claude-agent-sdk

name

claude-agent-sdk

description

Build agents, tools, and multi-agent systems with the Claude Agent SDK (formerly Claude Code SDK). Use this skill whenever the user wants to plan, design, implement, debug, or extend anything involving the Claude Agent SDK, claude-agent-sdk, @anthropic-ai/claude-agent-sdk, claude_agent_sdk, agent automation, programmatic Claude usage, MCP tool servers, agent hooks, subagents, or SDK-based agent orchestration — even if they just mention "agent SDK", "building agents programmatically", or want to create tools for Claude to use. Covers the complete API for TypeScript (default) and Python.

Claude Agent SDK

Anthropic's framework for building autonomous AI agents. Same tools, agent loop, and context management that power Claude Code — as a library.

Versions: TypeScript v0.2.76 (@anthropic-ai/claude-agent-sdk) | Python v0.1.48 (claude-agent-sdk)

All examples are TypeScript. If the user needs Python, read references/python/python-api.md and references/python/python-examples.md.

Quick Start

npm install @anthropic-ai/claude-agent-sdk
export ANTHROPIC_API_KEY=your-key

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

for await (const message of query({
  prompt: "Find and fix the bug in auth.ts",
  options: {
    allowedTools: ["Read", "Edit", "Bash"],
    permissionMode: "acceptEdits",
    maxTurns: 10,
  },
})) {
  if (message.type === "result") {
    console.log(message.result);
    console.log(`Cost: $${message.totalCostUsd?.toFixed(4)}`);
  }
}

Also supports Bedrock (CLAUDE_CODE_USE_BEDROCK=1), Vertex AI (CLAUDE_CODE_USE_VERTEX=1), Azure (CLAUDE_CODE_USE_FOUNDRY=1).

---

Capability Map

Capability	Summary	Deep Dive
query()	Async iterator API — send prompt, get messages	This file
Configuration	25+ options: model, tools, limits, sandbox, env	references/configuration.md
Built-in Tools	17 tools: Read, Write, Edit, Bash, Glob, Grep, Web...	references/tools-and-mcp.md
Custom Tools	Define tools as in-process MCP servers	references/tools-and-mcp.md
MCP Integration	4 transports: stdio, SSE, HTTP, SDK	references/tools-and-mcp.md
Hooks	16 events to intercept/modify agent behavior	references/hooks.md
Permissions	5 modes + allow/deny lists + callback	references/permissions.md
Sessions	Multi-turn, resume, fork conversations	references/sessions.md
Subagents	Isolated child agents with restricted tools	references/subagents.md
Streaming	Real-time output + input streaming	references/streaming-and-output.md
Structured Output	Force typed JSON via Zod / JSON Schema	references/streaming-and-output.md
Error Handling	5 error types + result subtypes	This file
Cost Tracking	Per-step and per-model usage	This file
Hosting	Docker, cloud, sandbox providers, security	references/hosting-and-security.md
Advanced	Thinking, 1M context, dynamic config, plugins	references/advanced.md
Examples	WebSocket chat, multi-agent, AIClient wrapper	references/examples.md

---

Core API: query()

Returns an async iterator yielding messages as the agent works through its loop (prompt → Claude responds → execute tools → repeat):

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

for await (const message of query({
  prompt: "Your task description",
  options: { /* see Configuration */ },
})) {
  switch (message.type) {
    case "system":    // Session lifecycle (init, compact_boundary)
    case "assistant": // Claude's responses — text + tool_use blocks in message.content
    case "user":      // Tool results — tool_result blocks in message.content
    case "result":    // Final: result, subtype, sessionId, totalCostUsd, usage
  }
}

Result Subtypes

Subtype	Meaning	Recovery
success	Completed normally	Process result
error_max_turns	Hit maxTurns limit	Resume with sessionId
error_max_budget_usd	Hit cost cap	Increase budget or resume
error_during_execution	Runtime error	Check logs, retry
error_max_structured_output_retries	Schema validation failed	Fix schema or prompt

---

Configuration (Key Options)

const options = {
  // Tools
  allowedTools: ["Read", "Edit", "Bash"],       // Auto-approve
  disallowedTools: ["Write"],                    // Always deny
  permissionMode: "acceptEdits",                 // See Permissions section

  // Limits
  maxTurns: 20,
  maxBudgetUsd: 1.0,

  // Model
  model: "claude-opus-4-6",
  effort: "high",                                // "low" | "medium" | "high" | "max"

  // Session
  resume: "session-id",                          // Resume previous session
  forkSession: true,                             // Branch when resuming

  // System
  systemPrompt: "You are a security auditor.",
  cwd: "/path/to/project",
  env: { GITHUB_TOKEN: "..." },

  // Extensions
  mcpServers: { /* MCP server configs */ },
  hooks: { /* Hook configs */ },
  agents: { /* Subagent definitions */ },
  outputFormat: { /* Structured output schema */ },
};

For all 25+ options with types and defaults, read references/configuration.md.

---

Built-in Tools

Tool	Category	What It Does
Read	File I/O	Read files (images, PDFs, notebooks)
Write	File I/O	Create or overwrite files
Edit	File I/O	Precise string-replacement edits
Bash	Execution	Terminal commands, scripts, git
Glob	Search	Find files by pattern (**/*.ts)
Grep	Search	Regex content search (ripgrep)
WebSearch	Web	Search the web
WebFetch	Web	Fetch and parse web pages
AskUserQuestion	Interaction	Ask clarifying questions
Agent	Orchestration	Spawn subagents
Skill	Orchestration	Invoke filesystem skills
TodoWrite	Tracking	Task status tracking
NotebookEdit	File I/O	Jupyter cell edits
ToolSearch	Discovery	Load deferred tools on-demand

---

Custom Tools (MCP)

Define tools as in-process MCP servers. Tool naming: mcp__<server>__<tool>

import { query, tool, createSdkMcpServer } from "@anthropic-ai/claude-agent-sdk";
import { z } from "zod";

const server = createSdkMcpServer({
  name: "my-tools",
  tools: [
    tool("get_weather", "Get temperature for a location", {
      latitude: z.number(),
      longitude: z.number(),
    }, async (args) => ({
      content: [{ type: "text", text: `72°F at ${args.latitude}, ${args.longitude}` }],
    })),
  ],
});

for await (const msg of query({
  prompt: "Weather at 40.7, -74.0?",
  options: {
    mcpServers: { "my-tools": server },
    allowedTools: ["mcp__my-tools__*"],   // Wildcard approval
  },
})) { /* ... */ }

For external servers (stdio, SSE, HTTP), tool schemas, and patterns, read references/tools-and-mcp.md.

---

Hooks

Intercept and modify agent behavior at 16 event points:

Key Events	Trigger
PreToolUse / PostToolUse	Before/after tool execution
Stop	Agent stopping
SubagentStart / SubagentStop	Subagent lifecycle
PreCompact	Before context compaction
Notification	Status messages
SessionStart / SessionEnd	Session lifecycle

const options = {
  hooks: {
    PreToolUse: [{
      matcher: "Write|Edit",
      hooks: [async (input, toolUseId, ctx) => {
        if (input.tool_input.file_path?.includes(".env"))
          return {
            hookSpecificOutput: {
              hookEventName: "PreToolUse",
              permissionDecision: "deny",
              permissionDecisionReason: "Cannot modify .env files",
            },
          };
        return {};  // Allow
      }],
    }],
  },
};

For all events, output options, matchers, and chaining, read references/hooks.md.

---

Permissions

Mode	Behavior
default	Unmatched tools → canUseTool callback
acceptEdits	Auto-approve Edit, Write, mkdir, rm, mv, cp
dontAsk	Deny anything not in allowedTools
bypassPermissions	Approve everything (use with caution)
plan	No tool execution — planning only

Evaluation order: Hooks → disallowedTools → Mode → allowedTools → canUseTool callback

For canUseTool callback and common patterns, read references/permissions.md.

---

Sessions

Use Case	How
One-shot	Single query() call
Resume	options: { resume: sessionId }
Branch	options: { resume: sessionId, forkSession: true }
Multi-turn	TS V2 preview: createSession() + send()

// Capture session ID
let sessionId: string;
for await (const msg of query({ prompt: "Analyze auth", options })) {
  if (msg.type === "result") sessionId = msg.sessionId;
}

// Resume later
for await (const msg of query({
  prompt: "Now refactor it",
  options: { resume: sessionId },
})) { /* full context preserved */ }

For TS V2 preview, fork patterns, and cross-host resume, read references/sessions.md.

---

Subagents

Isolated child agents with their own context, restricted tools, and model override:

const options = {
  allowedTools: ["Read", "Grep", "Agent"],    // Agent tool required
  agents: {
    "code-reviewer": {
      description: "Expert code reviewer for quality and security.",
      prompt: "You are a code review specialist. Be thorough but concise.",
      tools: ["Read", "Glob", "Grep"],        // Restricted
      model: "sonnet",                         // Cheaper model
    },
  },
};

Key properties: isolated context, parent sees only final message, no nesting, resumable via agent ID.

For orchestration patterns and cost optimization, read references/subagents.md.

---

Streaming & Structured Output

Streaming

for await (const msg of query({
  prompt: "...",
  options: { includePartialMessages: true },
})) {
  if (msg.type === "stream_event" && msg.event.delta?.type === "text_delta") {
    process.stdout.write(msg.event.delta.text);
  }
}

Structured Output

import { z } from "zod";
import { zodToJsonSchema } from "zod-to-json-schema";

const Plan = z.object({
  steps: z.array(z.object({ title: z.string(), description: z.string() })),
  risks: z.array(z.string()),
});

for await (const msg of query({
  prompt: "Plan the feature",
  options: { outputFormat: { type: "json_schema", schema: zodToJsonSchema(Plan) } },
})) {
  if (msg.type === "result") console.log(msg.structuredOutput);  // Parsed JSON
}

For stream event types and input streaming, read references/streaming-and-output.md.

---

Error Handling

for await (const msg of query({ prompt: "...", options })) {
  if (msg.type === "result") {
    if (msg.subtype === "success") {
      console.log(msg.result);
    } else if (msg.subtype === "error_max_turns") {
      // Resume: query({ prompt: "Continue", options: { resume: msg.sessionId } })
    } else if (msg.subtype === "error_max_budget_usd") {
      console.log(`Budget exceeded: $${msg.totalCostUsd}`);
    }
  }
}

---

Cost Tracking

if (msg.type === "result") {
  console.log(`Cost: $${msg.totalCostUsd?.toFixed(4)}`);
  // Per-model breakdown
  for (const [model, usage] of Object.entries(msg.modelUsage ?? {})) {
    console.log(`${model}: $${usage.costUSD.toFixed(4)}`);
  }
}

Optimization tips: Use model: "haiku" for subagents, set maxBudgetUsd, limit maxTurns, use effort: "low" for simple tasks.

---

Python Support

If the user's project is Python, load these references:

• references/python/python-api.md — Imports, types, ClaudeSDKClient, Pydantic
• references/python/python-examples.md — All patterns in Python

Key differences: snake_case options, ClaudeAgentOptions class, async for iteration, Pydantic for structured output.

---

Reference Loading Guide

Load the reference file that matches the topic you need:

Need	Read
Full config options table	references/configuration.md
Tool schemas, custom tools, MCP transports	references/tools-and-mcp.md
Hook events, callbacks, patterns	references/hooks.md
Permission modes, canUseTool callback	references/permissions.md
Session resume, fork, TS V2	references/sessions.md
Subagent orchestration	references/subagents.md
Streaming details, structured output	references/streaming-and-output.md
Docker, sandbox, security	references/hosting-and-security.md
Thinking, 1M context, plugins, migration	references/advanced.md
Full real-world code examples	references/examples.md
Python API and types	references/python/python-api.md
Python code examples	references/python/python-examples.md