Anthropic Python SDK

The Anthropic Python SDK is the official client library for calling the Claude API from Python applications. It exposes a fully typed, ergonomic interface over the Messages API, streaming, tool use, batch processing, and model management, so you can integrate Claude without hand-writing HTTP requests or response parsing.

Generated from Anthropic’s OpenAPI specification and hand-augmented with helpers, it ships synchronous and asynchronous clients, automatic retries with backoff, rich error types, and first-class support for running on Amazon Bedrock and Google Vertex AI. Pydantic-backed models give editor autocompletion and static type checking across every request and response.

What You Get

• Synchronous (Anthropic) and asynchronous (AsyncAnthropic) clients over the full Claude API surface
• First-class streaming helpers with typed events for incremental message output
• Built-in tool use, a tool runner, and MCP integration for agentic workflows
• Automatic retries with exponential backoff, configurable timeouts, and typed error hierarchy
• Cloud-provider clients for Amazon Bedrock and Google Vertex AI alongside the direct API

Common Use Cases

• Building chat and assistant experiences powered by Claude models
• Streaming long-form generations token-by-token to a UI or terminal
• Orchestrating tool-calling and agent loops with function definitions and MCP servers
• Running large offline jobs through the Message Batches API
• Deploying Claude on enterprise infrastructure via Bedrock or Vertex AI

Under The Hood

Architecture

The public entry points Anthropic and AsyncAnthropic in src/anthropic/_client.py (~1,100 lines) subclass SyncAPIClient/AsyncAPIClient in src/anthropic/_base_client.py (~2,650 lines), which centralizes the httpx transport, request building, retry-with-backoff logic, pagination, and streaming decode. API surface is organized into resource namespaces under src/anthropic/resources/ (messages, models, completions, and a beta tree), each mapping to typed request/response models in types/. Streaming lives in _streaming.py, while agentic features (tool runner, MCP integration, session runners) sit in lib/tools/, and hand-written auth plus cloud variants live in lib/credentials, lib/bedrock, and lib/vertex.

Tech Stack

Targets Python 3.9+ and is built on httpx for HTTP, Pydantic (v1.9-v3) for validated models, anyio for async, and jiter for fast JSON decoding, with distro and sniffio as runtime helpers. Optional extras pull in aiohttp, boto3/botocore (Bedrock/AWS), google-auth (Vertex), and the mcp package. The core is generated from Anthropic’s OpenAPI spec by Stainless and hand-augmented; the dev toolchain uses uv for locking plus ruff, pyright, mypy, and pytest.

Code Quality

The library is fully typed and ships py.typed, with strict type checking enforced via pyright and mypy in CI. Testing is substantial — roughly 103 test files under tests/ using pytest, pytest-asyncio, respx for HTTP mocking, and inline-snapshot. Error handling is well-structured: _exceptions.py defines a granular hierarchy keyed to HTTP status codes, and retry/timeout behavior is configurable and consistent across sync and async paths.

API Design

The developer experience is a strong point: a client is instantiated with an API key (auto-read from ANTHROPIC_API_KEY) and used through resource-oriented calls like client.messages.create(...), with sync and async interfaces kept symmetric. Getting started requires almost no boilerplate, and the repo backs the API with 30-plus runnable scripts in examples/ plus dedicated api.md, helpers.md, and tools.md references, making the naming and usage patterns easy to discover.