Claude Agent Sdk

Claude Agent SDK

Build AI agents that autonomously read files, run commands, search the web, edit code, and more. The Agent SDK provides the same tools, agent loop, and context management that power Claude Code, programmable in Python and TypeScript.

Installation

Prerequisites

Install Claude Code runtime first:

# macOS/Linux/WSL
curl -fsSL https://claude.ai/install.sh | bash

# or via npm
npm install -g @anthropic-ai/claude-code

SDK Installation

# TypeScript
npm install @anthropic-ai/claude-agent-sdk

# Python
pip install claude-agent-sdk

API Key

export ANTHROPIC_API_KEY=your-api-key

Alternative authentication:

• Amazon Bedrock: CLAUDE_CODE_USE_BEDROCK=1
• Google Vertex AI: CLAUDE_CODE_USE_VERTEX=1
• Microsoft Foundry: CLAUDE_CODE_USE_FOUNDRY=1

Basic Usage

TypeScript

import { query } from "@anthropic-ai/claude-agent-sdk";

for await (const message of query({
  prompt: "Find and fix the bug in auth.py",
  options: { allowedTools: ["Read", "Edit", "Bash"] }
})) {
  if ("result" in message) console.log(message.result);
}

Python

import asyncio
from claude_agent_sdk import query, ClaudeAgentOptions

async def main():
    async for message in query(
        prompt="Find and fix the bug in auth.py",
        options=ClaudeAgentOptions(allowed_tools=["Read", "Edit", "Bash"])
    ):
        if hasattr(message, "result"):
            print(message.result)

asyncio.run(main())

Built-in Tools

Key Features

For detailed API reference and examples:

• TypeScript: See references/typescript-api.md
• Python: See references/python-api.md

Permission Modes

• default: Standard permission checks
• acceptEdits: Automatically approve file edits
• bypassPermissions: Skip all permission checks (use with caution)

Sessions

Resume conversations with full context:

// Capture session ID from init message
if (message.type === 'system' && message.subtype === 'init') {
  sessionId = message.session_id;
}

// Resume later
for await (const message of query({
  prompt: "Continue where we left off",
  options: { resume: sessionId }
})) { ... }

MCP Integration

Connect external tools via Model Context Protocol:

const response = query({
  prompt: "Open example.com and describe what you see",
  options: {
    mcpServers: {
      playwright: { command: "npx", args: ["@playwright/mcp@latest"] }
    }
  }
});

Subagents

Spawn specialized agents for focused tasks:

const response = query({
  prompt: "Use the code-reviewer agent to review this codebase",
  options: {
    allowedTools: ["Read", "Glob", "Grep", "Task"],
    agents: {
      "code-reviewer": {
        description: "Expert code reviewer for quality and security reviews.",
        prompt: "Analyze code quality and suggest improvements.",
        tools: ["Read", "Glob", "Grep"]
      }
    }
  }
});

Hooks

Run custom code at key points in the agent lifecycle:

• PreToolUse, PostToolUse
• Stop, SessionStart, SessionEnd
• UserPromptSubmit

Budget Control

const response = query({
  prompt: "Analyze the codebase",
  options: {
    maxBudgetUsd: 2.5  // Stop if cost exceeds $2.50
  }
});

Project Skills

Load custom skills from .claude/commands/:

const response = query({
  prompt: "Review the pull request using available skills",
  options: {
    settingSources: ["project"]  // Loads skills from .claude/commands/
  }
});

Message Types

The SDK streams various message types:

• system (subtype init): Session initialization with session_id
• assistant: Claude's text responses and tool calls
• user: Tool results and user input
• result: Final response with execution stats
• error: Error information

Available Models

• claude-opus-4-5-20251101 - Most capable
• claude-sonnet-4-20250514 - Balanced (default)
• claude-sonnet-4-5-20250929 - Latest Sonnet
• claude-haiku-4-5-20251001 - Fastest

SDK vs Client SDK

The Agent SDK handles tool loops autonomously:

// Agent SDK: Claude handles tools autonomously
for await (const message of query({ prompt: "Fix the bug in auth.py" })) {
  console.log(message);
}

The Client SDK requires you to implement tool execution:

// Client SDK: You implement the tool loop
let response = await client.messages.create({...});
while (response.stop_reason === "tool_use") {
  const result = yourToolExecutor(response.tool_use);
  response = await client.messages.create({ tool_result: result, ... });
}

Resources

• TypeScript SDK GitHub
• Python SDK GitHub
• Example Agents