The Agent SDK

The Claude Agent SDK is Claude Code as a library — the same agentic loop, tools, and context management, embedded in your Python or TypeScript application. Use it for CI bots, review pipelines, and production agents.

from claude_agent_sdk import query

async def main():
    async for message in query(
        prompt="Find and fix the failing test in this repo",
        options={
            "allowed_tools": ["Read", "Grep", "Bash", "Edit"],
            "permission_mode": "acceptEdits",
            "cwd": "/work/my-repo",
        },
    ):
        print(message)

The SDK inherits the Claude Code feature set: CLAUDE.md and skills can load into SDK agents, hooks intercept behavior at the same lifecycle points, subagents isolate context, and MCP servers connect external tools (including in-process custom tools — your own functions exposed to Claude). Sessions persist and can resume/fork; permissions are declarative allow/deny rules plus modes, exactly as in the CLI.

For quick scripting without the SDK, claude -p "prompt" runs headless from any shell.

Check your understanding

Question 1What is the Agent SDK, conceptually?

Question 2Which Claude Code features carry over to SDK agents?