Anthropic API

Generates text responses using Claude models (Opus, Sonnet, Haiku) with a 200,000 token context window, enabling processing of entire documents, codebases, or conversation histories in a single request. The Messages API accepts a `messages` array with role/content fields and returns structured responses with token usage metadata, supporting both streaming and batch processing modes for flexible integration patterns.

Enables Claude to call external functions via a schema-based tool registry, supporting both synchronous request-response loops and agentic patterns where the model iteratively calls tools, receives results, and decides next actions. The implementation uses strict tool use enforcement mode and supports parallel tool execution, with Tool Runner providing SDK-level abstraction for managing the call-response cycle and error propagation.

Enables Claude to write and execute Python code directly within the API, enabling computational tasks, data analysis, and verification of outputs. The model generates Python code, which is executed in a sandboxed environment, and results are returned to the model for further analysis or refinement. This creates a feedback loop where Claude can test code, see errors, and iterate on solutions.

Generates vector embeddings for text, enabling semantic search, similarity comparison, and clustering. The embeddings API converts text into high-dimensional vectors that capture semantic meaning, enabling downstream applications like RAG systems, recommendation engines, or semantic search. Embeddings are compatible with standard vector databases (Pinecone, Weaviate, Milvus, etc.) for scalable similarity search.

Automatically generates citations linking Claude's responses to source documents or web results, improving transparency and enabling users to verify claims. Citations include source references (document names, URLs, page numbers) and can be used to trace information back to original sources. This is particularly useful for research, journalism, and compliance applications where source attribution is critical.

Streams response tokens in real-time as they are generated, enabling progressive display of output without waiting for the entire response to complete. The streaming API uses Server-Sent Events (SSE) or similar mechanisms to deliver tokens incrementally, reducing perceived latency and enabling interactive applications. Streaming works with all Claude features (vision, tool use, structured outputs) and includes streaming refusals for safety.

Integrates with MCP servers to access external tools, data sources, and services through a standardized protocol. Anthropic originated MCP and provides native support for both local and remote MCP servers, enabling Claude to interact with custom tools, databases, APIs, and services without requiring API-level integration. MCP servers can be registered and managed through the SDK or configuration files.

Enables Claude to interact with graphical user interfaces by accepting screenshots as input and executing actions (mouse clicks, keyboard input, scrolling) to automate GUI-based workflows. The model analyzes visual context from screenshots and generates structured action commands that are executed by the client, creating a feedback loop for multi-step automation tasks without requiring API-level GUI automation frameworks.

Processes images (JPEG, PNG, GIF, WebP) alongside text in the same request, enabling Claude to analyze visual content, extract information, answer questions about images, and generate descriptions. The vision capability is integrated into the Messages API — images are passed as content blocks with optional text annotations, and the model returns text analysis without separate vision API calls.

Constrains Claude's output to match a specified JSON schema, ensuring responses conform to predefined structure for downstream processing. The model generates text that is parsed and validated against the schema before returning to the client, preventing hallucinated fields or type mismatches. This enables reliable extraction of structured data without post-processing or regex parsing.

Caches large, reusable prompt segments (system prompts, documents, code context) at the API level, reducing token costs for subsequent requests that reference the same cached content. The caching mechanism uses content-based hashing to identify reusable blocks and stores them server-side, enabling cost savings of up to 90% on cached tokens while reducing latency for cache hits. Works seamlessly with tool use and structured outputs.

Processes multiple API requests asynchronously in batches, optimizing throughput and reducing per-request costs. Requests are submitted to a batch queue, processed in the background, and results are retrieved via polling or webhook callbacks. Batch processing is ideal for non-latency-sensitive workloads like data processing, content generation, or analysis tasks where cost optimization is prioritized over immediate response.

Enables Claude to perform extended internal reasoning before generating responses, using a 'thinking' phase to work through complex problems step-by-step. The model allocates computational resources to deliberation, which improves accuracy on reasoning-heavy tasks like mathematics, logic puzzles, and code analysis. Thinking tokens are counted separately from output tokens, with transparent cost tracking for the reasoning overhead.

Automatically adjusts the amount of internal reasoning effort based on task complexity, allocating more computational resources to difficult problems and less to straightforward queries. Unlike extended thinking (which is always-on), adaptive thinking dynamically determines when reasoning is beneficial, reducing wasted tokens on simple tasks while ensuring complex problems receive adequate deliberation.

Integrates real-time web search capability into Claude's responses, enabling the model to retrieve current information from the internet and cite sources. When activated, Claude can search the web for recent news, data, or information not in its training data, and includes citations linking to source URLs. This is implemented as a built-in tool that Claude can invoke during generation.