The Claude Agent SDK gives developers access to Claude Code's capabilities, allowing them to create applications that can handle natural language prompts and manage conversations over multiple turns. Key features include the ability to stream responses in real-time, use built-in tools for file operations and command execution, and define custom tools through the Model Context Protocol (MCP). The SDK maintains session persistence, which helps in managing ongoing conversations efficiently. The communication between the SDK and Claude Code occurs through a command-line interface (CLI). When an application constructs a query, it spawns the CLI as a subprocess, which streams JSON messages back to the application. This system allows applications to process messages as they arrive, enabling real-time interaction. The CLI can operate in various modes, such as emitting standard output in JSON format or plain text, depending on the specified flags. Messages exchanged between the application and Claude Code are structured in JSON format. Each message has a type, such as "system," "assistant," or "user," facilitating clear communication. The initial message in a session sets up the context, including the active model and available tools. Responses from Claude can include text and tool usage, with detailed information on token usage for input and output, which can help developers optimize their queries and manage costs effectively. Finally, the documentation outlines various flags and configuration options for customizing interactions, such as session management, tool permissions, and model selection. This level of detail ensures that developers can tailor the SDK to fit specific needs and use cases, making it a flexible solution for integrating AI-driven functionalities into applications.