🗓️ 19092026 2037
The Claude Agent SDK is Anthropic's library for building applications where Claude works through multiple steps and tools. It provides an agent runtime so you do not need to implement every model-and-tool interaction yourself.
Mental model
You give the runtime a goal, allowed capabilities, and configuration. It coordinates model responses and tool use until the task finishes or a configured limit is reached.
Use it when you want to build a Claude-powered workflow as an application. It is different from MCP: the SDK runs or hosts an agent workflow, while MCP connects that workflow to compatible external capability servers.
Decide before choosing it
Choose an agent SDK when a multi-step model workflow is the product feature. Use an MCP server when the goal is to expose a reusable external system. Build your own loop when you need unusual control over branching, approvals, or execution policy.
The choices can compose: an Agent SDK application can consume MCP servers.