent's LP classification, the chosen interface, a short runnable file, and a solver status with values you can inspect. A missing runtime is a successful boundary check when the agent reports it clearly and leaves the project unchanged.

The skill also supports AMPL, GAMS/GAMSPy, PuLP, JuMP, Pyomo, and CVXPY as modeling backends. If your project already uses one of those tools, ask the agent to preserve the existing model and point its solver configuration at cuOpt instead of porting the model immediately. 7

Second test: a read-only Jetson diagnostic

Use this path only when you have a network-reachable Jetson with JetPack and SSH access from the host running the coding agent. The independent Seeed Studio walkthrough uses NVIDIA skills with Claude Code for device checks, package selection, memory inspection, prototype generation, and demo packaging. 8

Install the diagnostic skill into Claude Code:

npx skills@latest add nvidia/skills \
  --skill jetson-diagnostic \
  --agent claude-code \
  --yes

Then begin with a read-only request:

/jetson-diagnostic Check whether this Jetson is ready for a real-time vision prototype and summarize any blockers.

Run read-only checks first. Report the Jetson model, JetPack/L4T version, memory pressure, thermal state, relevant services, and hardware endpoints. List any command that would change the device before running it. Stop after the diagnostic report.

The walkthrough's next steps use jetson-package for ARM-compatible dependency choices and jetson-memory-audit for memory pressure. A later prototype step uses deepstream-import-vision-model, but the article presents that workflow as exploratory: the output needs a clear input, output, rerun command, and known bottleneck before anyone treats it as a deployment. 8

Seeed's walkthrough uses a manual repository copy in its setup. The official npx skills route above is the current starting point because NVIDIA's README documents it as the catalog install path. 1 8

What users and benchmarks report

The hands-on Jetson account reports less repeated bring-up work around device checks, compatibility, installation, model deployment, output validation, and demo packaging. It also reports that a single workflow can chain several skills. Those are observations from one independent walkthrough, so use them as a workflow lead rather than a reliability score. 8

A public LinkedIn post by Moshe Abramovitch announced NVIDIA skills in Cursor and displayed 798 reactions and 23 comments when fetched. The post describes a catalog of more than 300 skills across more than 30 NVIDIA products. The post is a project-side announcement and engagement signal; the public page did not expose comment text for an independent review. 9

NVIDIA's SkillEvaluator compares agent runs with a skill installed against baseline runs for the same task. In its August 12, 2026 snapshot, NVIDIA reported these average scores across published skill-harness pairs:

Dimension	Baseline run	Skill-installed run	Reported lift
Correctness	46	87	+41 10
Discoverability	42	82	+40 10
Effectiveness	39	78	+39 10
Efficiency	43	78	+35 10
Security	97	98	+1 10

Treat those values as NVIDIA's benchmark scores for specialized tasks. The source does not identify the underlying model names or report confidence intervals. 85% of skills with published results used one attempt, while 15% used two. The catalog is evaluated continuously, so the snapshot can change. 10

The same report gives a useful cost boundary. cuopt-install rose from 25,227 to 55,582 tokens and from 34.0 to 41.1 seconds in the cited example after the skill was installed. Skills can improve task guidance while adding context and execution work. 10

The caveats that matter

The catalog is a live mirror, and its verification pipeline can delay individual items. NVIDIA pull request #476 records an automated sync in which numerous skills were reverted after content changed without refreshed signatures; two skills had internally mismatched signatures, one lacked a signature, and four Dynamo skills failed because their directories were empty or missing. The record describes a pipeline holdback, not a blanket failure of the catalog. 11

Jetson work adds hardware and architecture boundaries. Generic x86 installation instructions can fail on ARM, and a working prototype may still need calibration, stronger tracking, monitoring, a structured DeepStream pipeline, or TensorRT optimization before production use. 8

The catalog also separates ownership. NVIDIA's README directs product-specific skill issues to the corresponding upstream product repository, while catalog, synchronization, signing, and distribution issues belong in NVIDIA/skills. Read the skill card before installation, and use the individual skill's own license and dependency fields for deployment decisions. 1

Use it now or keep it in a sandbox

Install it now when all four conditions hold:

Your task uses an NVIDIA product covered by a named skill.
Your agent host is Claude Code, Codex, Cursor, or another host you have tested with the CLI.
You can start with a temporary project and a small, inspectable task.
You can review the generated files, commands, dependencies, and runtime output before sharing access.

Keep it in a sandbox when any of these conditions apply:

The first task touches production credentials, a remote device, or a deployment target.
The host is Cline or another generic CLI target and you have not verified how that host discovers installed skills.
The skill requires a private dependency, a hardware-specific runtime, or a package path you have not confirmed.
The installed directory fails strict signature verification or the sync state reports a missing or mismatched signature.
You want benchmark-level certainty about your own task; the published scores describe specialized evaluation cases, not your project.

The smallest useful action today is the cuOpt LP prompt above. It tests discovery, interface selection, runtime availability, generated code, and solver output while keeping the agent away from production systems.

Quick reference

Field	Pick details
Pick	`NVIDIA/skills`, starting with `cuopt-numerical-optimization-api` 1 7
Best for	NVIDIA product workflows that need product-specific instructions, references, and agent actions 4
Hosts	Claude Code, Codex, Snowflake CoCo, Cursor, and Kiro in NVIDIA's README; Cline appears in the generic CLI host list 1 5
Install	Run `npx skills@latest add nvidia/skills --skill cuopt-numerical-optimization-api --agent claude-code --yes` 1
Prerequisite	`skills` CLI 1.5.16 or newer; reload the session after installation 1
Catalog signal	368 listed skills and 386.1K total installs on skills.sh; GitHub shows about 3.1K stars and 582 commits 1 2
Maintainer	NVIDIA; the selected skill identifies the NVIDIA cuOpt Team as author 1 7
License	Selected cuOpt skill: Apache-2.0; repository README separates CC BY 4.0 for documentation and skills from Apache 2.0 for source code 1 7
First test	A local continuous-variable shipping LP; stop on missing cuOpt runtime and preserve the project 7
Main caveat	Catalog sync can hold back skills with stale, mismatched, or missing signatures 11