Anthropic’s docs subtitle it “Build production AI agents with Claude Code as a library”, packaging what powers Claude Code (Claude Agent SDK docs). It records over 33 million PyPI downloads per month as of August 2026, mostly below the radar of framework comparisons.
What do you actually get in the harness?
The SDK inherits everything that makes Claude Code usable day to day: built-in tools for reading files, editing code, running commands, and searching the web; hooks that run custom code at points in the agent lifecycle; subagents for specialized tasks; MCP connections to external systems (what MCP is); granular permissions controlling which tools run automatically versus which need approval; and sessions that maintain context across exchanges. It is a complete AI agent harness rather than a set of parts.
Where is the boundary?
The tradeoff is the flip side of the lineage. The SDK is Claude native on purpose, so teams that need model portability weigh it against provider-neutral frameworks. Its high download count against modest community buzz fits the pattern of a vendor embedding a package across its own tooling surface. Used for what it is, a runtime with production guardrails battle-tested inside another product, it is one of the strongest starting points in the ecosystem.