@anthropic-ai/vertex-sdk

Initializes authenticated HTTP clients for Google Cloud Vertex AI endpoints using Application Default Credentials (ADC) or explicit service account credentials. The SDK wraps Google's auth libraries to automatically handle token refresh, credential discovery from environment variables, and GAPIC client configuration for Vertex-specific endpoints, eliminating manual OAuth2 setup.

Routes Claude API requests (text generation, vision, tool use) through Google Cloud Vertex AI's managed endpoints instead of Anthropic's direct API. The SDK translates standard Anthropic SDK method calls into Vertex AI-compatible gRPC/REST payloads, maintaining API parity while leveraging Vertex's infrastructure, scaling, and audit logging.

Enables submitting multiple API requests to Vertex AI's batch processing endpoint for asynchronous execution at reduced cost (typically 50% discount). Handles request batching, polling for completion, and result retrieval without blocking on individual request latency.

Provides runtime detection of available Claude models on Vertex AI, their capabilities (vision, tool use, context window size), and version information. Allows applications to select models dynamically based on required features or cost constraints.

Implements streaming token-by-token responses from Claude models via Vertex AI using Server-Sent Events (SSE) or gRPC streaming, buffering and parsing Vertex-specific event formats into standard Anthropic SDK event objects. Handles backpressure, connection drops, and partial message recovery automatically.

Processes images (base64-encoded, URLs, or GCS paths) through Claude's vision capabilities via Vertex AI, automatically handling image format validation, size constraints, and Vertex-specific image encoding. Supports multi-image inputs and mixed text-image prompts in a single API call.

Enables Claude to request tool execution through Vertex AI by defining tools as JSON schemas, parsing Claude's tool_use content blocks, and routing tool calls through Vertex-managed infrastructure. Supports parallel tool calls, nested tool use, and automatic argument validation against schemas.

Manages multi-turn conversation state by maintaining message history (user and assistant messages) and passing it to Vertex AI in subsequent API calls. Handles message role validation, content concatenation, and context window management to prevent exceeding Vertex AI's token limits.

Estimates token consumption for prompts and messages before sending to Vertex AI using Claude's tokenizer, enabling cost prediction and context window validation. Supports counting tokens for text, images, and tool definitions separately.

Implements automatic retry logic for transient Vertex AI failures (rate limits, temporary outages) with exponential backoff, while distinguishing between retryable errors (429, 503) and permanent failures (401, 400). Provides detailed error messages mapping Vertex AI error codes to actionable remediation steps.

Provides hooks for logging and monitoring all API requests and responses to Vertex AI, including latency metrics, token usage, and error rates. Integrates with standard logging frameworks and allows custom middleware for observability integration (e.g., OpenTelemetry, Datadog).

Provides full TypeScript type definitions for all Vertex AI API parameters, responses, and message types, enabling IDE autocomplete, compile-time type checking, and inline documentation. Types are generated from Anthropic's API schema and kept in sync with Vertex AI's supported models.