agent-sdk-builder

Claude Agent SDK

The Claude Agent SDK lets you run Claude Code programmatically -- build AI agents that read files, write code, execute commands, search the web, and orchestrate subagents, all from your application code.

Key distinction: The Agent SDK (claude-agent-sdk) runs the full Claude Code agent loop with built-in tools. The Anthropic Client SDK (anthropic) is for raw API calls. Use the Agent SDK when you need autonomous tool-using agents.

Quick Reference

	TypeScript	Python
Package	@anthropic-ai/claude-agent-sdk	claude-agent-sdk
Install	npm install @anthropic-ai/claude-agent-sdk	pip install claude-agent-sdk
Auth	ANTHROPIC_API_KEY env var	ANTHROPIC_API_KEY env var
Core function	query()	query()
GitHub	anthropics/claude-agent-sdk-typescript	anthropics/claude-agent-sdk-python

The CLI package @anthropic-ai/claude-code is bundled inside the SDK -- no separate install needed.

---

1. Installation & Auth

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

# Python
pip install claude-agent-sdk
# or with uv
uv add claude-agent-sdk

Authentication via environment variable:

export ANTHROPIC_API_KEY=sk-ant-...

Alternative providers:

• Amazon Bedrock: CLAUDE_CODE_USE_BEDROCK=1 + AWS credentials
• Google Vertex AI: CLAUDE_CODE_USE_VERTEX=1 + GCP credentials
• Microsoft Azure: CLAUDE_CODE_USE_FOUNDRY=1 + Azure credentials

---

2. Core API -- query()

Both SDKs expose query() as the primary entry point. It returns an async iterator streaming SDKMessage objects. Claude handles the entire tool loop autonomously -- you do NOT implement tool execution.

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"],
    maxTurns: 10,
  },
})) {
  if (message.type === "assistant" && message.content) {
    for (const block of message.content) {
      if (block.type === "text") process.stdout.write(block.text);
    }
  }
  if ("result" in message) {
    console.log("\nFinal:", message.result);
    console.log("Cost:", message.total_cost_usd);
  }
}

Python

import asyncio
from claude_agent_sdk import query, ClaudeAgentOptions

async def main():
    options = ClaudeAgentOptions(
        allowed_tools=["Read", "Edit", "Bash"],
        max_turns=10,
    )
    async for message in query(prompt="Find and fix the bug in auth.py", options=options):
        if hasattr(message, "result"):
            print(f"Final: {message.result}")
            print(f"Cost: ${message.total_cost_usd:.4f}")

asyncio.run(main())

---

3. Configuration Options

Full Options Reference

Option (TS / Py)	Type	Description
allowedTools / allowed_tools	string[]	Tools to auto-approve without user confirmation
disallowedTools / disallowed_tools	string[]	Tools to always deny
permissionMode / permission_mode	string	Permission strategy (see Permissions section)
systemPrompt / system_prompt	string	Custom system prompt or "claude_code" for default
model	string	Model ID (e.g., "claude-fable-5", "claude-opus-4-8", "claude-sonnet-4-6") -- short aliases resolve to the latest date-slugged release (e.g., "claude-haiku-4-5" resolves to "claude-haiku-4-5-20251001"); pin a full slug for reproducibility
maxTurns / max_turns	number	Maximum agentic loop iterations
maxBudgetUsd / max_budget_usd	number	Spending cap in USD
effort	string	"low", "medium", "high", "max"
cwd	string	Working directory for file operations
mcpServers / mcp_servers	object	MCP server configurations
hooks	object	Lifecycle hook callbacks
agents	object	Subagent definitions
resume	string	Session ID to resume
continue / continue_conversation	boolean	Continue most recent session
forkSession / fork_session	string	Fork from an existing session
settingSources / setting_sources	string[]	Load settings from ["user", "project", "local"]
plugins	string[]	Local plugin directory paths
sandbox	object	Sandbox/isolation settings
thinking	object	Extended thinking: "adaptive", {type: "enabled", budget: N}, "disabled"
outputFormat / output_format	object	JSON schema for structured output
env	object	Environment variables passed to agent
canUseTool / can_use_tool	function	Runtime permission callback
includePartialMessages / include_partial_messages	boolean	Enable token-level streaming
spawnClaudeCodeProcess	function	Custom process spawner (VMs, containers, remote)
agentProgressSummaries	boolean	Enable periodic AI-generated progress summaries for running subagents
debug / debug	boolean	Enable programmatic debug logging
debugFile / debug_file	string	File path for debug log output

Example -- Full Configuration

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

for await (const msg of query({
  prompt: "Refactor the auth module to use JWT tokens",
  options: {
    model: "claude-sonnet-4-6",
    allowedTools: ["Read", "Write", "Edit", "Bash", "Glob", "Grep"],
    disallowedTools: ["WebSearch", "WebFetch"],
    permissionMode: "bypassPermissions",
    maxTurns: 25,
    maxBudgetUsd: 1.0,
    effort: "high",
    cwd: "/home/user/project",
    systemPrompt: "You are a senior backend engineer. Follow the project's coding standards.",
    thinking: "adaptive",
    env: { NODE_ENV: "development" },
  },
})) {
  // process messages
}

---

4. Built-in Tools

The agent has access to these tools by default:

Tool	Purpose
Read	Read files from filesystem
Write	Create new files
Edit	Precise string replacements in existing files
Bash	Execute shell commands
Glob	Find files by pattern
Grep	Search file contents with regex
WebSearch	Search the web
WebFetch	Fetch and parse web pages
Agent	Spawn subagents (required for multi-agent)
Skill	Invoke skills from plugins
AskUserQuestion	Request user input
TodoWrite	Manage task lists
ToolSearch	Discover deferred tools

Control which tools the agent can use:

// Only allow read-only operations
options: {
  allowedTools: ["Read", "Glob", "Grep"],
  disallowedTools: ["Bash", "Write", "Edit"],
}

---

5. Custom Tools via MCP

Create custom tools using the SDK's MCP server helpers. Tools are defined with schemas and handlers, then exposed as in-process MCP servers.

TypeScript

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

// Define tools
const getWeather = tool(
  "get_weather",
  "Get current weather for a city",
  { city: z.string(), units: z.enum(["celsius", "fahrenheit"]).default("celsius") },
  async ({ city, units }) => ({
    content: [{ type: "text", text: JSON.stringify({ city, temp: 22, units }) }],
  })
);

const searchDatabase = tool(
  "search_db",
  "Search the application database",
  { query: z.string(), limit: z.number().default(10) },
  async ({ query: q, limit }) => {
    const results = await db.search(q, limit);
    return { content: [{ type: "text", text: JSON.stringify(results) }] };
  }
);

// Create MCP server
const server = createSdkMcpServer({
  name: "app-tools",
  tools: [getWeather, searchDatabase],
});

// Use in query
for await (const msg of query({
  prompt: "What's the weather in Rome and find related travel posts?",
  options: {
    mcpServers: { app: server },
    allowedTools: ["mcp__app__get_weather", "mcp__app__search_db"],
  },
})) {
  // Custom tools are called automatically by the agent
}

Python

from claude_agent_sdk import tool, create_sdk_mcp_server, query, ClaudeAgentOptions

@tool("get_weather", "Get current weather for a city", {"city": str, "units": str})
async def get_weather(args):
    return {"content": [{"type": "text", "text": f"Weather in {args['city']}: 22C"}]}

@tool("search_db", "Search the database", {"query": str, "limit": int})
async def search_db(args):
    results = await db.search(args["query"], args["limit"])
    return {"content": [{"type": "text", "text": str(results)}]}

server = create_sdk_mcp_server(name="app-tools", tools=[get_weather, search_db])

options = ClaudeAgentOptions(
    mcp_servers={"app": server},
    allowed_tools=["mcp__app__get_weather", "mcp__app__search_db"],
)

async for msg in query(prompt="Weather in Rome?", options=options):
    pass

MCP Tool Naming Convention

Custom tools follow the pattern: mcp__<server-name>__<tool-name>

Example: server named "mytools" with tool "search" becomes mcp__mytools__search

External MCP Servers

Connect to external MCP servers via stdio or HTTP:

options: {
  mcpServers: {
    // stdio transport (local process)
    localServer: {
      command: "node",
      args: ["./mcp-server.js"],
    },
    // HTTP/SSE transport (remote)
    remoteServer: {
      url: "https://mcp.example.com/sse",
      headers: { Authorization: "Bearer token" },
    },
  },
}

---

6. Subagents -- Multi-Agent Orchestration

Subagents let you define specialized agents that the main agent can spawn for parallel or delegated work.

Defining Subagents

for await (const msg of query({
  prompt: "Review the codebase for quality and security issues",
  options: {
    allowedTools: ["Read", "Grep", "Glob", "Agent"],  // Agent tool required
    agents: {
      "security-reviewer": {
        description: "Security expert. Use for vulnerability scanning, auth review, injection detection.",
        prompt: "You are a security auditor. Find vulnerabilities, auth issues, and injection vectors.",
        tools: ["Read", "Glob", "Grep"],
        model: "opus",
      },
      "code-quality": {
        description: "Code quality reviewer. Use for style, patterns, complexity, dead code.",
        prompt: "Review code quality: naming, complexity, patterns, duplication.",
        tools: ["Read", "Glob", "Grep"],
        model: "sonnet",
      },
      "test-runner": {
        description: "Runs tests and reports results.",
        prompt: "Execute test suites and analyze failures.",
        tools: ["Bash", "Read", "Grep"],
        model: "haiku",
      },
    },
  },
})) {
  // Claude automatically decides when to spawn subagents
  // based on their descriptions
}

Subagent Properties

Property	Type	Description
description	string	Required. When to use this agent (Claude decides based on this)
prompt	string	Required. System prompt for the subagent
tools	string[]	Restricted tool set
model	string	"sonnet", "opus", "haiku", or "inherit"
disallowedTools	string[]	Tools to block (TS only)
mcpServers	object	MCP servers available to subagent
skills	string[]	Skills the subagent can invoke
memory	object	Memory configuration for the subagent
maxTurns	number	Turn limit for this subagent (TS only)

Subagent Behavior

• Context isolation -- each subagent gets a fresh conversation; only its final message returns to the parent
• Parallel execution -- multiple subagents run concurrently when spawned together
• No nesting -- subagents cannot spawn their own subagents
• Resumable -- subagents can be resumed by ID from tool results
• Cost isolated -- each subagent's token usage is tracked separately
• Progress summaries -- enable agentProgressSummaries: true to receive periodic AI-generated progress updates from running subagents

---

7. Session Management

Sessions persist conversation history to disk, enabling multi-turn workflows.

Resume a Session (TypeScript)

let sessionId: string | undefined;

// First query -- capture session ID
for await (const msg of query({
  prompt: "Read the authentication module",
  options: { allowedTools: ["Read", "Glob"] },
})) {
  if (msg.type === "system" && msg.subtype === "init") {
    sessionId = msg.session_id;
  }
}

// Second query -- resume with full context
for await (const msg of query({
  prompt: "Now refactor it to use JWT",
  options: { resume: sessionId },
})) {
  if ("result" in msg) console.log(msg.result);
}

Continue Most Recent Session

// TypeScript
for await (const msg of query({
  prompt: "Continue where we left off",
  options: { continue: true },
})) { /* ... */ }

# Python
options = ClaudeAgentOptions(continue_conversation=True)
async for msg in query(prompt="Continue where we left off", options=options):
    pass

Session Client (Python)

Python provides ClaudeSDKClient for managed multi-turn conversations:

from claude_agent_sdk import ClaudeSDKClient

async with ClaudeSDKClient() as client:
    # First turn
    await client.query("What's the project structure?")
    async for msg in client.receive_response():
        pass  # process

    # Second turn -- context retained automatically
    await client.query("Now find all API endpoints")
    async for msg in client.receive_response():
        pass  # process

    # Interrupt current generation
    await client.interrupt()

List, Inspect, and Manage Sessions

import {
  listSessions, getSessionInfo, getSessionMessages,
  forkSession, tagSession, renameSession,
} from "@anthropic-ai/claude-agent-sdk";

// List sessions
const sessions = await listSessions({ dir: "/path/to/project", limit: 10 });
for (const session of sessions) {
  console.log(session.sessionId, session.createdAt, session.tag);
}

// Single-session metadata lookup
const info = await getSessionInfo(sessionId);
console.log(info.tag, info.createdAt);

// Read conversation history (includes parallel tool results)
const messages = await getSessionMessages(sessionId);

// Branch a conversation from a specific point
const forked = await forkSession(sessionId);

// Organize sessions with tags and renames
await tagSession(sessionId, "auth-refactor");
await renameSession(sessionId, "auth-refactor-v2");

from claude_agent_sdk import (
    list_sessions, get_session_info, get_session_messages,
    fork_session, tag_session, rename_session,
)

sessions = await list_sessions(dir="/path/to/project", limit=10)
for session in sessions:
    messages = await get_session_messages(session.session_id)

info = await get_session_info(session_id)
await tag_session(session_id, "auth-refactor")
await rename_session(session_id, "auth-refactor-v2")

Session State Events

Session state change events are opt-in as of v0.2.83. Enable with environment variable:

export CLAUDE_CODE_EMIT_SESSION_STATE_EVENTS=1

Exit Reasons

The ExitReason type includes: "end_turn", "max_turns", "budget", "interrupt", "resume".

---

8. Introspection Utilities

import { supportedAgents, getSettings } from "@anthropic-ai/claude-agent-sdk";

// Discover available subagents
const agents = await supportedAgents();

// Inspect runtime-resolved settings (includes applied model and effort)
const settings = await getSettings();
console.log(settings.applied.model, settings.applied.effort);

---

9. Permissions

Control what the agent can do at runtime.

Permission Modes

Mode	Behavior
"default"	Unmatched tools trigger canUseTool callback or user prompt
"dontAsk"	Deny anything not pre-approved (TypeScript only)
"acceptEdits"	Auto-accept file mutations (Edit, Write, mkdir, touch, rm, mv, cp)
"bypassPermissions"	All tools run without prompts (use with caution)
"plan"	No execution -- planning/analysis only

Permission Evaluation Order

1. Hooks (PreToolUse)
2. Deny rules (disallowedTools) -- overrides everything, including bypassPermissions
3. Permission mode
4. Allow rules (allowedTools) -- does NOT constrain bypassPermissions
5. canUseTool callback

Important: disallowedTools is the only hard block. Even bypassPermissions cannot override it. Use it for safety-critical restrictions.

Runtime Permission Callback

for await (const msg of query({
  prompt: "Deploy the application",
  options: {
    allowedTools: ["Read", "Bash"],
    canUseTool: async (toolName, input) => {
      // Block destructive commands
      if (toolName === "Bash" && input.command?.includes("rm -rf")) {
        return { behavior: "deny", message: "Destructive commands not allowed" };
      }
      // Allow everything else
      return { behavior: "allow" };
    },
  },
})) { /* ... */ }

Callback return values:

• { behavior: "allow" } -- approve the tool call
• { behavior: "deny", message: "reason" } -- reject with explanation
• { behavior: "ask" } -- fall through to user prompt (default mode)

---

10. Hooks -- Lifecycle Events

Hooks intercept agent lifecycle events for logging, validation, or control flow.

Available Hook Events

Hook	When	Can modify?
PreToolUse	Before a tool executes	Yes -- allow, deny, or modify input
PostToolUse	After a tool completes	Yes -- modify output
PostToolUseFailure	After a tool fails	Log errors
UserPromptSubmit	When user sends a prompt	Modify prompt text
Stop	Agent is about to stop	Force continue or modify
SubagentStart	Subagent is starting	Modify subagent config
SubagentStop	Subagent completed	Process results
PreCompact	Before context compaction	Log or modify
Notification	Agent sends a notification	Display or forward
PermissionRequest	Tool needs permission	Auto-approve or deny
TaskCompleted	A task has been completed	Process results
TeammateIdle	A teammate agent is idle	Reassign or notify
ConfigChange	Configuration changed at runtime	Security auditing
SessionStart (TS)	Session initialized	Setup actions
SessionEnd (TS)	Session completed	Cleanup actions

Hook Matchers

Hooks can use regex matchers to target specific tools, and support async (fire-and-forget) mode:

hooks: {
  PreToolUse: [
    // Matcher targets specific tools by regex on tool name
    {
      matcher: /^(Write|Edit)$/,
      hooks: [async (event) => {
        if (event.input.file_path?.includes(".env")) {
          return { behavior: "deny", message: "Cannot modify .env files" };
        }
        return { behavior: "allow" };
      }],
      timeout: 30,  // seconds, default 60
    },
    // Async hook -- fire-and-forget logging (does not block)
    {
      matcher: /.*/,
      hooks: [{ async: true, asyncTimeout: 5, handler: async (event) => {
        await fetch("https://logs.example.com/webhook", {
          method: "POST",
          body: JSON.stringify({ tool: event.toolName, time: Date.now() }),
        });
      }}],
    },
  ],
}

Hook Return Values

Hook outputs can include:

• behavior: "allow", "deny", or "ask" (deny > ask > allow when multiple hooks conflict)
• systemMessage: inject a system message into conversation
• continue: force the agent to continue (for Stop hooks)
• hookSpecificOutput: event-specific modifications

Hook Example

for await (const msg of query({
  prompt: "Analyze the codebase",
  options: {
    hooks: {
      PreToolUse: async (event) => {
        console.log(`Tool: ${event.toolName}, Input: ${JSON.stringify(event.input)}`);
        // Block writes to production config
        if (event.toolName === "Write" && event.input.file_path?.includes("production")) {
          return { behavior: "deny", message: "Cannot modify production files" };
        }
        return { behavior: "allow" };
      },
      PostToolUse: async (event) => {
        console.log(`Tool ${event.toolName} completed in ${event.duration_ms}ms`);
      },
      Stop: async (event) => {
        console.log(`Agent stopping. Result: ${event.result}`);
      },
    },
  },
})) { /* ... */ }

---

11. Streaming

Message Types

Messages streamed from query() include:

Type	Description
system (subtype: init)	Session initialized -- contains session_id
system (subtype: api_retry)	API retry info -- attempt count, max retries, delay, error status
assistant	Claude's response with content blocks (text, tool_use)
result	Final result with result text, total_cost_usd, usage
task_progress	Real-time usage metrics for running agents
rate_limit	Rate limit event with retry timing (Python: RateLimitEvent)
stream_event	Partial token (when includePartialMessages: true)

Token-Level Streaming

for await (const msg of query({
  prompt: "Explain the auth flow",
  options: {
    includePartialMessages: true,
    allowedTools: ["Read"],
  },
})) {
  if (msg.type === "stream_event") {
    process.stdout.write(msg.delta?.text ?? "");
  }
}

---

12. Structured Output

Force the agent to return JSON conforming to a schema:

for await (const msg of query({
  prompt: "Analyze this codebase and list all API endpoints",
  options: {
    outputFormat: {
      type: "json_schema",
      json_schema: {
        name: "api_analysis",
        schema: {
          type: "object",
          properties: {
            endpoints: {
              type: "array",
              items: {
                type: "object",
                properties: {
                  method: { type: "string" },
                  path: { type: "string" },
                  handler: { type: "string" },
                },
                required: ["method", "path", "handler"],
              },
            },
          },
          required: ["endpoints"],
        },
      },
    },
  },
})) {
  if ("result" in msg) {
    const analysis = JSON.parse(msg.result);
    console.log(analysis.endpoints);
  }
}

---

13. Plugins and Skills

Load local plugins to give the agent access to custom skills, agents, and commands:

for await (const msg of query({
  prompt: "Review the frontend code",
  options: {
    plugins: ["/path/to/my-plugin"],
    allowedTools: ["Read", "Glob", "Grep", "Skill"],
  },
})) { /* agent can now invoke skills from the plugin */ }

Load filesystem settings (CLAUDE.md, .claude/ configs):

options: {
  settingSources: ["user", "project", "local"],
}

---

14. Cost Tracking

for await (const msg of query({ prompt: "Analyze auth.py", options: {} })) {
  if ("result" in msg) {
    console.log(`Total cost: $${msg.total_cost_usd}`);
    console.log(`Input tokens: ${msg.usage?.input_tokens}`);
    console.log(`Output tokens: ${msg.usage?.output_tokens}`);
    // Per-model breakdown (TypeScript)
    if (msg.modelUsage) {
      for (const [model, usage] of Object.entries(msg.modelUsage)) {
        console.log(`${model}: $${usage.cost_usd}`);
      }
    }
  }
}

Use maxBudgetUsd to set a hard spending cap:

options: { maxBudgetUsd: 0.50 }  // stop after $0.50

---

15. Hosting & Deployment Patterns

Ephemeral Sessions

New container per task, destroy on completion. Best for CI/CD, one-shot tasks.

// In a serverless function or CI step
const result = [];
for await (const msg of query({
  prompt: taskDescription,
  options: {
    maxTurns: 15,
    maxBudgetUsd: 0.25,
    permissionMode: "bypassPermissions",
    cwd: "/workspace",
  },
})) {
  if ("result" in msg) result.push(msg);
}

Long-Running Sessions

Persistent containers with multiple agent interactions. Best for interactive applications.

// Resume across container restarts
const sessionId = await loadSessionId();
for await (const msg of query({
  prompt: userInput,
  options: { resume: sessionId },
})) { /* ... */ }

Custom Process Spawning

Run agents in VMs, containers, or remote environments:

for await (const msg of query({
  prompt: "Analyze the repo",
  options: {
    spawnClaudeCodeProcess: async (options) => {
      // Spawn in a Docker container, VM, or remote server
      const container = await docker.createContainer({
        Image: "node:20",
        Cmd: ["npx", "@anthropic-ai/claude-code", ...options.args],
      });
      await container.start();
      return container.stream;
    },
  },
})) { /* ... */ }

Sandbox Isolation

Use Anthropic's sandbox runtime for secure execution:

import { createSandbox } from "@anthropic-ai/sandbox-runtime";

const sandbox = await createSandbox({ image: "node:20" });
for await (const msg of query({
  prompt: "Run the test suite",
  options: { sandbox },
})) { /* ... */ }

---

16. Security Best Practices

1. Always set allowedTools -- restrict to minimum necessary tools
2. Use maxBudgetUsd -- prevent runaway costs
3. Use maxTurns -- prevent infinite loops
4. Use canUseTool callbacks -- validate dangerous operations at runtime
5. Sandbox untrusted code -- use container isolation for user-submitted tasks
6. Never pass secrets in prompts -- use env option or MCP tools for credential access
7. Use disallowedTools -- explicitly block tools you never want used
8. Proxy credentials -- use a proxy pattern for API keys the agent needs

// Secure configuration example
options: {
  allowedTools: ["Read", "Glob", "Grep"],     // read-only
  disallowedTools: ["Bash", "Write", "Edit"],  // no execution or mutation
  maxTurns: 10,
  maxBudgetUsd: 0.25,
  permissionMode: "dontAsk",                   // deny anything not listed
  canUseTool: async (tool, input) => {
    // Additional runtime validation
    if (input.file_path?.includes("..")) {
      return { behavior: "deny", message: "Path traversal blocked" };
    }
    return { behavior: "allow" };
  },
}

---

17. Common Patterns

CI/CD Code Review Agent

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

async function reviewPR(diff: string): Promise<string> {
  let review = "";
  for await (const msg of query({
    prompt: `Review this PR diff for bugs, security issues, and style:\n\n${diff}`,
    options: {
      allowedTools: ["Read", "Glob", "Grep"],
      maxTurns: 15,
      maxBudgetUsd: 0.50,
      outputFormat: {
        type: "json_schema",
        json_schema: {
          name: "pr_review",
          schema: {
            type: "object",
            properties: {
              issues: { type: "array", items: { type: "object", properties: {
                severity: { type: "string" }, file: { type: "string" },
                line: { type: "number" }, description: { type: "string" },
              }}},
              summary: { type: "string" },
              approved: { type: "boolean" },
            },
            required: ["issues", "summary", "approved"],
          },
        },
      },
    },
  })) {
    if ("result" in msg) review = msg.result;
  }
  return review;
}

Multi-Agent Research Pipeline

for await (const msg of query({
  prompt: "Research best practices for rate limiting in Node.js APIs",
  options: {
    allowedTools: ["Read", "Grep", "Glob", "WebSearch", "WebFetch", "Agent"],
    agents: {
      "codebase-analyst": {
        description: "Analyzes the local codebase for existing patterns.",
        prompt: "Search the codebase for rate limiting implementations and patterns.",
        tools: ["Read", "Glob", "Grep"],
        model: "sonnet",
      },
      "web-researcher": {
        description: "Researches best practices and documentation online.",
        prompt: "Search for current best practices, libraries, and patterns.",
        tools: ["WebSearch", "WebFetch"],
        model: "sonnet",
      },
    },
  },
})) { /* ... */ }

Interactive Chat Application

from claude_agent_sdk import ClaudeSDKClient

async def chat_loop():
    async with ClaudeSDKClient() as client:
        while True:
            user_input = input("You: ")
            if user_input.lower() == "exit":
                break
            await client.query(user_input)
            async for msg in client.receive_response():
                if hasattr(msg, "content"):
                    for block in msg.content:
                        if block.get("type") == "text":
                            print(f"Agent: {block['text']}")

---

18. Migration from claude-code-sdk

The old claude-code-sdk / @anthropic-ai/claude-code-sdk packages are deprecated. Migration:

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

# Python
pip uninstall claude-code-sdk
pip install claude-agent-sdk

Update imports:

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

# Old
from claude_code_sdk import query, ClaudeCodeOptions
# New
from claude_agent_sdk import query, ClaudeAgentOptions

Breaking Changes

The API surface is mostly identical, but two critical defaults changed:

1. System prompt no longer defaults to Claude Code's prompt -- the new SDK uses a minimal system prompt. To restore Claude Code behavior:

   systemPrompt: { type: "preset", preset: "claude_code" }
   

2. Settings sources no longer loaded by default -- CLAUDE.md, .claude/ configs, and user settings are NOT loaded unless you opt in:

   settingSources: ["user", "project", "local"]
   

To fully restore old claude-code-sdk behavior, use both options together.

---

19. TypeScript V2 Preview

A simplified session-based API is available as a preview:

import {
  unstable_v2_createSession,
  unstable_v2_resumeSession,
  unstable_v2_prompt,
} from "@anthropic-ai/claude-agent-sdk";

// Create a session
await using session = await unstable_v2_createSession({
  model: "claude-sonnet-4-6",
  allowedTools: ["Read", "Edit", "Bash"],
});

// Send a prompt and stream response
const response = session.send("Find bugs in auth.py");
for await (const event of response) {
  // process streaming events
}

// Or use stream() for full message objects
const stream = session.stream("Now fix them");
for await (const msg of stream) {
  // process messages
}

// Resume a previous session
const resumed = await unstable_v2_resumeSession(sessionId);

// One-shot helper (no session management)
for await (const msg of unstable_v2_prompt("Quick question", { model: "claude-haiku-4-5" })) {
  // process
}

await using (TypeScript 5.2+) automatically closes the session when the scope exits. The V2 API does not yet support session forking.

---

20. ClaudeSDKClient Methods (Python)

The Python ClaudeSDKClient provides additional runtime control methods:

async with ClaudeSDKClient() as client:
    # Core conversation
    await client.query("Analyze the codebase")
    async for msg in client.receive_response():
        pass

    # Runtime controls
    await client.set_permission_mode("bypassPermissions")
    await client.set_model("claude-sonnet-4-6")

    # File management
    await client.rewind_files()  # undo file changes made by agent

    # MCP server management
    status = await client.get_mcp_status()
    await client.add_mcp_server("new-server", {"command": "node", "args": ["server.js"]})
    await client.remove_mcp_server("old-server")
    info = await client.get_server_info()

    # Interrupt current generation
    await client.interrupt()

    # Clean disconnect
    await client.disconnect()

---

Official Resources

• Agent SDK Overview
• TypeScript Reference
• Python Reference
• Subagents Guide
• Permissions Guide
• Hooks Guide
• Custom Tools / MCP
• Sessions
• Hosting
• Secure Deployment
• Migration Guide
• V2 Preview (TypeScript)
• Structured Output
• Cost Tracking
• Demo Apps