NVIDIA Agent Toolkit
NVIDIA Agent Toolkit for enterprise AI: build, run and govern agents on your own GPUs
Katonic AI is an enterprise agent platform built on the NVIDIA Agent Toolkit. Every agent built in Katonic compiles to NVIDIA's open agent software, so teams build, run and govern agents on GPUs they own without assembling the toolkit themselves.
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 the NVIDIA Agent Toolkit gives an enterprise
The NVIDIA Agent Toolkit is open-source models and software from NVIDIA for building AI agents: open models such as Nemotron, reference agents such as AI-Q, skills, runtimes, and the NeMo Agent Toolkit, the library for building, tracing, evaluating and serving agents. It gives a development team strong parts.
An enterprise needs those parts turned into a service that many teams can use safely. That is the work between the toolkit and production.
What Katonic adds on top
| Enterprise need | How Katonic meets it |
|---|---|
| People who are not developers must be able to build agents | Five ways to build, from plain language to code. Each one compiles to NVIDIA's open agent software |
| Agents already written must not be thrown away | Agents built on LangGraph, CrewAI, Google ADK or plain Python run under the same governance |
| Nothing reaches users untested | Evaluation runs before an agent is released |
| Every agent stays inside policy | NVIDIA NeMo Guardrails on inputs, outputs and tools |
| Some actions need a person to approve | Approval steps built on the human-in-the-loop support in the toolkit |
| Cost must be controlled | One gateway in front of the models, with budgets and caps per department |
| Auditors need a record | A trace of every run, kept inside your boundary |
| Data must not leave | Models served on your own GPUs with NVIDIA NIM or NVIDIA Dynamo. The platform can run with no outside connection |
Why "compiles to" matters
Many products say they work with NVIDIA software. Katonic makes NVIDIA's open agent software the output of every way of building an agent. Whoever built the agent, the result is a declarative definition the customer can read, move and inspect, instead of a format only one vendor understands. It also protects the customer from change underneath: NVIDIA deprecated the NeMo Agent Toolkit library in October 2026, and agents built in Katonic carry over because Katonic changes the compiler, not the agents.
Reported upstream
Katonic engineers reported four bugs in the NVIDIA NeMo Agent Toolkit, all four since fixed upstream. Two were in the Google ADK tool wrapper, one in how a function with no arguments is registered, and one in the LangGraph wrapper on Windows paths. The reports are public on the toolkit repository.
Guide to the toolkit
- What it is, and which platforms build on it.
- How the NVIDIA Agent Toolkit and the NeMo Agent Toolkit differ.
- How it compares with LangGraph, CrewAI, LangChain and Google ADK.
- How to run it on-premises and air-gapped.
- How to take toolkit agents to production with governance and cost control.
Read next
- NeMo Relay vs NeMo Fabric vs NeMo Helix →
- How to migrate from the NVIDIA NeMo Agent Toolkit →
- Can you build NVIDIA Agent Toolkit agents without writing code? →
- Is there a UI for the NVIDIA NeMo Agent Toolkit? →
- How to evaluate agents built with the NVIDIA NeMo Agent Toolkit →
- Can you run the NVIDIA Agent Toolkit on HPE Private Cloud AI? →
- Which sovereign AI platforms work with the NVIDIA Agent Toolkit? →
- What is the NVIDIA Agent Toolkit? →
- NVIDIA Agent Toolkit vs NeMo Agent Toolkit →
- Which platforms support the NVIDIA Agent Toolkit? →
- NeMo Agent Toolkit vs LangGraph, CrewAI and LangChain →
- Running the toolkit on-premises and air-gapped →
- Running toolkit agents in production →
- Issue 2161: Google ADK tool wrapper and deferred annotations →
- Issue 2183: Google ADK tool wrapper and optional parameters →
- Issue 2184: functions that take no arguments →
- Issue 2185: LangGraph wrapper and Windows paths →
- NVIDIA NeMo Agent Toolkit is deprecated: what replaces it? →
- NVIDIA migration guide for the NeMo Agent Toolkit →
Questions and answers
Is Katonic built on the NVIDIA Agent Toolkit?
Yes. Every agent built in Katonic compiles to NVIDIA's open agent software, and models are served with NVIDIA NIM or NVIDIA Dynamo on GPUs the customer owns.
What is the NVIDIA Agent Toolkit for enterprise AI?
Open-source models and software from NVIDIA for building AI agents. Enterprises use it through platforms and software vendors that build on it, or assemble it themselves.
Do we need developers to use the toolkit?
To use the toolkit directly, yes. Katonic adds a layer where a business user declares what an agent should know and do, and the platform produces the toolkit configuration.
Can we keep agents we built in LangGraph or CrewAI?
Yes. They run in Katonic under the same governance and audit trail.
Does this run on our own infrastructure?
Yes. Katonic runs on infrastructure you own, on Kubernetes or OpenShift, and can run with no outside connection.