NeMo Fabric
What is NVIDIA NeMo Fabric?
NVIDIA NeMo Fabric is an open-source runtime, under the Apache 2.0 licence, that lets an application run many different AI agents through one interface. Each kind of agent is connected by an adapter. The application describes what it wants in one configuration, and Fabric starts the right adapter, runs the agent and returns a result in a common shape. It replaces the part of the NeMo Agent Toolkit that configured and ran agents.
The pieces
| Piece | What it is |
|---|---|
| Configuration | One description of the run: which adapter, which model, the instructions, limits and telemetry |
| Adapter | A small program that knows how to start one kind of agent and report back |
| Descriptor | A file that tells Fabric what an adapter is called, what settings it takes and what it can do |
| Runtime | The part that starts adapters and keeps a session going across messages |
| Consumers | Applications that use Fabric. NVIDIA lists Harbor, NeMo Gym and NeMo Helix |
Which agents it runs
The catalogue that ships with Fabric lists adapters for coding agents and agent harnesses, including Claude, Codex, Deep Agents, Hermes, OpenClaw, OpenHands, Strands and others, plus a Remote Agent adapter that calls an agent already running behind an HTTP endpoint. For your own agent, you write an adapter.
How an adapter works
- It is a class with three operations: start, invoke and stop.
- A descriptor file names it and lists its settings and capabilities.
- Fabric gives it the model, the instructions and the limits in a common form, and the adapter turns those into whatever the agent needs.
- Token streaming is optional. An adapter that supports it declares the capability and sends chunks as they arrive.
- The adapter contract was at an alpha version in November 2026, so expect it to change.
What we found
We installed NeMo Fabric 0.6.1 on its own. NVIDIA's custom LangGraph example ran in every mode we tried: plan, run, with a tool server, with NeMo Relay, and as a two-message session. A small adapter of our own streamed tokens through it.
- There is no general LangGraph loader. Each agent, or each family of agents, needs its own adapter.
- Fabric starts the adapter under the system Python unless a virtual environment is activated or a variable is set. Installed with common tools, that fails on the first run. We reported the cause and a fix.
- The Remote Agent adapter declared a streaming capability it did not implement. We reported it with a fix.
- Setting a turn limit of one, as the quick start does, stops a tool-calling agent before it can answer. Leave it unset.
A Katonic adapter for NeMo Fabric
Katonic has built an adapter that reads a Katonic agent definition and runs it through NeMo Fabric, with token streaming and optional NeMo Relay telemetry. We have proposed it to NVIDIA as a contribution. It is how Katonic agents carry over from the NeMo Agent Toolkit to the software that replaces it.
Read next
- What is NVIDIA NeMo Relay? →
- What is NVIDIA NeMo Helix? →
- NeMo Relay vs NeMo Fabric vs NeMo Helix →
- Issue 636: Katonic adapter proposal →
- Issue 635: adapter starts under system Python →
- Issue 634: Remote Agent streaming capability →
- NeMo Fabric on GitHub →
Questions and answers
What is NVIDIA NeMo Fabric?
An open-source runtime that runs different AI agents through one interface, using an adapter for each kind of agent.
What is a NeMo Fabric adapter?
A small program with start, invoke and stop operations, plus a descriptor file, that connects one kind of agent to Fabric.
Does NeMo Fabric run LangGraph agents?
Yes, through an adapter. There is no general LangGraph loader, so you write an adapter for your agent. NVIDIA ships an example.
Does NeMo Fabric support streaming?
Yes, as an optional adapter capability. An adapter that declares it sends tokens as they arrive.
Is NeMo Fabric open source?
Yes. It is on GitHub under the Apache 2.0 licence.
Can I contribute an adapter to NeMo Fabric?
Outside contributors open an issue with a design or patch, and a maintainer can adopt it. That is the route Katonic used for its adapter.