AI agents for NVIDIA
NIM endpoints and GPU-served models in your workflows. AGNT connects to NVIDIA so agents can read what happens there, reason about it, and act — with a receipt for every run.
NVIDIA and AGNT, in brief
| What it does | Lets an AGNT agent read from and act in NVIDIA. |
|---|---|
| Connects with | Bearer — held in the local vault |
| AGNT node | custom-api, or the zapier-action bridge |
| Runs when | A schedule or another workflow fires it |
| Typical build time | Ten minutes |
| Where data goes | Nowhere by default — AGNT runs on your machine and credentials never leave it. |
What you can automate
Agents can…
- call NIM model endpoints
- run GPU-served inference
Teams use it for
- self-hosted model serving behind agents
- specialised models as workflow steps
- local acceleration for heavy jobs
The NVIDIA workflow, ready to paste
Most NVIDIA builds begin the same way. A run starts when a scheduled run begins. The agent reads what arrived, judges it against criteria you wrote in plain language, and acts only where that judgment says it should — typically to call NIM model endpoints. That is what makes self-hosted model serving behind agents something you can hand over rather than merely schedule.
Three nodes carry that: something happens in NVIDIA, an agent reads it and forms a view, and it acts only where that view says it should. The definition below is the whole thing — paste it into a new workflow and it loads exactly as drawn.
{
"id": "f01ff69a-7a23-450b-a18e-a70563daf2ba",
"name": "NVIDIA — read, decide, act",
"nodes": [
{
"id": "af070d92-38c7-4ef6-a375-bb828add50ac",
"text": "Run on a schedule",
"x": 512,
"y": 144,
"isEditing": false,
"type": "trigger-timer",
"icon": "clock",
"category": "trigger",
"isSelected": false,
"parameters": {
"interval": "15",
"unit": "minutes"
},
"description": "Checks NVIDIA every fifteen minutes.",
"error": null,
"isActive": false,
"output": null,
"outputs": {}
},
{
"id": "137ed1f6-ff18-479b-ac5b-f879583f08be",
"text": "Work out what should change",
"x": 512,
"y": 336,
"isEditing": false,
"type": "agnt-agent",
"icon": "agnt",
"category": "action",
"isSelected": false,
"parameters": {
"instructions": "Decide what needs to happen in NVIDIA. Explain your reasoning."
},
"description": "Decides what needs to happen in NVIDIA, and why.",
"error": null,
"isActive": false,
"output": null,
"outputs": {}
},
{
"id": "d13deed0-4cd1-436f-a8f6-0b572f1adcc2",
"text": "Call the NVIDIA API",
"x": 512,
"y": 528,
"isEditing": false,
"type": "custom-api",
"icon": "connect",
"category": "action",
"isSelected": false,
"parameters": {
"url": "https://integrate.api.nvidia.com/v1/chat/completions",
"method": "POST",
"authType": "Bearer",
"headers": "{\"Content-Type\": \"application/json\"}"
},
"description": "POST against the NVIDIA API, authenticated with bearer.",
"error": null,
"isActive": false,
"output": null,
"outputs": {}
}
],
"edges": [
{
"id": "ff9534bb-3ca5-4a7c-afcf-8663a8799a4c",
"start": {
"id": "af070d92-38c7-4ef6-a375-bb828add50ac",
"type": "output"
},
"end": {
"id": "137ed1f6-ff18-479b-ac5b-f879583f08be",
"type": "input"
},
"startX": 800,
"startY": 168,
"endX": 512,
"endY": 360
},
{
"id": "8abbfd92-128f-4710-a06b-c346207f85cc",
"start": {
"id": "137ed1f6-ff18-479b-ac5b-f879583f08be",
"type": "output"
},
"end": {
"id": "d13deed0-4cd1-436f-a8f6-0b572f1adcc2",
"type": "input"
},
"startX": 800,
"startY": 360,
"endX": 512,
"endY": 552
}
],
"zoomLevel": 1,
"canvasOffsetX": 0,
"canvasOffsetY": 0,
"isTinyNodeMode": false
}Build this NVIDIA workflow
AGNT Community Core is free and runs on your own machine. Nothing about this workflow leaves it.
Self-hosted model serving behind your agents
NIM endpoints let a team serve models on their own GPUs with a standard interface, which suits organisations that have made the hardware investment and want workflows pointed at it rather than at a public API.
Specialised models as workflow steps
Not every step wants a general chat model. Vision, speech and domain-specific models served locally can slot in as individual steps, with the general model handling the reasoning around them.
How to reach NVIDIA from a workflow
There is no first-party NVIDIA node yet, and a page implying otherwise would cost you an afternoon. What works today is the generic custom-api node pointed at NVIDIA’s own API, authenticating with bearer.
| What you need | Value |
|---|---|
| Endpoint | https://integrate.api.nvidia.com/v1/chat/completions |
| Method | POST |
| Authentication | Bearer |
| Header | Authorization: Bearer {nvapiKey} |
| Reference | NVIDIA API documentation |
Worth knowing: The request and response shapes are OpenAI-compatible, so an existing OpenAI body works with only the host changed.
One thing to know before you build
The request and response shapes are OpenAI-compatible, so an existing OpenAI body works with only the host changed.
It is worth getting right first, because the failure is quiet rather than loud — you get a response, it just is not the one you wanted. Once past it, NVIDIA is a straightforward API to automate against.
Connect NVIDIA in two minutes
- Download AGNT Community Core — free, local-first, no account needed to run.
- Connect by pasting an API key into AGNT’s vault — stored encrypted on your machine, never uploaded.
- Drop the NVIDIA node into a workflow or hand it to an agent — the first run produces a receipt you can read line by line.
NVIDIA + AGNT — common questions
What is a NIM endpoint?
A standardised way to serve a model on your own GPU infrastructure, callable like any other provider.
When is this worth it?
When you already have GPU capacity, or when data residency rules out hosted inference.
Can it be mixed with hosted models?
Yes — different steps in one workflow can use different providers entirely.
