@core-ai/anthropic-vertex

Vertex AI Anthropic (Claude) provider package for @core-ai/core-ai. It uses the @anthropic-ai/vertex-sdk client and shares its request, streaming, tool-calling, structured-output, and reasoning behavior with @core-ai/anthropic.

Installation

npm install @core-ai/core-ai @core-ai/anthropic-vertex zod

Usage

import { generate } from '@core-ai/core-ai';
import { createAnthropicVertex } from '@core-ai/anthropic-vertex';

const anthropicVertex = createAnthropicVertex({
    projectId: process.env.GOOGLE_VERTEX_PROJECT,
    region: 'europe-west1',
});
const model = anthropicVertex.chatModel('claude-sonnet-4-6');

const result = await generate({
    model,
    messages: [{ role: 'user', content: 'Hello!' }],
});

console.log(result.content);

Authentication

By default, the provider uses Google Application Default Credentials (ADC):

const anthropicVertex = createAnthropicVertex({
    projectId: process.env.GOOGLE_VERTEX_PROJECT,
    region: 'europe-west1',
});

To authenticate with an explicit service account instead, parse its JSON key and pass it as credentials:

const rawCredentials = process.env.GOOGLE_APPLICATION_CREDENTIALS_JSON;
const credentialsJson = rawCredentials?.trim().startsWith('{')
    ? rawCredentials
    : Buffer.from(rawCredentials ?? '', 'base64').toString('utf8');

const anthropicVertex = createAnthropicVertex({
    projectId: process.env.GOOGLE_VERTEX_PROJECT,
    region: 'europe-west1',
    credentials: JSON.parse(credentialsJson),
});

You can also inject a preconfigured client (for example a custom AnthropicVertex instance, or one wired up for testing):

import { AnthropicVertex } from '@anthropic-ai/vertex-sdk';

const anthropicVertex = createAnthropicVertex({
    client: new AnthropicVertex({ projectId: 'my-project', region: 'eu' }),
});

Model IDs and regions

Pass the Vertex model ID (as listed in the Vertex AI Model Garden) to chatModel(). Model availability and naming vary by region — for example, some Claude models are only published to Vertex AI's eu multi-region while others are published to region-specific endpoints like europe-west1. Since a provider instance targets a single region, create separate providers if you need to reach models hosted in different regions:

const anthropicVertexEu = createAnthropicVertex({
    projectId: process.env.GOOGLE_VERTEX_PROJECT,
    region: 'eu',
});
const anthropicVertexEuWest1 = createAnthropicVertex({
    projectId: process.env.GOOGLE_VERTEX_PROJECT,
    region: 'europe-west1',
});

This package supports chat models only.

Caveat: Pass the unversioned Vertex model id (e.g. claude-sonnet-4-6) rather than a version-pinned id (e.g. claude-sonnet-4-6@20250929). Reasoning-effort and sampling-restriction capability detection (getAnthropicModelCapabilities and friends in @core-ai/anthropic) only recognizes the unversioned form today, so a version-pinned id silently falls back to standard capabilities instead of the model's actual ones.

Provider options and model capabilities

This provider shares its request, streaming, and reasoning behavior with @core-ai/anthropic. Provider-options schemas (anthropicGenerateProviderOptionsSchema), reasoning metadata types (AnthropicReasoningMetadata), and model capability helpers (getAnthropicModelCapabilities) are available from @core-ai/anthropic and apply equally to Vertex-hosted Claude models.