Claude Agent SDK — Build Agents with Claude Code Power

Claude Agent SDK — Features & API

Core Capabilities

V2 Session API (New)

// Create a persistent session
const session = await claude.unstable_v2_createSession({
  workingDirectory: "./project",
});

// Send messages within the session
const response = await claude.unstable_v2_prompt(session.id, {
  prompt: "What files are in this project?",
});

// Continue the conversation
const followUp = await claude.unstable_v2_prompt(session.id, {
  prompt: "Now refactor the main module",
});

Available Tools

The SDK exposes the same 53 tools that Claude Code uses internally:

• File tools: Read, Write, Edit, Glob, Grep
• Shell tools: Bash command execution
• Search tools: Web search, web fetch
• Agent tools: Spawn sub-agents for parallel tasks
• Planning tools: Task decomposition and approval

Use Cases

1. Custom Coding Agents — Build your own Claude Code with custom behaviors
2. Email Assistants — Agents that read, classify, and respond to emails
3. Research Agents — Multi-agent systems that search, analyze, and synthesize
4. CI/CD Agents — Autonomous code review and fix workflows
5. Document Processors — Analyze and transform documents at scale

Migration from Claude Code SDK

The package was renamed from @anthropic-ai/claude-code-sdk to @anthropic-ai/claude-agent-sdk. Update your import:

// Old
import { Claude } from "@anthropic-ai/claude-code-sdk";
// New
import { Claude } from "@anthropic-ai/claude-agent-sdk";

FAQ

Q: What is the Claude Agent SDK? A: Anthropic's official TypeScript SDK that provides the same capabilities as Claude Code (file editing, command execution, codebase understanding) in a programmable API for building custom AI agents.

Q: Is it free? A: The SDK is free to install. You pay for Anthropic API usage (Claude model tokens). Governed by Anthropic's Commercial Terms of Service.

Q: How is this different from the regular Anthropic SDK? A: The regular SDK (@anthropic-ai/sdk) sends messages to Claude models. The Agent SDK adds the full agent loop — tool execution, file system access, shell commands, and multi-step autonomous workflows.