Claude Agent SDK

Claude Agent SDK

Complete guide to the Claude Agent SDK for building custom AI agents.

Overview

The Claude Agent SDK allows you to programmatically spawn and control Claude Code sessions from TypeScript/JavaScript or Python applications.

Installation

TypeScript/JavaScript

npm install @anthropic-ai/claude-agent-sdk
# or
pnpm add @anthropic-ai/claude-agent-sdk

Python

# Using uv (recommended)
uv init && uv add claude-agent-sdk

# Using pip
python3 -m venv .venv && source .venv/bin/activate
pip3 install claude-agent-sdk

Prerequisite: Claude Code CLI must be installed:

curl -fsSL https://claude.ai/install.sh | bash

Required env var: ANTHROPIC_API_KEY

Basic Usage

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

// Simple query
const response = await claude("What files are in this directory?");
console.log(response.text);

// With options
const response = await claude("Refactor this function", {
  model: "claude-sonnet-4-6",
  maxTurns: 10,
  cwd: "/path/to/project",
  allowedTools: ["Read", "Write", "Edit", "Glob", "Grep"],
});

Python SDK

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"],
            permission_mode="acceptEdits",
        ),
    ):
        print(message)

asyncio.run(main())

Python Options

ClaudeAgentOptions(
    allowed_tools=["Read", "Edit", "Bash"],
    permission_mode="acceptEdits",    # default, acceptEdits, dontAsk, bypassPermissions, plan
    system_prompt="Custom prompt",
    agents={"reviewer": AgentDefinition(
        description="Expert code reviewer",
        prompt="Analyze code quality.",
        tools=["Read", "Glob", "Grep"],
    )},
    mcp_servers={"playwright": {...}},
)

TypeScript Streaming

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

// Stream responses
for await (const message of query({
  prompt: "Explain this codebase",
  options: {
    allowedTools: ["Read", "Glob", "Grep"],
  }
})) {
  console.log(message);
}

Event Types

interface TextEvent {
  type: "text";
  text: string;
}

interface ToolUseEvent {
  type: "tool_use";
  name: string;
  input: Record<string, unknown>;
  id: string;
}

interface ToolResultEvent {
  type: "tool_result";
  tool_use_id: string;
  content: string;
  is_error: boolean;
}

interface StopEvent {
  type: "stop";
  reason: "end_turn" | "max_turns" | "error";
}

Options

interface ClaudeOptions {
  // Model selection
  model?: string;                    // Claude model to use

  // Conversation control
  maxTurns?: number;                 // Max agentic turns
  systemPrompt?: string;            // Override system prompt
  appendSystemPrompt?: string;      // Append to system prompt

  // Working directory
  cwd?: string;                     // Working directory for file operations

  // Permissions
  allowedTools?: string[];          // Allowed tools list
  disallowedTools?: string[];       // Denied tools list
  permissionMode?: "default" | "plan" | "bypassPermissions";

  // Session management
  sessionId?: string;               // Resume existing session
  continue?: boolean;               // Continue last session

  // MCP
  mcpConfig?: string;               // Path to MCP config file

  // Environment
  env?: Record<string, string>;     // Additional env vars

  // Input
  inputFormat?: "text" | "stream-json";

  // Abort
  abortSignal?: AbortSignal;        // Cancel the request
}

Built-in Sub-Agent Types

When using the Agent tool within Claude Code, these sub-agent types are available:

Type	Purpose	Tools Available
general-purpose	Multi-step tasks, code search	All tools
Explore	Fast codebase exploration	Read-only tools
Plan	Architecture & implementation planning	Read-only tools
claude-code-guide	Claude Code feature questions	Read-only + web
researcher	Deep research with web access	Read-only + web
test-writer	Write comprehensive tests	All tools
code-reviewer	Code review and quality	Read-only + Bash
debugger	Debug errors and failures	All tools
doc-writer	Documentation generation	Write tools
security-reviewer	Security audit	Read-only + Bash
regex-expert	Regular expression design	Read/Write tools
prisma-specialist	Prisma ORM operations	All tools
redis-specialist	Redis caching patterns	All tools
graphql-specialist	GraphQL API design	All tools
infrastructure-specialist	Distributed systems, k8s	All tools
docker-ops	Docker build and registry	Read-only + Bash
k8s-image-auditor	K8s deployment auditing	Read-only + Bash
ansible-specialist	Ansible automation	All tools
pulumi-specialist	Pulumi IaC	All tools
coverage-analyzer	Test coverage analysis	All tools
spec-validator	API spec validation	All tools
rabbitmq-specialist	RabbitMQ messaging	All tools
event-streaming-architect	CQRS/event sourcing	All tools
kafka-specialist	Apache Kafka streaming	All tools
plugin-manager	Claude Code plugin management	All tools
statusline-setup	Claude Code status line config	Read + Edit

Agent Lifecycle & Orchestration

When spawning agents, Claude should prefer to orchestrate — delegate work to specialist agents rather than doing it directly. Key principles:

1. Track all agent IDs — record every ID returned from Agent tool calls
2. Check in on background agents — use SendMessage(to: agent_id) every 2 min
3. Audit every output — spawn a second agent to review the first agent's work
4. Clean up idle agents — terminate agents idle >3 min with no planned follow-up
5. Use run_in_background: true for independent work you don't need immediately
6. Use isolation: "worktree" for agents that write code to prevent conflicts

Agent Naming for Resumption

Name your agents for easy resumption:

Agent(name="builder-1", subagent_type="general-purpose", ...)
Agent(name="auditor-1", subagent_type="code-reviewer", ...)

# Later, resume:
SendMessage(to: "builder-1", message: "Fix the issues found in audit")

Multi-Agent Patterns

Parallel Research

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

// Run multiple agents in parallel
const [security, performance, docs] = await Promise.all([
  claude("Review security of auth module", {
    allowedTools: ["Read", "Glob", "Grep"],
  }),
  claude("Profile and optimize the API routes", {
    allowedTools: ["Read", "Glob", "Grep", "Bash"],
  }),
  claude("Generate API documentation", {
    allowedTools: ["Read", "Glob", "Grep", "Write"],
  }),
]);

Pipeline Pattern

// Step 1: Research
const research = await claude("Analyze the authentication system");

// Step 2: Plan based on research
const plan = await claude(`Based on this analysis: ${research.text}\n\nCreate an implementation plan`);

// Step 3: Implement based on plan
const impl = await claude(`Implement this plan: ${plan.text}`, {
  allowedTools: ["Read", "Write", "Edit", "Bash"],
});

Supervisor Pattern

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

async function supervisedTask(task: string) {
  // Worker does the task
  const result = await claude(task, {
    maxTurns: 20,
    allowedTools: ["Read", "Write", "Edit", "Bash", "Glob", "Grep"],
  });

  // Reviewer checks the work
  const review = await claude(
    `Review this work for quality and correctness:\n\nTask: ${task}\n\nResult: ${result.text}`,
    {
      allowedTools: ["Read", "Glob", "Grep"],
    }
  );

  return { result, review };
}

Spawning Agents within Claude Code

Inside Claude Code conversations, use the Agent tool:

Agent tool parameters:
- prompt: The task description
- subagent_type: Agent specialization (see table above)
- description: Short 3-5 word summary
- run_in_background: true/false for async execution
- isolation: "worktree" for isolated git worktree
- mode: "plan", "acceptEdits", "bypassPermissions", "default", "dontAsk"
- resume: Agent ID to resume previous agent

Background Agents

// Launch in background - get notified on completion
Agent(subagent_type="researcher", run_in_background=true, ...)

// Resume a completed agent
Agent(resume="agent-id-here", ...)

Worktree Isolation

// Run agent in isolated git worktree
Agent(subagent_type="test-writer", isolation="worktree", ...)
// Agent gets its own copy of the repo
// Changes are on a separate branch

Custom Agents

Create custom agent definitions in .claude/agents/:

---
name: my-specialist
description: Domain-specific expert
tools:
  - Read
  - Write
  - Edit
  - Bash
model: claude-sonnet-4-6
---

# My Specialist Agent

You are an expert in [domain]. When activated, you should:

1. Analyze the current context
2. Apply domain-specific knowledge
3. Provide actionable recommendations

## Guidelines
- Always check existing code before suggesting changes
- Follow project conventions
- Include tests for new functionality

Agent Teams

Agent teams allow multiple Claude instances to collaborate on tasks.

Configuration

# Enable agent teams
export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1

# Start as teammate
claude --teammate-mode "task description"

# Set team name
export CLAUDE_CODE_TEAM_NAME="my-team"

# Require plan mode for teammates
export CLAUDE_CODE_PLAN_MODE_REQUIRED=1

Team Patterns

// Coordinator spawns specialized teammates
const result = await claude("Coordinate the following tasks", {
  agents: {
    "frontend-dev": {
      description: "Frontend specialist",
      prompt: "You handle React components",
      tools: ["Read", "Write", "Edit", "Bash"],
    },
    "backend-dev": {
      description: "Backend specialist",
      prompt: "You handle API routes",
      tools: ["Read", "Write", "Edit", "Bash"],
    },
  },
});

Remote Sessions

# Create a cloud session
claude --remote "implement feature X"

# Resume a web session locally
claude --teleport

# Start remote control
/remote-control

Error Handling

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

try {
  const response = await claude("risky operation", {
    maxTurns: 5,
  });

  if (response.exitCode !== 0) {
    console.error("Agent encountered an error");
  }
} catch (error) {
  if (error.name === "AbortError") {
    console.log("Agent was cancelled");
  } else {
    console.error("Unexpected error:", error);
  }
}

CI/CD Integration

# GitHub Actions example
- name: Code Review with Claude
  run: |
    npx claude -p "Review the changes in this PR and report issues" \
      --output-format json \
      --max-turns 10 \
      > review.json
  env:
    ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}