NVIDIA: 415 Signed Agent Skills Now Ship for Claude Code and Codex

By Indie Kings | October 17, 2026

Updated October 17, 2026: NVIDIA now publishes an official, verified catalog of agent skills for coding assistants, and Anthropic maintains the official plugin directory for Claude Code. Together they move agent skills from a community pattern to vendor-backed, cryptographically signed distribution, with NVIDIA's own bundle already flowing into both channels.

The one sentence a friend would repeat: agent skills just grew up, because NVIDIA is now shipping a catalog of hundreds of signed, verified skills that teach Claude Code, Codex, and other coding agents how to drive its software, while Anthropic runs the curated directory that installs plugins, including NVIDIA's bundle, straight into Claude Code.

NVIDIA's verified skills catalog

The NVIDIA Agent Skills repository describes itself, per the official README, as "Official, NVIDIA-verified Agent Skills for Claude Code, Codex, and other coding agents." Skills are portable instruction sets that teach AI agents how to use NVIDIA software: physical AI and robotics workflows, simulation, CUDA-X libraries, retrieval pipelines, and platform tools. The catalog repository mirrors skills from their product repos daily through an automated sync pipeline, so the catalog grows continuously rather than in scheduled batches.

As of October 17, 2026, the README's skill catalog table lists 53 NVIDIA products and 415 unique skill names, from optimization libraries to Jetson board support packages to medical imaging suites. Documentation for the catalog lives at docs.nvidia.com/skills, which NVIDIA names as the official docs host in the README. A companion NVIDIA developer blog post, referenced from the README under the title "NVIDIA Verified Agent Skills: Capability Governance for AI Agents," frames the whole effort as governance for what agents are allowed to do, not just a bigger pile of prompts.

Licensing is explicit. The README carries an SPDX header reading "Apache-2.0 AND CC-BY-4.0" with the line "Copyright (c) 2026 NVIDIA Corporation." Per the license section, source code is under Apache 2.0 and the documentation and skills themselves are under CC-BY-4.0, in NVIDIA's own words. One framing sentence in our own words: agent skills extend what AI coding assistants can do reliably, in a repeatable and governed way; NVIDIA's contribution here is treating that reliability as a distribution problem, the way it treats driver signing.

What the skills actually do

The catalog spans consumer and enterprise NVIDIA software alike. Five examples from the README's catalog table give the flavor: (per the NVIDIA skills README)

• rtx-remix-modding, listed under RTX Remix, lets an agent mod and remaster classic games: manage projects and USD layers and replace textures and models through the Toolkit MCP server.
• nvidia-app, listed under NVIDIA App, checks drivers and manages and optimizes games over an MCP bridge, including laptop features.
• nvidia-broadcast applies audio and video effects, processes local media, and changes camera resolution through Broadcast's MCP gateway.
• g-assist-mcp-skill, under G-Assist, checks or changes resolution, refresh rate, V-Sync, G-SYNC, brightness, color, and GPU performance on the machine.
• cuopt-numerical-optimization-api guides an agent through the cuOpt Python API, the quickstart example being a request to solve a linear programming problem.

On the enterprise side, the same table lists skills for BioNeMo drug discovery pipelines, DeepStream video analytics development, Jetson board support package builds that prepare, customize, and flash images, and diagnostics on live Jetson devices after boot. NVIDIA even ships a skill about skills: skill-card-generator reads a skill's source files and produces a skill card plus a review table for governance purposes.

Signing, skill cards, and the governance pipeline

The verification machinery is the headline. Per the README's Verifying Skills section, every published skill ships with a detached OMS signature, verifiable against NVIDIA's trust anchor certificate, and the sync pipeline drops any skill missing its required artifacts before it ever reaches the catalog. Each catalog skill carries four things: the SKILL.md instructions the agent consumes, a skill-card.md with identity and governance metadata, the detached OMS signature file, and a Tier 3 evaluation dataset, with a benchmark report generated alongside it so consumers can see measured uplift data rather than a vendor claim.

NVIDIA's roadmap section marks its major governance milestones as completed: a public catalog of verified skills, daily automated sync, security scanning covering instruction safety and supply chain integrity, signing for every published skill, universal and task specific evaluation criteria, machine readable skill cards, sync time compliance gates including signature drift detection, and syndication to external marketplaces. The only open roadmap item, per the README, is syndication to additional MCP hubs and partner channels. A successful signature verification confirms, in the README's words, that skill contents have not been modified since signing by NVIDIA.

The catalog also records how a skill is expected to behave, not just what it does. The skill card carries, per the README's repository description of skill cards, machine readable metadata for identity, provenance, quality, and behavioral boundaries. Behavioral boundaries published as metadata is the part most directly aimed at the trust problem that has shadowed agent tooling: an agent reading arbitrary instructions from the internet has always been a supply chain risk, and NVIDIA's answer is to make every instruction set a signed, evaluated, revocable artifact.

Anthropic's official plugin directory

Anthropic runs the other half of the ecosystem. Per the official README, Claude Code Plugins Directory is "A curated directory of high-quality plugins for Claude Code." Its structure has two halves: an internal directory of plugins developed and maintained by Anthropic's team, and an external directory of third party plugins from partners and the community. External submissions must meet quality and security standards for approval before inclusion, and partners submit through Anthropic's plugin directory submission form.

Installation runs entirely inside Claude Code: as described per the README, users install a plugin by name with a single plugin install command, or browse the catalog in the plugin menu's Discover tab. A plugin can bundle slash commands, agent definitions, and skills, and a skill bundle entry lets a marketplace include an upstream repo's skill files directly even when that repo ships skills without a plugin manifest of its own. Names are immutable slugs once published, because renaming breaks existing installs, per the README.

Anthropic's own caution sits at the top of the README, and it is worth quoting honestly in paraphrase: the README tells users to make sure they trust a plugin before installing, updating, or using it, and states that Anthropic does not control what MCP servers, files, or other software plugins include and cannot verify that they work as intended or that they will not change. In other words, the directory answers the trust problem with curation and disclosure, while NVIDIA's catalog answers it with cryptography. Both maintainers are, in effect, publishing the trust documentation they wished the ecosystem already had.

Where the two pieces fit together

The two announcements connect, and the connection is structural. NVIDIA's catalog adheres to the Agent Skills specification published at agentskills.io, a spec NVIDIA's README names directly and which defines skills as portable directories with a SKILL.md file, YAML frontmatter metadata with required name and description fields, and a progressive disclosure model where lightweight metadata loads at startup and full instructions load on activation. Anthropic co-produced that spec, and Anthropic's platform now operates the ecosystem directory where skills and plugins reach end users, with plugin names registered as immutable slugs and skill bundles addressed as plugin prefixed skills.

NVIDIA's own README shows the loop closing inside Anthropic's channel: the repository brands itself with an Agent Skills Specification badge pointing at agentskills.io, carries Claude Code marketplace metadata alongside agent marketplace metadata for other clients, packages a curated NVIDIA skills bundle for Claude Code and Codex, and marks syndication to Skills.sh, the Codex plugin channel, the Claude Code plugin channel, ClawHub, and Hermes Hub as completed roadmap items. One vendor therefore publishes a governed catalog under a spec produced with the marketplace operator, and the marketplace's own guardrails, the trust caution, the immutable slugs, the quality standards for external submissions, provide the ecosystem side of the same trust story. Agent skills have graduated from a pattern hobbyists copied between configs into shipped, provisioned capability surfaces with root certificates and review pipelines attached.

The two catalogs at a glance

FAQ

What is an agent skill?
A portable instruction set, per the Agent Skills specification, packaged as a directory with a SKILL.md file that teaches an AI coding agent how to use a specific tool or workflow. Metadata loads lightly at startup and full instructions load when the agent hits a relevant task.

Are NVIDIA's agent skills free to use?
Per the README, the catalog is dual licensed: source code under Apache 2.0 and documentation plus skills under CC-BY-4.0, both with the 2026 NVIDIA copyright. That makes the catalog open to read, reuse, and contribute to under those license terms.

Which coding agents can use NVIDIA's skills?
NVIDIA's install flow targets Claude Code, Codex, Snowflake CoCo, Cursor, and Kiro, with a full client list maintained by the open skills CLI the catalog installs through. Because skills follow the portable Agent Skills spec, any compliant agent can consume them.

How do I know an NVIDIA skill has not been tampered with?
Every published skill ships a detached OMS signature verifiable against NVIDIA's published root certificate, per the README. A successful verification confirms the skill contents have not been modified since NVIDIA signed them, and the sync pipeline rejects skills missing signatures or eval artifacts.

What is the difference between a skill and a Claude Code plugin?
A skill is one instruction directory; a plugin is a package that can bundle skills together with slash commands, agent definitions, and MCP server configuration, per Anthropic's README. Anthropic's marketplace can also register an upstream repo's skills directly as a skill bundle plugin.

Is it safe to install third party plugins from Anthropic's directory?
Anthropic's README tells users to trust a plugin before installing, updating, or using it and says Anthropic does not control what MCP servers, files, or other software plugins include and cannot verify they will work as intended or stay unchanged. External submissions must meet quality and security standards before approval.

Bottom Line

Agent skills crossed from a community habit into shipped, governed infrastructure this month, and the two moves that made it real come from the same design partnership. NVIDIA's catalog gives coding agents signed, evaluated, minimally 415 strong verified instructions for its software stacks, from RTX Remix modding to Jetson flashing to drug discovery pipelines, every one carrying an OMS signature against NVIDIA's root certificate. Anthropic's directory gives Claude Code users a curated place to install plugins, including NVIDIA's bundle, with the maintainers' own trust documentation printed where people will actually read it. The equipment question for agent tooling is changing from "what can it do" to "who vouches for it," and both of these catalogs exist to answer the second question first.

Sources:

• NVIDIA Agent Skills repository README (NVIDIA)
• Claude Code Plugins Directory README (Anthropic)

Image: NVIDIA Agent Skills repository card. Credit: NVIDIA.