NVIDIA Agent Toolkit is NVIDIA's umbrella name for its collection of open models, agent blueprints, libraries, skills and runtimes for building and running AI agents; it is not the same thing as the NVIDIA NeMo Agent Toolkit, a single Apache-licensed Python library (formerly AgentIQ) that is one of the pieces NVIDIA and its partners place under the umbrella. NVIDIA announced the Agent Toolkit at its GTC conference on 16 March 2026, describing it as open source models and software for enterprises and developers building tools that scale productivity "by autonomously determining how to complete assigned tasks."[1] At launch NVIDIA named four kinds of content: open models such as Nemotron, open agents such as the AI-Q blueprint, open skills such as cuOpt, and open runtimes such as NVIDIA OpenShell.[1] Over the following months NVIDIA also placed the NVIDIA NemoClaw blueprints, CUDA-X libraries packaged as agent skills, NVIDIA Omniverse libraries, PhysicsNeMo and BioNeMo under the same label, and launched a separately named life-sciences offshoot, the BioNeMo Agent Toolkit, built with Agent Toolkit technologies.[5][6][8][10][11] There is no single "Agent Toolkit" package or repository: each component ships separately, under its own license, and several predate the name.[16][17][18][19][20]

Field	Value
Developer	NVIDIA
Type	Umbrella collection of open models, agent blueprints, libraries, skills and a secure runtime for AI agents
Announced	16 March 2026, GTC[1]
Launch components named by NVIDIA	Nemotron models, AI-Q blueprint, cuOpt skills, OpenShell runtime[1]
Components NVIDIA lists as of September 2026	Nemotron, BioNeMo, Omniverse, PhysicsNeMo, NeMo, OpenShell, AI-Q and NemoClaw[11]
License	No single license; components are licensed individually (see the licensing section below)
Distribution	build.nvidia.com, GitHub, Hugging Face and partner clouds; no single package[1][5][17][18]
Not to be confused with	NVIDIA NeMo Agent Toolkit (`nvidia-nat` Python library)

What NVIDIA means by the name

NVIDIA uses "Agent Toolkit" for a stack rather than a product with its own installer. In the 16 March 2026 GTC press release it wrote that "NVIDIA Agent Toolkit provides open source models and software for enterprises and developers building tools that scale productivity by autonomously determining how to complete assigned tasks," and that it "now includes NVIDIA OpenShell," an open source runtime enforcing policy-based security, network and privacy guardrails for autonomous agents, which NVIDIA also calls "claws."[1] An NVIDIA technical blog published the same day described the Agent Toolkit as providing "the full deployment stack" of models, tools, evaluation and runtimes for building, testing and optimizing long-running agents, and called OpenShell a runtime "which is part of the NVIDIA Agent Toolkit."[3]

Later NVIDIA material settled on a four-part description. In a 23 June 2026 blog post, Justin Boitano called the toolkit "an open foundation comprising models, tools, skills and a secure runtime for AI agents," and listed three building blocks: Nemotron open models for reasoning, NemoClaw blueprints for tools, skills and safer agent behavior, and the OpenShell runtime.[7] NVIDIA's agentic AI solutions page calls it "an open source platform for building digital coworkers that enterprises can customize and control."[12] The most complete list is in the FAQ on NVIDIA's NemoClaw product page, which defines the Agent Toolkit as the collection of NVIDIA open models, libraries, runtimes and blueprints, naming Nemotron, BioNeMo, Omniverse, PhysicsNeMo, NeMo, OpenShell, AI-Q and NemoClaw.[11]

NVIDIA's "harness" framing explains how the parts fit. In its 31 May 2026 GTC Taipei release, the company said autonomous agents "start with a model" and then need a harness, a software layer providing orchestration, context, memory, tool use and security. It positioned Nemotron models and NemoClaw blueprints as the pieces that connect popular harnesses, OpenShell as the layer that sets policy and privacy controls, and CUDA-X libraries as skills that agents can call.[5] The toolkit is meant to work with third-party harnesses; Boitano's post names Hermes Agent and OpenClaw.[7]

Components

The table lists only components that NVIDIA itself has placed under the Agent Toolkit name, with the date NVIDIA first did so in the sources reviewed.

Component	What it is	Placed under Agent Toolkit	License (as published)
Nemotron open models	NVIDIA's open-weight model family; supplies the reasoning layer	16 March 2026[1]	Per model. Nemotron 3 Super uses the NVIDIA Nemotron Open Model License; Nemotron 3 Ultra uses OpenMDW 1.1[22][23]
AI-Q Blueprint	Open reference agent for research over enterprise data; NVIDIA says its hybrid design uses frontier models for orchestration and Nemotron for research	16 March 2026[1]	Apache 2.0 (GitHub repository created February 2025)[19]
cuOpt	GPU-accelerated decision optimization library, offered as an agent skill	16 March 2026[1]	Apache 2.0[20]
NVIDIA OpenShell	Sandboxed runtime with policy controls over files, networks, credentials and tools	16 March 2026[1][3]	Apache 2.0[3][17]
NVIDIA NemoClaw	Open blueprints that package a model, an agent harness and OpenShell; launched as a stack for OpenClaw	Described as using Agent Toolkit software 16 March 2026; listed as a component by 31 May 2026[2][5]	Apache 2.0[18]
CUDA-X libraries as skills	cuDF, cuOpt, AI-Q, NeMo, PhysicsNeMo and CUDA-Q exposed to agents as domain skills	31 May 2026[5]	Per library; the NVIDIA skills repository is Apache 2.0[21]
NVIDIA NeMo	Tools for agent optimization, evaluation, governance and model customization[5]	Named in partner descriptions 16 March 2026; listed in NVIDIA's FAQ[1][11]	Per library
NVIDIA BioNeMo (via the BioNeMo Agent Toolkit)	Life-sciences tools and skills for agents	23 June 2026[6][11]	Per component
NVIDIA Omniverse libraries	ovrtx (RTX sensor simulation), ovphysx (GPU physics) and CAD-to-SimReady skills	20 July 2026[8]	Per library
PhysicsNeMo libraries and new and updated CUDA-X libraries	PhysicsNeMo re-architected into agent-friendly libraries, plus the new cuISS solver library, cuDSS and cuEST (quantum chemistry)	PhysicsNeMo exposed as a skill 31 May 2026; re-architected libraries and cuISS, cuDSS and cuEST added 26 July 2026[5][10]	Per library

NVIDIA's agentic AI page groups the same material into models (Nemotron and NVIDIA Cosmos, deployable as NVIDIA NIM microservices), tools and skills (NVIDIA verified skills, AI-Q, and "NVIDIA NeMo Agent Tools" for data generation, model customization and continuous improvement), blueprints (NemoClaw) and runtime (OpenShell).[12]

History

Date	Event
16 March 2026	NVIDIA announces the Agent Toolkit at GTC alongside OpenShell and NemoClaw, with 17 named software companies[1][2]
16 March 2026	LangChain announces an enterprise agent platform combining its tools with the Agent Toolkit[4]
31 May 2026	At GTC Taipei, NVIDIA adds NemoClaw blueprints and CUDA-X libraries as agent skills; OpenShell described as in early preview[5]
23 June 2026	NVIDIA announces the BioNeMo Agent Toolkit for life sciences; Boitano's blog restates the Agent Toolkit's scope[6][7]
20 July 2026	At SIGGRAPH, NVIDIA adds Omniverse libraries and promotes a local Agent Toolkit stack on DGX Station[8][9]
26 July 2026	NVIDIA adds PhysicsNeMo and new CUDA-X libraries for engineering agents[10]
28 September 2026	NVIDIA launches the Open Agent Safety Platform, with OpenShell as its software layer; the release does not use the Agent Toolkit name[13]

Launch at GTC 2026

NVIDIA announced the Agent Toolkit on 16 March 2026 in a press release titled "NVIDIA Ignites the Next Industrial Revolution in Knowledge Work With Open Agent Development Platform."[1] Jensen Huang framed it around agent adoption: "Claude Code and OpenClaw have sparked the agent inflection point," he said, adding that agents extend AI beyond generation and reasoning into action.[1] The release's news summary headlined three things: OpenShell as a new Agent Toolkit component, the AI-Q Blueprint, and a list of enterprise software adopters.[1] NVIDIA said it used AI-Q to develop the top-ranking agent on the DeepResearch Bench and DeepResearch Bench II leaderboards, and that the hybrid design "can cut query costs by more than 50%."[1] Both are NVIDIA's claims.

A separate release that day announced NemoClaw, a stack for the OpenClaw agent platform that installs Nemotron models and OpenShell in a single command. NVIDIA wrote that "NemoClaw uses NVIDIA Agent Toolkit software to optimize OpenClaw in a single command."[2] SiliconANGLE, which called the Agent Toolkit "another open-source project that Nvidia debuted today," described OpenShell and Nemotron as "the two core components of NemoClaw" that "also ship with" the Agent Toolkit.[24]

NVIDIA said developers could explore the Agent Toolkit and OpenShell on build.nvidia.com, running on inference providers and NVIDIA Cloud Partners including Baseten, CoreWeave, DeepInfra, Together AI and Vultr. OpenShell could also be downloaded from GitHub to run locally on GeForce RTX PCs, RTX workstations, DGX Station and DGX Spark.[1]

GTC Taipei expansion (May 2026)

On 31 May 2026, at GTC Taipei, NVIDIA announced "New NVIDIA Agent Toolkit software," naming NemoClaw blueprints, Nemotron models, the OpenShell secure runtime and CUDA-X libraries with agent skills.[5] NemoClaw was now described as blueprints for autonomous "digital coworkers" rather than only an OpenClaw add-on. Cadence, Dassault Systèmes, Siemens and Synopsys were named among the first to use it for engineering agents.[5] The same release introduced Nemotron 3 Ultra, which NVIDIA called a 550-billion-parameter mixture-of-experts model for long-running agents, and exposed cuDF, cuOpt, AI-Q, NeMo, PhysicsNeMo and CUDA-Q to agents as skills. NVIDIA said the verified skills were available in the Claude Code plug-in marketplace and the Hermes Skills Hub.[5] At that point NVIDIA described NemoClaw as available and OpenShell as "available in early preview."[5] SiliconANGLE called the announcement "the latest iteration of its Agent Toolkit."[26]

Domain extensions (June and July 2026)

On 23 June 2026 NVIDIA announced the NVIDIA BioNeMo Agent Toolkit, a separately named toolkit of life-sciences tools and skills. NVIDIA said it was built by "harnessing NVIDIA Agent Toolkit technologies" such as Nemotron models, the NeMo RL library and NemoClaw blueprints, with OpenShell as the execution environment.[6] The same day, Boitano's blog post pointed readers to both the Agent Toolkit and the BioNeMo Agent Toolkit.[7]

On 20 July 2026, at SIGGRAPH, NVIDIA announced that the Agent Toolkit "now includes NVIDIA Omniverse libraries," covering ovrtx for RTX sensor simulation, ovphysx for GPU-accelerated physics and CAD-to-SimReady skills for turning CAD data into simulation-ready OpenUSD assets.[8] The release described the Agent Toolkit as helping "software makers build AI agents that connect tools, skills and data sources."[8] An NVIDIA blog post that day described a local configuration in which the Agent Toolkit "brings together NVIDIA NemoClaw, the NVIDIA Nemotron 3 Ultra open model, NVIDIA Omniverse libraries as agent-accessible tools and skills, and a secure runtime in a single local system" on DGX Station.[9]

On 26 July 2026 NVIDIA announced "an expansion of NVIDIA Agent Toolkit for engineering," adding PhysicsNeMo, re-architected into agent-friendly libraries, and CUDA-X libraries including the new cuISS iterative sparse solver, cuDSS and cuEST for quantum chemistry.[10] Timothy Costa, NVIDIA's vice president and general manager of computational engineering, said: "With NVIDIA Agent Toolkit, developers can build agentic engineers that reason using physics, run complex simulations and generate high-fidelity data to become a new engine for innovation in chip and system design."[10]

Open Agent Safety Platform (September 2026)

On 28 September 2026 NVIDIA launched the NVIDIA Open Agent Safety Platform, made up of OpenShell open source software and the NVIDIA Sentry reference system design. The release said OpenShell is "Now broadly available," but it does not use the name "Agent Toolkit."[13] OpenShell therefore appears in two NVIDIA groupings as of September 2026: the Agent Toolkit collection for building agents, and the Open Agent Safety Platform for governing them. The NemoClaw product page lists OpenShell as an Agent Toolkit component and also links to the Open Agent Safety Platform.[11]

Relationship to the NeMo Agent Toolkit

The two names differ by one word and are easy to confuse.

NeMo Agent Toolkit is a specific open source Python library for connecting, profiling, evaluating and optimizing agents across frameworks, distributed on PyPI as nvidia-nat, and licensed under Apache 2.0. Its repository dates to March 2025.[15][16] It was released as AgentIQ, renamed the Agent Intelligence toolkit (AIQ), and on 11 June 2025 renamed the NVIDIA NeMo Agent toolkit. NVIDIA said at the time: "The core technologies, performance, and roadmap remain unchanged."[14]
NVIDIA Agent Toolkit is the 2026 umbrella described in this article. It has no package or repository of its own.[1][11]

The NeMo Agent Toolkit sits inside the umbrella. LangChain's 16 March 2026 announcement lists "NVIDIA Agent Toolkit, including NVIDIA Nemotron models, NVIDIA NeMo Agent Toolkit profiling and optimization, NVIDIA NIM microservices, and NVIDIA Dynamo."[4] NVIDIA's own component lists usually name the broader NVIDIA NeMo family rather than the library.[11][12] NVIDIA's GTC release uses both names separately, for example saying Synopsys is building with "NVIDIA Nemotron and Nemo Agent Toolkit" while other partners use "NVIDIA Agent Toolkit software."[1]

NVIDIA's own copy sometimes blurs the line. The SIGGRAPH blog says "NVIDIA Agent Toolkit also supports MCP integration, including an MCP client for connecting to remote MCP servers and an MCP server for publishing tools to any MCP client."[9] The NeMo Agent Toolkit documentation describes the same Model Context Protocol client and server features as capabilities of that library.[15]

Relationship to NemoClaw and OpenShell

NemoClaw and OpenShell are both components of the Agent Toolkit, at different layers. NVIDIA's NemoClaw FAQ describes NemoClaw as "the full agent deployment package" of models, harness, tools and runtime, and OpenShell as "the secure runtime inside it."[11] NVIDIA says each NemoClaw blueprint "includes NVIDIA Agent Toolkit components," naming Nemotron and other frontier models, NeMo for specialization and optimization, and OpenShell for runtime policy controls.[11] In short, the Agent Toolkit is the catalog, NemoClaw is a packaged reference build drawn from it, and OpenShell is the runtime that both NemoClaw and non-NVIDIA harnesses can run inside.[3][11]

Licensing and openness

NVIDIA markets the toolkit as open, and most of its software components are published on GitHub under Apache 2.0: OpenShell, NemoClaw, the AI-Q Blueprint, cuOpt, the NeMo Agent Toolkit and the NVIDIA skills repository.[16][17][18][19][20][21] The models use other terms. As published on Hugging Face, Nemotron 3 Super is released under the NVIDIA Nemotron Open Model License, while Nemotron 3 Ultra uses the OpenMDW 1.1 license.[22][23] The toolkit also sits on NVIDIA's CUDA software stack. VentureBeat observed that "The models are open, but they are optimized for Nvidia's CUDA libraries," and summed up the strategy as "open to all, owned by one."[25] Anyone adopting "the Agent Toolkit" is therefore accepting a set of separately licensed components, not one license.

Adoption

NVIDIA's GTC release named 16 software companies in its summary (Adobe, Atlassian, Amdocs, Box, Cadence, Cisco, Cohesity, CrowdStrike, Dassault Systèmes, IQVIA, Red Hat, SAP, Salesforce, Siemens, ServiceNow and Synopsys) and described Palantir's Nemotron work in the body, which is why coverage counted 17 adopters.[1][25] The commitments ranged from adoption to exploration:

Company	What NVIDIA's 16 March 2026 release says
Adobe	Will adopt Agent Toolkit software as a foundation for hybrid, long-running creativity, productivity and marketing agents[1]
Amdocs	Uses AI-Q and Nemotron for its Cognitive Core agent platform[1]
Atlassian	Working with Agent Toolkit software and OpenShell for its Rovo AI strategy[1]
Box	Using the Agent Toolkit for agents that run long business processes on the Box file system[1]
Cadence	Will use the Agent Toolkit and Nemotron with its ChipStack AI SuperAgent[1]
Cisco	Cisco AI Defense will provide AI security protection for OpenShell[1]
Cohesity	Adding OpenShell support and extending its Gaia platform with AI-Q[1]
CrowdStrike	Secure-by-Design AI Blueprint for agents built on AI-Q and OpenShell; uses Nemotron for detection and response agents[1]
Dassault Systèmes	Exploring Agent Toolkit software and Nemotron for its Virtual Companions[1]
IQVIA	Integrating Nemotron and other Agent Toolkit software with IQVIA.ai[1]
Palantir	Using Nemotron for agents on its sovereign AI Operating System Reference Architecture[1]
Red Hat	Integrating Agent Toolkit software into Red Hat AI Factory with NVIDIA[1]
Salesforce	Working with Agent Toolkit software including Nemotron for Agentforce, with a Slack-centered reference architecture[1]
SAP	Using Agent Toolkit software including NeMo for agents in Joule Studio[1]
ServiceNow	Autonomous Workforce of AI Specialists uses Agent Toolkit software, AI-Q, Nemotron and ServiceNow Apriel models[1]
Siemens	Fuse EDA AI Agent uses Nemotron[1]
Synopsys	Multi-agent framework built on AgentEngineer, using Nemotron and the NeMo Agent Toolkit[1]

LangChain said it would integrate Agent Toolkit components into its Deep Agents library, joined NVIDIA's Nemotron Coalition, and described AI-Q as "the flagship result of this collaboration."[1][4] Later releases added more names. ChipAgents is building chip-design agents with the toolkit, and Synopsys described using it with NIM, Nemotron, NeMo Gym and NemoClaw.[10] Omniverse-library adopters included SideFX, PTC, ForgeCAD, Lightwheel, Moonlake AI and Palatial.[8] NVIDIA said more than 50 companies were using the BioNeMo Agent Toolkit at its launch.[6]

Reception

VentureBeat treated the launch as NVIDIA moving from supplying hardware to owning the platform layer for enterprise agents. It argued that the free, open components are designed to generate demand for NVIDIA GPUs and compared the strategy to Google's approach with Android. It also cautioned that many partner statements used hedged language such as "exploring" and "working with," noting: "Adoption announcements are not deployment announcements."[25] VentureBeat added that OpenShell's security claims, "while architecturally sound, remain unproven at scale."[25] SiliconANGLE's GTC Taipei coverage described the toolkit as "a comprehensive suite of software, open-source models and blueprints" for long-running digital coworkers.[26]

See also

References

NVIDIA, "NVIDIA Ignites the Next Industrial Revolution in Knowledge Work With Open Agent Development Platform," NVIDIA Newsroom, 16 March 2026. https://nvidianews.nvidia.com/news/ai-agents
NVIDIA, "NVIDIA Announces NemoClaw for the OpenClaw Community," NVIDIA Newsroom, 16 March 2026. https://nvidianews.nvidia.com/news/nvidia-announces-nemoclaw
Ali Golshan, Alex Watson and John Myers, "Run Autonomous, Self-Evolving Agents More Safely with NVIDIA OpenShell," NVIDIA Technical Blog, 16 March 2026. https://developer.nvidia.com/blog/run-autonomous-self-evolving-agents-more-safely-with-nvidia-openshell/
LangChain, "LangChain Announces Enterprise Agentic AI Platform Built with NVIDIA," LangChain Blog, 16 March 2026. https://blog.langchain.com/nvidia-enterprise/
NVIDIA, "Enterprise Software Leaders Build AI Agents With NVIDIA," NVIDIA Newsroom, 31 May 2026. https://nvidianews.nvidia.com/news/enterprise-software-leaders-build-ai-agents-with-nvidia
NVIDIA, "NVIDIA Announces BioNeMo Agent Toolkit: Tools for Agents to Accelerate Scientific Discovery," NVIDIA Newsroom, 23 June 2026. https://nvidianews.nvidia.com/news/nvidia-launches-bionemo-agent-toolkit-giving-ai-agents-the-tools-to-accelerate-scientific-discovery
Justin Boitano, "How Businesses Are Building Specialized AI They Can Trust," NVIDIA Blog, 23 June 2026. https://blogs.nvidia.com/blog/nvidia-agent-toolkit-open-models-tools-skills-secure-runtime-ai-agents/
NVIDIA, "NVIDIA Agent Toolkit Expands With New Omniverse Libraries, Putting AI Agents to Work Building Simulation-Ready Worlds," NVIDIA Newsroom, 20 July 2026. https://nvidianews.nvidia.com/news/nvidia-agent-toolkit-expands-with-new-omniverse-libraries-putting-ai-agents-to-work-building-simulation-ready-worlds
NVIDIA, "At SIGGRAPH, NVIDIA Advances Graphics and Simulation With Agentic and Physical AI," NVIDIA Blog, 20 July 2026. https://blogs.nvidia.com/blog/siggraph-news-2026/
NVIDIA, "NVIDIA Expands NVIDIA Agent Toolkit With NVIDIA PhysicsNeMo and CUDA-X Libraries to Transform How the World Engineers, Designs and Builds," NVIDIA Newsroom, 26 July 2026. https://nvidianews.nvidia.com/news/nvidia-expands-nvidia-agent-toolkit-with-nvidia-physicsnemo-and-cuda-x-libraries-to-transform-how-the-world-engineers-designs-and-builds
NVIDIA, "NVIDIA NemoClaw" (product page and FAQ), accessed 28 September 2026. https://www.nvidia.com/en-us/ai/nemoclaw/
NVIDIA, "Agentic AI Solutions," accessed 28 September 2026. https://www.nvidia.com/en-us/solutions/ai/agentic-ai/
NVIDIA, "NVIDIA Launches Open Agent Safety Platform to Secure Agents From Testing to Deployment," NVIDIA Newsroom, 28 September 2026. https://nvidianews.nvidia.com/news/open-agent-safety-platform
NVIDIA, "NVIDIA Agent Intelligence toolkit is now the NVIDIA NeMo Agent toolkit," NVIDIA Developer Forums, 11 June 2025. https://forums.developer.nvidia.com/t/nvidia-agent-intelligence-toolkit-is-now-the-nvidia-nemo-agent-toolkit/335881
NVIDIA, "NVIDIA NeMo Agent Toolkit Overview," NVIDIA NeMo Agent Toolkit (1.8) documentation. https://docs.nvidia.com/nemo/agent-toolkit/latest/index.html
NVIDIA, "NeMo-Agent-Toolkit," GitHub. https://github.com/NVIDIA/NeMo-Agent-Toolkit
NVIDIA, "OpenShell," GitHub. https://github.com/NVIDIA/OpenShell
NVIDIA, "NemoClaw," GitHub. https://github.com/NVIDIA/NemoClaw
NVIDIA AI Blueprints, "aiq," GitHub. https://github.com/NVIDIA-AI-Blueprints/aiq
NVIDIA, "cuopt," GitHub. https://github.com/NVIDIA/cuopt
NVIDIA, "skills," GitHub. https://github.com/NVIDIA/skills
NVIDIA, "NVIDIA-Nemotron-3-Super-120B-A12B-BF16," Hugging Face model card. https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16
NVIDIA, "NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16," Hugging Face model card. https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16
Maria Deutscher, "Nvidia launches NemoClaw, Agent Toolkit to enhance AI agents," SiliconANGLE, 16 March 2026. https://siliconangle.com/2026/03/16/nvidia-launches-nemoclaw-agent-toolkit-enhance-ai-agents/
VentureBeat, "Nvidia launches enterprise AI agent platform with Adobe, Salesforce, SAP among 17 adopters at GTC 2026," VentureBeat, 2026. https://venturebeat.com/technology/nvidia-launches-enterprise-ai-agent-platform-with-adobe-salesforce-sap-among
Mike Wheatley, "Nvidia gives developers the tools to build secure, autonomous AI workers that scale," SiliconANGLE, 1 June 2026. https://siliconangle.com/2026/06/01/nvidia-gives-developers-tool-build-secure-autonomous-ai-workers-scale/

Editing: NVIDIA Agent Toolkit