Vertex AI

Use AnthropicVertexClient when Claude is provisioned through Google Vertex AI. It returns the same AnthropicCompletionModel used by the direct client, so agents and completions do not need Vertex-specific code.

Configure Application Default Credentials

Set the project, region, and credentials in the server environment:

export ANTHROPIC_VERTEX_PROJECT_ID="my-gcp-project"
export CLOUD_ML_REGION="global"
export GOOGLE_APPLICATION_CREDENTIALS="/absolute/path/service-account.json"

Then create the client:

import { AnthropicVertexClient } from '@anvia/anthropic'

export const vertexAnthropic = new AnthropicVertexClient({
  projectId: process.env.ANTHROPIC_VERTEX_PROJECT_ID,
  region: process.env.CLOUD_ML_REGION ?? 'global',
})

export const model = vertexAnthropic.completionModel(
  'claude-sonnet-5',
)

The official Vertex SDK resolves Application Default Credentials. In Google-hosted environments, prefer the workload's attached service identity over a long-lived JSON key.

Use explicit Google authentication

AnthropicVertexClient also accepts the official Vertex SDK's googleAuth, authClient, accessToken, and transport options. This is useful for service-account impersonation or an application-managed authentication client.

import { GoogleAuth } from 'google-auth-library'
import { AnthropicVertexClient } from '@anvia/anthropic'

const googleAuth = new GoogleAuth({
  credentials: serviceAccountJson,
  scopes: 'https://www.googleapis.com/auth/cloud-platform',
})

const vertexAnthropic = new AnthropicVertexClient({
  projectId: 'my-gcp-project',
  region: 'global',
  googleAuth,
})

Install google-auth-library directly when application code imports it. Validate externally supplied credential objects and never commit service-account JSON.

Vertex-specific boundaries

• Construction fails when neither region nor CLOUD_ML_REGION is available.
• Model availability and model IDs depend on the configured Vertex project and region.
• Vertex credentials, IAM permissions, quotas, and provider failures surface through the official SDK.
• Model listing is not available on AnthropicVertexClient.

Run an authenticated startup check or smoke test in each deployment environment. Local ADC success does not prove that the production workload identity has access to the selected Claude model.