Anthropic Claude Agent SDK

The Claude Agent SDK is a Python and TypeScript developer library for building AI agents that read files, run commands, and call outside tools, without developers writing that control loop themselves. It is built on the same agent loop, tools, and context management used in Claude Code, Anthropic's coding tool. It supports the Model Context Protocol, or MCP, for connecting to outside data, subagents for delegating parts of a task, and hooks for running custom code at set points. Permission settings control which actions run automatically and which need approval. It is made by Anthropic, which renamed it from the earlier 'Claude Code SDK' and provides a migration guide. Use is governed by Anthropic's commercial terms of service.

This fits an engineering team that wants strong built-in guardrails, like requiring approval before risky actions, rather than assembling that logic from a lower-level API, and it works especially well for coding-adjacent agents such as one that reviews pull requests or triages bug reports. Start with a narrow, well-defined task, use the permission settings to require human approval on anything that changes data, and expand scope only after that works. There is no separate license fee; cost comes from Anthropic API usage billed by token, plus engineering time. Because the product was recently renamed, some online tutorials may still reference the older 'Claude Code SDK' name, so check that any guide you follow matches the current SDK.