NeMo Helix
What is NVIDIA NeMo Helix?
NVIDIA NeMo Helix is an open-source control plane, under the Apache 2.0 licence, for evaluating, fine-tuning and improving AI agents. It comes with a web interface called NeMo Studio and a command-line tool called nemo. It is one of the three projects, with NeMo Relay and NeMo Fabric, that replace the NeMo Agent Toolkit. It is a workbench for engineers who already have agents, not a visual tool for building them.
What is inside NeMo Studio
These are the sections of NeMo Studio in release 0.9.1, which we installed on 14 November 2026.
| Area | Sections | What they do |
|---|---|---|
| Observability | Insights, Traces | Show each agent run step by step, and analyse runs for patterns |
| Components | Agents | List registered agents and their deployments, with a chat tab |
| Models | Model Catalog, Evaluations, Fine-tuning, Deployments, Virtual Models | Register models and providers, score them, train them and serve them |
| Evaluations | Experiments | Compare agents or models on a dataset with metrics |
| Data | Datasets, Anonymizer, Data Designer, Safe Synthesizer | Hold test data, remove personal data, and generate synthetic data |
| System | Filesets, Jobs, Settings | Files, background jobs and configuration |
How an agent gets into Helix
- An agent is a configuration file that names a harness, a model, tools and telemetry.
- You register and deploy it with the nemo command, or ask a coding assistant to do it using the skills Helix ships.
- Helix runs the agent through NeMo Fabric and records the run through NeMo Relay.
- The sample agent that ships with Helix is an email security triage agent built on the Deep Agents harness.
What NeMo Helix is not
- It has no visual or drag-and-drop agent builder. The empty screens point you to a coding assistant or the command line.
- It is not a multi-tenant platform for a whole organisation. It has workspaces, and it can sit behind a sign-in service, but tenancy, quotas and chargeback are not its job.
- Its model deployment is a single-replica launcher for vLLM or NIM containers, without autoscaling.
What it needs to run
- Linux, Python, the uv tool and Docker. Helix starts a ClickHouse container to store traces.
- No GPU if your agents call a hosted model API. We ran it on four CPU cores.
- GPUs and a container job backend for model deployment and fine-tuning.
What we found
- The evaluation and trace tooling is good. An experiment comparing models on a sample dataset ran end to end and showed scores per model.
- The recommended install needed five fixes on a clean machine before the sample agent answered.
- Runs that fail are recorded as successful in the trace list, so failures cannot be filtered. We reported this to NVIDIA.
Where Katonic fits
Helix is where an engineer measures and improves an agent. Katonic is where an organisation builds, governs and runs agents for many teams. Every agent built in Katonic is a declarative definition that compiles to NVIDIA's open agent software, so the two are designed to work together and not to compete.
Read next
- How to install NVIDIA NeMo Helix →
- Model deployment and fine-tuning in NeMo Studio →
- NVIDIA NeMo Helix, Relay and Fabric: field guide →
- Issue 4091: failed runs stored as success →
- NeMo Helix on GitHub →
- NeMo Studio documentation →
Questions and answers
What is NVIDIA NeMo Helix?
An open-source control plane for evaluating, fine-tuning and improving AI agents, with a web interface called NeMo Studio.
What is NeMo Studio?
The web interface of NeMo Helix. It has sections for traces, agents, models, evaluations, fine-tuning, deployments, experiments and data.
Does NeMo Helix have a visual agent builder?
No. Agents are configuration files, registered from the command line or by a coding assistant using the skills Helix ships.
Is NeMo Helix free?
Yes. It is open source under the Apache 2.0 licence.
Does NeMo Helix replace the NeMo Agent Toolkit?
In part. Helix takes over evaluation, optimization and hardening. NeMo Relay and NeMo Fabric take over the rest.
Does NeMo Helix need a GPU?
Not for running and evaluating agents against a hosted model API. Model deployment and fine-tuning need GPUs.