Anthropic Console

Interactive web-based interface for testing Claude prompts in real-time without writing code. Users compose prompts, adjust parameters (temperature, max tokens, model selection), and receive immediate responses with token counting and cost estimation. The Workbench maintains conversation history within a session and allows A/B testing of prompt variations side-by-side, with results persisted for comparison.

Console-based key management system for generating, revoking, and rotating API keys with granular control over key permissions and expiration policies. Keys are scoped to specific projects or applications, with audit logging of key creation and usage. The system supports automatic key rotation schedules and revocation of compromised keys without requiring account-level credential changes.

API support for streaming responses from Claude token-by-token in real-time, using Server-Sent Events (SSE) or WebSocket connections. Streaming enables lower perceived latency and allows applications to display responses as they are generated, rather than waiting for the complete response. Streaming responses include delta updates (new tokens) and metadata updates (tool calls, stop reasons).

API endpoint for generating dense vector embeddings from text, enabling semantic search, similarity comparison, and clustering. The embeddings API accepts text input and returns fixed-size vectors (dimension size unknown from docs) that capture semantic meaning. Embeddings can be stored in vector databases for retrieval-augmented generation (RAG) or used directly for similarity calculations.

API feature that enables Claude to engage in extended reasoning before generating a response, allowing the model to think through complex problems step-by-step. Extended thinking mode allocates additional computational resources to reasoning, resulting in longer response times but potentially higher-quality outputs for complex tasks. The API returns both the internal reasoning process and the final response.

Pre-built tools available to Claude for accessing external systems without requiring custom tool definitions. Built-in tools include web search (for current information), code execution (Python sandbox), bash shell access, text editor, and computer use (screenshot and interaction). These tools are automatically available in Claude's context and can be invoked without explicit tool definitions in the API request.

Official SDKs for 8 programming languages (Python, TypeScript, Go, Java, Ruby, PHP, C#, and CLI) that provide consistent API interfaces across all languages. Each SDK abstracts HTTP/REST details and provides language-native abstractions (async/await, iterators, type hints). SDKs handle authentication, request formatting, response parsing, and error handling, enabling developers to use Claude API idiomatically in their language of choice.

Integration with major cloud providers' AI platforms, enabling Claude API access through AWS Bedrock, Google Cloud Vertex AI, and Microsoft Azure Foundry. These integrations allow organizations to use Claude through their existing cloud provider accounts, with unified billing, IAM, and compliance frameworks. The API remains consistent across cloud providers, but authentication and deployment models differ.

Dashboard displaying API call metrics, token consumption (input and output), and estimated costs in USD. The monitoring system aggregates usage across all API calls made with keys in the account, with filtering by date range, model, and API endpoint. Cost calculations account for prompt caching savings and different pricing tiers per model version, with daily/monthly cost projections based on current usage patterns.

System for defining and managing cached prompt segments (system prompts, long context, repeated instructions) that are stored server-side and reused across multiple API calls. Cached segments are identified by content hash and reused when identical content is submitted in subsequent requests, reducing token costs by 90% for cached tokens. The console provides configuration UI for setting cache TTL (time-to-live), monitoring cache hit rates, and estimating savings.

Built-in evaluation tools in the console for systematically testing prompts against test cases, comparing outputs, and measuring quality metrics. The evaluation framework supports defining test datasets (input-output pairs), running batch evaluations across multiple prompt variants, and generating comparison reports. Evaluations can use custom scoring functions or built-in metrics (exact match, semantic similarity, token efficiency).

REST API endpoint for sending prompts to Claude and receiving responses, with support for managing multi-turn conversations by explicitly passing message history in each request. The API is stateless — the server does not maintain conversation state, so clients must manage the full message history (system prompt, user messages, assistant responses) and include it in each subsequent request. Responses include token usage metadata and tool call information.

Framework for defining custom tools that Claude can invoke during API calls, using JSON Schema to specify tool parameters and return types. Tools are defined in the API request with name, description, and input schema, and Claude can call tools by returning structured tool_use blocks. The system supports parallel tool execution (multiple tools in a single response), strict tool use mode (forcing Claude to use tools), and tool result handling via the messages API.

API support for processing images (JPEG, PNG, GIF, WebP) and PDF documents as input to Claude, with automatic image encoding and PDF text extraction. Images are sent as base64-encoded data or URLs, and PDFs are processed to extract text and visual elements. The system supports mixed text and image inputs in a single request, enabling Claude to analyze images, extract text from PDFs, and answer questions about visual content.

Asynchronous API for submitting large batches of requests to Claude and retrieving results later, with lower per-token costs than real-time API calls. Batch requests are submitted as JSONL files (one request per line), processed in the background, and results are retrieved via polling or webhook callbacks. Batch processing is optimized for cost reduction (typically 50% discount) rather than latency, making it suitable for non-time-sensitive workloads.

API feature that enforces Claude's responses to conform to a specified JSON schema, ensuring structured and predictable output format. The schema is provided in the API request, and Claude is constrained to generate only valid JSON matching the schema. This enables reliable parsing and downstream processing of Claude's responses without manual validation or error handling.