Get started

Install the adapter with Core:

pnpm add @anvia/core @anvia/openai

Create one client at the server boundary. Construction requires apiKey unless an initialized OpenAI client is supplied.

import { AgentBuilder } from '@anvia/core'
import { OpenAIClient } from '@anvia/openai'

const openai = new OpenAIClient({
  apiKey: process.env.OPENAI_API_KEY,
})

const agent = new AgentBuilder(
  'support',
  openai.completionModel('gpt-5'),
)
  .instructions('Answer support questions clearly.')
  .defaultMaxTurns(4)
  .build()

const result = await agent.prompt('Draft a concise reply to this ticket.').send()
console.log(result.output)

completionModel() uses the Responses adapter by default. It returns an Anvia StreamingCompletionModel, so the same object works with agents and direct completion APIs.

Stream a run

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

The agent stream includes runtime events around the provider stream: turns, tools, usage, completion, and errors. For one provider call without an agent loop, use Completions.

Add another model capability

Create only the capability objects the process needs:

const embeddings = openai.embeddingModel('text-embedding-3-small')
const image = openai.imageGenerationModel()
const speech = openai.audioGenerationModel()
const transcription = openai.transcriptionModel()

These objects share the underlying OpenAI SDK client but do not share request state.

Before production

• Keep the API key on the server.
• Set explicit model IDs instead of relying on defaults across releases.
• Bound agent turns, tool access, and request timeouts in application policy.
• Inspect provider errors rather than retrying every failure.
• Record normalized usage through an Anvia observer.
• Review Responses and compatible endpoints before targeting a non-OpenAI service.