Get started

Install the adapter with Core:

pnpm add @anvia/core @anvia/azure

Create one client at the server boundary. Provide exactly one endpoint form (endpoint or baseUrl) and exactly one credential (apiKey or azureADTokenProvider), unless you inject an initialized SDK client.

import { Agent } from '@anvia/core'
import { AzureOpenAIClient } from '@anvia/azure'

const azure = new AzureOpenAIClient({
  endpoint: process.env.AZURE_OPENAI_ENDPOINT!,
  apiKey: process.env.AZURE_OPENAI_API_KEY!,
})

const agent = new Agent({
  id: 'support',
  model: azure.completionModel({
    modelId: process.env.AZURE_OPENAI_DEPLOYMENT!,
    api: 'responses',
  }),
  instructions: 'Answer support questions clearly.',
  maxTurns: 4,
})

const result = await agent.generate({
  prompt: 'Draft a concise reply to this ticket.',
})

if (result.type === 'response') {
  console.log(result.output)
}

modelId is the Azure deployment name. completionModel() returns an Anvia StreamingCompletionModel, so the same object works with agents and direct completion APIs. Omit api for Chat Completions, which is the default.

Stream a run

for await (const event of agent.stream({ prompt: 'Explain the resolution.' })) {
  if (event.type === 'text_delta') {
    process.stdout.write(event.delta)
  }
}

Add another capability

const embeddings = azure.embeddingModel({ modelId: 'my-embedding-deployment' })
const images = azure.imageGenerationModel({ modelId: 'my-image-deployment' })
const speech = azure.speechGenerationModel({ modelId: 'my-speech-deployment' })
const transcription = azure.transcriptionModel({ modelId: 'my-transcription-deployment' })

These objects share the underlying SDK client but not request state, and each needs a matching deployment on your endpoint.

Before production

• Keep the key or token provider on the server.
• Use one credential variable set per provider, such as AZURE_OPENAI_*, rather than reusing OpenAI variable names.
• Use deployment names as model IDs and keep them in deployment configuration.
• Set explicit contextLimits and controls for custom deployment names.
• Bound agent turns, tool access, and request timeouts in application policy; the managed client does not retry.
• Smoke test streaming and tool calls against the real endpoint. See Configuration.