NeMo Relay, Fabric, Helix
NeMo Relay vs NeMo Fabric vs NeMo Helix: what each one does
They do three different jobs and are meant to be used together. NVIDIA NeMo Relay sits inside an agent run and observes and controls each model and tool call. NVIDIA NeMo Fabric gives applications one interface for running different agents. NVIDIA NeMo Helix is a control plane that evaluates, optimizes and hardens agents. Together they replace the NeMo Agent Toolkit, which NVIDIA deprecated in October 2026.
Side by side
| NeMo Relay | NeMo Fabric | NeMo Helix | |
|---|---|---|---|
| One-line job | See and control every model and tool call | Run any supported agent through one interface | Evaluate, optimize and harden agents |
| What it is | A runtime library with middleware and plugins | An SDK with adapters for agent harnesses | A control plane with services, a command line and a web interface |
| How you use it | Attach it to an agent or instrument your own code | Describe a run in configuration and pick an adapter | Run it locally or on Kubernetes and call it by command line, API or SDK |
| Takes over from the toolkit | Tracing, telemetry, middleware, governance | Configuring and running agents | Evaluation, optimization, safety testing |
| Licence | Apache 2.0 | Apache 2.0 | Apache 2.0 |
NeMo Relay
- Wraps the model and tool calls of an agent so they can be observed and controlled without rewriting the agent.
- Captures model calls, tool calls, turns and sub-agents as one stream of events.
- Exports to OpenTelemetry and OpenInference, so it feeds the observability tool you already use.
- Middleware can block, redact or rewrite a call before it runs.
- Has integrations for agent frameworks and software kits for Python, Node.js and Rust.
- It is not a dashboard. It feeds one.
NeMo Fabric
- Gives applications, evaluation systems and rollout platforms one interface for running agents.
- A run is described once: models, tools, skills, Model Context Protocol servers and settings.
- Adapters translate that description for each agent harness, so the caller does not change when the agent does.
- Checks a configuration before it runs and reports what an adapter does not support.
- Returns results, errors and artefacts in one format so runs can be compared.
NeMo Helix
- An open-source control plane that runs on your own infrastructure, locally or on Kubernetes.
- Evaluates agents against benchmarks and test sets.
- Optimizes prompts and configuration, and supports fine-tuning of open models.
- Tests safety with guardrails and red-teaming tools.
- Reached through a command line, programming interfaces and a web interface.
- Uses NeMo Fabric and NeMo Relay to connect to agents.
Which do you need?
| If you want to | Start with |
|---|---|
| See what your agents do and what they cost | NeMo Relay |
| Enforce a policy on tool or model calls | NeMo Relay |
| Run several kinds of agent from one application | NeMo Fabric |
| Score an agent before release | NeMo Helix |
| Improve accuracy, speed or cost with evidence | NeMo Helix |
| Replace everything the NeMo Agent Toolkit did | All three |
What none of them is
These are tools for teams that build and operate agents. They do not give an organisation a builder for business users, budgets per department, approval before release, or a workspace where staff use approved agents. That remains the job of an agent platform. All three were pre-1.0 releases in November 2026, so their interfaces may change.
Where Katonic fits
Katonic is an agent platform that runs on infrastructure you own. 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.
Read next
- NVIDIA NeMo Agent Toolkit is deprecated: what replaces it? →
- How to migrate from the NVIDIA NeMo Agent Toolkit →
- NVIDIA Agent Toolkit for enterprise AI →
- NeMo Relay on GitHub →
- NeMo Fabric on GitHub →
- NeMo Helix on GitHub →
Questions and answers
What is NVIDIA NeMo Relay?
A library that wraps the model and tool calls of an agent so you can observe, control and tune them without rewriting the agent. It exports to OpenTelemetry and OpenInference.
What is NVIDIA NeMo Fabric?
A consistent interface for running agent harnesses and custom agents. Adapters handle the differences between agents.
What is NVIDIA NeMo Helix?
An open-source control plane for evaluating, optimizing and hardening production agents. It runs on your own infrastructure.
Do I need all three?
Only if you used the NeMo Agent Toolkit for tracing, running agents and evaluation. Otherwise choose the one that matches what you need.
Is NeMo Fabric the new NeMo Agent Toolkit?
It takes over the part that configures and runs agents. Tracing moved to NeMo Relay and evaluation to NeMo Helix.
Are they production ready?
All three were pre-1.0 releases in November 2026, and NVIDIA describes NeMo Relay as beta software. Pin versions and read the release notes before upgrading.