Claude Code Agent SDK

Claude Agent SDK documentation — build production AI agents with Claude Code as a library in Python or TypeScript. Use when building, configuring, or debugging agents with the Claude Agent SDK.

Build production-grade AI agents using Claude Code as a library. Supports Python and TypeScript.

Covers

• Quickstart and setup
• Agent loop and message types
• Tool configuration and built-in tools
• Hooks for intercepting agent behavior
• Session management (resume, fork, continue)
• MCP server integration
• Sub-agents and parallel execution
• Permissions and security
• Custom tools
• Streaming and structured outputs
• Cost tracking and observability
• Claude Code features (skills, commands, memory)
• Hosting and deployment
• Python and TypeScript API references.

Quick Navigation

Reference files are split by topic in references/. Load only what you need:

Core

• overview.md — SDK overview, installation, authentication, built-in tools, permission modes
• quickstart.md — Quick start: building a bug fix agent
• agent-loop.md — Agent loop: message types, tool execution, context window, compression

Agent Control

• hooks.md — Hooks: PreToolUse/PostToolUse/Stop, matchers, callback functions
• permissions.md — Permissions: allowedTools/disallowedTools, permission modes
• user-input.md — User input: approval prompts, AskUserQuestion

Sessions & State

• sessions.md — Session management: continue/resume/fork, ClaudeSDKClient
• file-checkpointing.md — File checkpointing: track and restore file changes

Tools & Integration

• custom-tools.md — Custom tools: in-process MCP server
• mcp.md — MCP: connect external tools + tool search extending to thousands of tools

Sub-agents & Parallelism

• subagents.md — Sub-agents: AgentDefinition, context isolation, parallel tasks + todo tracking

Streaming & Output

• streaming-output.md — Streaming responses + input schemas + structured outputs (JSON Schema/Zod/Pydantic)

Configuration & Features

• claude-code-features.md — Claude Code features: settingSources, CLAUDE.md, skills
• modifying-system-prompts.md — System prompts: claude_code preset vs custom
• skills.md — Agent Skills
• slash-commands.md — Slash commands
• plugins.md — Plugin system

Operations

• hosting.md — Hosting + secure deployment: Docker, cloud, CI/CD, isolation, credential management
• cost-tracking.md — Cost tracking: token usage, budgeting
• observability.md — Observability: OpenTelemetry
• migration-guide.md — Migration guide

API Reference

• python.md — Python SDK API reference
• typescript.md — TypeScript SDK API reference (includes V2 deprecation notes)

Key Patterns

Basic agent

from claude_agent_sdk import query, ClaudeAgentOptions

async for message in query(
    prompt="Fix the bug in auth.py",
    options=ClaudeAgentOptions(allowed_tools=["Read", "Edit", "Bash"])
):
    print(message)

With hooks

options = ClaudeAgentOptions(
    hooks={"PreToolUse": [HookMatcher(matcher="Bash", hooks=[validate_command])]}
)

With MCP servers

options = ClaudeAgentOptions(
    mcp_servers={"playwright": {"command": "npx", "args": ["@playwright/mcp@latest"]}}
)

With sub-agents

options = ClaudeAgentOptions(
    allowed_tools=["Read", "Glob", "Agent"],
    agents={"reviewer": AgentDefinition(
        description="Code reviewer",
        prompt="Review code quality",
        tools=["Read", "Glob", "Grep"]
    )}
)

Multi-turn with sessions

async with ClaudeSDKClient(options=options) as client:
    await client.query("Analyze the auth module")
    async for msg in client.receive_response(): print(msg)
    await client.query("Now refactor it")  # auto-continues
    async for msg in client.receive_response(): print(msg)