agent-sdk-guide

Build autonomous agents using the Claude Agent SDK by installing the TypeScript or Python package, authenticating with your provider, and configuring tools like Read, Write, Bash, and WebSearch. The SDK handles multi-step task execution, context management, and error handling for production deployments.

Claude agents support Read, Write, Bash, and WebSearch tools, plus MCP-based extensions. Fine-grained permissions let you restrict each tool's access—for example, limiting file writes to specific directories or controlling which commands Bash can execute.

The Agent SDK supports multiple authentication providers configured via file-based settings or environment variables. Choose your provider during setup, store credentials securely, and the SDK handles token refresh and session management automatically.

Yes, the Claude Agent SDK provides installation and code examples for both TypeScript and Python. Both implementations support the same core features—tool permissions, authentication, error handling, prompt caching, and MCP extensions.

The SDK includes error handling, automatic retries, session management, observability hooks for monitoring, and prompt caching to reduce latency and costs. These features ensure agents run reliably at scale with predictable behavior.

Extend agents using MCP (Model Context Protocol) to add custom tools and data sources. The SDK's MCP support lets you integrate specialized services, proprietary APIs, and domain-specific capabilities while maintaining permission controls.