Azure AI Foundry v1.19.0+
Every model deployed on your Azure resource, OpenAI’s and others’ such as Llama, DeepSeek, Mistral, Grok and Phi, with chat, streaming, tools, vision and embeddings.
Table of contents
The Azure provider talks to the models deployed on an Azure AI Foundry (or Azure OpenAI) resource, through Azure’s OpenAI-compatible APIs. Requests go to your resource, so they stay within your Azure tenant and region.
Configuration
Give it the address the Azure portal shows for your deployment or resource, and the resource’s key:
import { createLLM } from "@node-llm/core";
const llm = createLLM({
provider: "azure",
azureApiKey: process.env.AZURE_API_KEY,
azureApiBase: "https://my-resource.services.ai.azure.com/openai/v1"
});
const chat = llm.chat("my-deepseek"); // the deployment's name
Or via environment variables:
NODELLM_PROVIDER=azure
AZURE_API_KEY=...
AZURE_API_BASE=https://my-resource.services.ai.azure.com/openai/v1
The address
Any of the addresses the portal shows works:
| Address | Example |
|---|---|
| The resource’s v1 address | https://my-resource.services.ai.azure.com/openai/v1 |
| The resource itself | https://my-resource.services.ai.azure.com (Foundry’s model inference) |
| A model endpoint | https://my-resource.services.ai.azure.com/models/chat/completions?api-version=2024-05-01-preview |
| A deployment’s address | https://my-resource.openai.azure.com/openai/deployments/my-gpt?api-version=2025-04-01-preview |
| A serverless endpoint | https://my-llama.eastus2.models.ai.azure.com/chat/completions |
With a deployment’s address, the address decides which deployment answers. With any other, the model you pass is the deployment’s name.
Signing in without a key
Resources that turn key sign-in off take a Microsoft Entra ID token instead:
const llm = createLLM({
provider: "azure",
azureAuthToken: process.env.AZURE_AUTH_TOKEN,
azureApiBase: "https://my-resource.services.ai.azure.com/openai/v1"
});
The token is used as given; get a new one before it expires.
Deployment names
On Azure you choose each deployment’s name. When it is the model’s own name, such as gpt-4o or DeepSeek-V3.2, NodeLLM knows the model’s capabilities and price. When it isn’t, say which model each deployment runs:
const llm = createLLM({
provider: "azure",
azureApiKey: process.env.AZURE_API_KEY,
azureApiBase: "https://my-resource.services.ai.azure.com/openai/v1",
azureDeployments: { "team-llama": "llama-3.3-70b-instruct" }
});
const response = await llm.chat("team-llama").ask("Hello");
console.log(response.usage.cost); // priced as llama-3.3-70b-instruct
A deployment NodeLLM doesn’t know is assumed able to chat, call tools and return structured output: Azure itself refuses what the model can’t do.
Features
- Models: every model a resource deploys, OpenAI’s (GPT, o-series) and others’ (Llama, DeepSeek, Mistral, Grok, Phi, Kimi).
- Chat and streaming: on every chat model.
- Tools: on models that call tools. GPT-5.6 and GPT-6 take function tools only through the Responses API; as on OpenAI, NodeLLM sends those requests, and only those, to the resource’s
/openai/v1/responses, naming your deployment. - Structured output: JSON schema and JSON mode, where the model supports them.
- Vision: images on vision models such as
gpt-4o. - Embeddings: through an embedding deployment, such as
text-embedding-3-small. - Listing models:
llm.listModels()lists the models the resource’s region offers, which you can deploy.
Not yet supported on Azure: image generation, transcription, moderation, Cohere’s embed and rerank APIs, and Claude models, which Azure serves through Anthropic’s API.
Embeddings Example
const llm = createLLM({
provider: "azure",
azureApiKey: process.env.AZURE_API_KEY,
azureApiBase: "https://my-resource.services.ai.azure.com/openai/v1"
});
const result = await llm.embed("Hello world", { model: "text-embedding-3-small" });
console.log(result.vectors[0]);
Getting a key
Create a resource and deploy a model in the Azure AI Foundry portal. The resource’s key and addresses are on its overview page.