Agent SDK

The Agent SDK is Claude Code with the interactive shell removed and a programming interface in its place. You get the same agent loop, the same tools, the same permission model and the same extensibility — and you decide what the agent does.

Available for TypeScript and Python.

What you get that a raw API loop would not give you

Building an agent directly on the Messages API means writing the loop, the tool dispatch, the retry logic, the permission prompts, the session persistence and the cost tracking yourself. The SDK ships those:

• The agent loop, streaming input and output
• Custom tools, plus MCP and tool search for scaling past a handful of tools
• Skills, plugins, subagents and hooks
• Permission handling and approval flows
• Sessions, including persisting them to external storage
• File checkpointing, so changes can be rewound
• Structured outputs, todo tracking, cost tracking and OpenTelemetry observability

What it does not do

• It is not a hosted service. You host it. Anthropic documents hosting and secure deployment separately, and the secure-deployment page is the one to read before shipping.
• It is not the Messages API. It sits on top. If all you need is one completion, you do not need an agent SDK.
• It is not stable across major versions by default. The TypeScript V2 session API is documented as removed — worth checking the migration guide before upgrading anything old.

Migration

There is a dedicated “Migrate to Claude Agent SDK” guide, which is the trail left by the rename from the earlier Claude Code SDK. If you have code referencing the old package name, start there.