NVIDIA Agent Toolkit
What is the NVIDIA Agent Toolkit, and which platforms support it?
The NVIDIA Agent Toolkit is a set of open-source models and software from NVIDIA for building AI agents, announced at GTC in March 2026. It includes open models such as Nemotron, open agent blueprints such as AI-Q, open skills and open runtimes such as OpenShell, with the NeMo Agent Toolkit as the library for building, tracing, evaluating and serving agents. Katonic AI is built on it: every agent built in Katonic compiles to NVIDIA's open agent software.
Update, November 2026
NVIDIA has deprecated the NeMo Agent Toolkit library as of October 2026, with end of life at the end of 2026, and names NeMo Relay, NeMo Fabric and NeMo Helix as its successors. The NVIDIA Agent Toolkit announced in March 2026 continues. Every agent built in Katonic is a declarative definition that compiles to NVIDIA's open agent software. Today that software is the NeMo Agent Toolkit. When NVIDIA changes the runtime, Katonic changes the compiler and your agents carry over.
What it includes
| Part | What it is | Example |
|---|---|---|
| Open models | Language models you can host yourself | NVIDIA Nemotron |
| Open agents | Reference agents, published as blueprints | AI-Q, for research over company knowledge |
| Open skills | Capabilities an agent can call | cuOpt, for optimisation |
| Open runtimes | Where agents run under policy | OpenShell |
| Agent library | The code library for building, tracing, evaluating and serving agents | NeMo Agent Toolkit |
NVIDIA describes the toolkit as open source models and software for enterprises and developers. The parts can be used together or separately.
Agent Toolkit and NeMo Agent Toolkit
The two names are easy to confuse. The NVIDIA Agent Toolkit is the whole collection. The NeMo Agent Toolkit is one part of it: the library a developer uses to define an agent, connect its tools, trace what it does, evaluate it and expose it as a service. It works with agents written in common open-source frameworks, so a team does not have to rewrite what it already has.
Which platforms support it
Three groups build on it.
- Enterprise software companies. At the launch NVIDIA named Adobe, Atlassian, Box, Cisco, CrowdStrike, Palantir, Red Hat, Salesforce, SAP, ServiceNow, Siemens and others as building agents with it.
- Agent frameworks. The NeMo Agent Toolkit works with open-source frameworks including LangChain and LangGraph, LlamaIndex, CrewAI, Semantic Kernel and Google ADK. LangChain is integrating the toolkit into its own agent library.
- Agent platforms that install on your own infrastructure. Katonic AI uses the NeMo Agent Toolkit as its agent runtime.
What it leaves to you
The toolkit gives you the parts for building an agent. An organisation that wants many teams to run agents in production still has to supply the rest:
- GPU infrastructure and model serving for the models you choose.
- One gateway in front of the models, with access, budgets and logging.
- Your documents and data, connected with permissions intact.
- Guardrails and a test gate that every agent passes before release.
- Sign-in, roles and an audit trail across every team.
- Somewhere for people who are not developers to build and use agents.
How Katonic uses it
Katonic is built on the NVIDIA stack. Every agent a customer builds in Katonic, whichever way it was built, compiles to NVIDIA's open agent software. That gives the customer a portable, inspectable artefact instead of a format only one vendor can read.
- Five ways to build, from plain language to code, and one output.
- Agents already written in LangGraph, CrewAI, Google ADK or plain Python run under the same governance and audit trail.
- Evaluation runs before an agent is released.
- NVIDIA NeMo Guardrails apply to what goes in and what comes out.
- Models are served on your own GPUs with NVIDIA NIM or NVIDIA Dynamo.
- The whole platform runs on infrastructure you own and can run with no outside connection.
Read next
- NVIDIA Agent Toolkit for enterprise AI →
- Running NVIDIA Agent Toolkit agents in production →
- Katonic on the NVIDIA stack →
- From served model to working agent →
- Use case: moving agent pilots to production →
- NVIDIA announcement (nvidia.com) →
- NVIDIA NeMo Agent Toolkit is deprecated: what replaces it? →
- NVIDIA migration guide for the NeMo Agent Toolkit →
Questions and answers
Is the NVIDIA Agent Toolkit open source?
Yes. NVIDIA describes it as open source models and software for enterprises and developers.
Is the NVIDIA Agent Toolkit the same as the NeMo Agent Toolkit?
No. The NVIDIA Agent Toolkit is the whole collection of models, agents, skills and runtimes. The NeMo Agent Toolkit is the library inside it for building, tracing, evaluating and serving agents.
Does it replace LangGraph or CrewAI?
No. The NeMo Agent Toolkit works with agents written in those frameworks and adds tracing, evaluation and serving around them.
Do I need NVIDIA GPUs to use it?
The software is open source. In practice it is designed to run with NVIDIA models and serving software on NVIDIA GPUs.
Which platforms are built on the NVIDIA Agent Toolkit?
NVIDIA named enterprise software companies including Adobe, Atlassian, Salesforce, SAP and ServiceNow at the launch. Katonic AI, a platform for running agents on infrastructure you own, uses the NeMo Agent Toolkit as its agent runtime.
Can it run air-gapped?
The open-source parts can be hosted on your own infrastructure. Whether a deployment runs with no outside connection depends on the platform around it. Katonic can run with no outside connection.