NeMo Helix, Relay, Fabric
NVIDIA NeMo Helix, Relay and Fabric: a field guide from a hands-on test
NVIDIA NeMo Relay, NeMo Fabric and NeMo Helix are three open-source projects, all under the Apache 2.0 licence, that replace the NeMo Agent Toolkit NVIDIA deprecated in October 2026. Relay observes and controls model and tool calls inside an agent. Fabric runs different agents through one interface. Helix, with its NeMo Studio screens, evaluates, fine-tunes and improves them. Katonic installed and tested all three on 14 and 15 November 2026, and this page collects what we found.
The three projects at a glance
| NeMo Relay | NeMo Fabric | NeMo Helix | |
|---|---|---|---|
| Job | Observe and control model and tool calls | Run agents through one interface | Evaluate, fine-tune and improve agents |
| Shape | A library with a Rust runtime and Python, Node and Rust SDKs | A runtime plus adapters, one per agent harness | A control plane with a web UI called NeMo Studio |
| Version we tested | 0.11.1 | 0.6.1 | 0.9.1, and the main branch |
| Repository | NVIDIA/NeMo-Relay | NVIDIA/NeMo-Fabric | NVIDIA-NeMo/nemo-helix |
| Our experience | Worked first time on a LangGraph agent | Worked once the Python environment was activated | Useful once installed; the install needed five fixes |
How they fit together
- An application asks Fabric to run an agent. Fabric starts the right adapter for that agent.
- Inside the run, Relay records each model call and tool call, and can apply guardrails, redaction and cost tracking.
- Relay sends the record of the run to Helix, where it shows up as a trace you can inspect, score and compare.
- Helix bundles its own copies of Fabric and Relay, so installing Helix gives you all three.
How we tested
We used one clean Ubuntu 24.04 virtual machine with four CPU cores and no GPU, and called hosted models through an API. We installed Helix the way the documentation recommends, ran its sample agent, ran an evaluation and an experiment, and looked at every NeMo Studio screen. We then installed Relay and Fabric on their own in a separate environment and ran NVIDIA's examples, a LangGraph agent with Relay attached, and a small adapter of our own.
What worked well
- Relay attached to a LangChain and LangGraph agent with a few lines and no fixes. Tracing, a tool guardrail, redaction and cost tracking all ran.
- Fabric ran NVIDIA's custom LangGraph example in every mode we tried: plan, run, with a tool server, with Relay, and as a two-message session.
- Helix's evaluation and trace tooling is the strongest part of the stack. An experiment comparing models on a sample dataset ran end to end.
- Fabric supports token streaming as an optional adapter capability. A small test adapter streamed 33 chunks through it.
What to know before you start
- All three were before version 1.0 in November 2026. Pin versions.
- NeMo Studio has no visual agent builder. Agents are configuration files, registered from the command line or by a coding assistant.
- Model deployment in NeMo Studio is a single-replica launcher for vLLM or NIM containers. It does not use NVIDIA Dynamo.
- Fine-tuning in NeMo Studio is broad: three training backends, distillation and reinforcement learning.
- The recommended Helix install needed five fixes on a clean machine. We reported them to NVIDIA.
What we reported to NVIDIA
Katonic filed seven issues from this test: three on NeMo Fabric and four on NeMo Helix. Six are bug reports with a reproduction and a proposed fix. The seventh proposes a Katonic adapter for NeMo Fabric. They follow four earlier bug reports on the NeMo Agent Toolkit, all since fixed upstream.
Where Katonic fits
Every agent built in Katonic is a declarative definition that compiles to NVIDIA's open agent software. When NVIDIA changes the runtime, Katonic changes the compiler and your agents carry over. The move from the NeMo Agent Toolkit to Relay and Fabric is that change in progress: Katonic has built a NeMo Fabric adapter that runs a Katonic agent definition and has proposed it to NVIDIA. Helix is a workbench for engineers. Katonic is the governed platform an organisation runs agents on, with tenancy, access control and audit.
Read next
- What is NVIDIA NeMo Helix? →
- What is NVIDIA NeMo Relay? →
- What is NVIDIA NeMo Fabric? →
- How to install NVIDIA NeMo Helix →
- Model deployment and fine-tuning in NeMo Studio →
- How to write a NeMo Fabric adapter →
- How to use NeMo Relay with LangChain and LangGraph →
- NeMo Helix vs LangGraph, Dify and n8n →
- NeMo Helix for enterprise →
- NeMo Agent Toolkit end of life →
- NeMo Agent Toolkit is deprecated: what replaces it? →
- Proposal: Katonic adapter for NeMo Fabric →
- NeMo Relay on GitHub →
- NeMo Fabric on GitHub →
- NeMo Helix on GitHub →
Questions and answers
What are NVIDIA NeMo Helix, Relay and Fabric?
Three open-source projects that replace the NeMo Agent Toolkit. Relay observes and controls model and tool calls, Fabric runs agents through one interface, and Helix evaluates, fine-tunes and improves agents.
Are NeMo Helix, Relay and Fabric open source?
Yes. All three are on GitHub under the Apache 2.0 licence.
Do I need all three?
No. Each works on its own. Helix bundles Fabric and Relay, so installing Helix gives you all three.
Are they ready for production?
All three were before version 1.0 in November 2026. Relay was the most mature in our test. Pin versions and expect change.
Does NeMo Helix need a GPU?
Not to run agents and evaluations against a hosted model API. Deploying a model or fine-tuning one inside NeMo Studio needs GPUs.
How does Katonic relate to them?
Katonic agents are declarative definitions that compile to NVIDIA's open agent software. Katonic has built a NeMo Fabric adapter for those definitions and proposed it to NVIDIA.