Skills
skills/waltersumbon/claude-agent-sdk-skill/claude-agent-sdk-typescript

claude-agent-sdk-typescript

SKILL.md

Claude Agent SDK — TypeScript Guide

Production guidance for building AI agents with the Claude Agent SDK in TypeScript.

Naming: The Claude Code SDK was renamed to the Claude Agent SDK (v0.1.0+). Package: npm install @anthropic-ai/claude-agent-sdk · Import: import { query } from "@anthropic-ai/claude-agent-sdk"

Quick Reference — Two Interaction Modes

1. query() — Stateless, One-Shot

Best for: independent tasks, automation scripts, CI pipelines.

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

const options: ClaudeAgentOptions = {
  allowedTools: ["Read", "Edit", "Glob"],
  permissionMode: "acceptEdits",
};

for await (const message of query({
  prompt: "Review utils.ts for bugs. Fix any issues you find.",
  options,
})) {
  if (message.type === "assistant") {
    for (const block of message.message.content) {
      if ("text" in block) console.log(block.text);
      else if ("name" in block) console.log(`Tool: ${block.name}`);
    }
  }
  if (message.type === "result") {
    console.log(`Done: ${message.subtype}`);
  }
}

2. ClaudeSDKClient — Stateful, Multi-Turn

Best for: conversations, follow-up questions, interactive apps.

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

const client = new ClaudeSDKClient({
  options: {
    allowedTools: ["Read", "Write", "Bash"],
    permissionMode: "acceptEdits",
  },
});

try {
  await client.query("Analyze the codebase structure");
  for await (const msg of client.receiveMessages()) {
    console.log(msg);
  }
  // Continue the conversation with context preserved
  await client.query("Now refactor the largest file you found");
  for await (const msg of client.receiveMessages()) {
    console.log(msg);
  }
} finally {
  await client.close();
}

ClaudeAgentOptions — Complete Configuration

All options are optional. Key fields (all camelCase):

Field Type Description
allowedTools string[] Tools Claude can use. See Built-in Tools below.
disallowedTools string[] Explicitly block specific tools.
permissionMode string "default", "acceptEdits", or "bypassPermissions".
systemPrompt string | object Custom instructions. Use { type: "preset", preset: "claude_code" } for CC default.
model string e.g. "sonnet", "opus", "haiku", or full model string.
cwd string Working directory for the agent.
maxTurns number Maximum agentic loop iterations.
settingSources string[] ["user", "project"] to load Skills/CLAUDE.md from filesystem.
mcpServers Record<string, McpServerConfig> MCP server configurations.
agents Record<string, AgentDefinition> Named subagent definitions.
hooks object Lifecycle hook callbacks.

Built-in Tools

Tool names for allowedTools:

  • File ops: Read, Write, Edit, MultiEdit
  • Search: Glob, Grep
  • Execution: Bash
  • Web: WebSearch, WebFetch
  • Delegation: Task (required for subagents)
  • Skills: Skill (requires settingSources)

Custom Tools via SDK MCP Server

Define in-process tools without a separate MCP server process:

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

const searchOrders = tool(
  "search_orders",
  "Search orders by customer ID",
  { customer_id: "string", status: "string" },
  async (args) => {
    const results = await db.queryOrders(args.customer_id, args.status);
    return { content: [{ type: "text", text: JSON.stringify(results) }] };
  }
);

const sendEmail = tool(
  "send_email",
  "Send an email notification",
  { to: "string", subject: "string", body: "string" },
  async (args) => {
    await emailService.send(args.to, args.subject, args.body);
    return { content: [{ type: "text", text: `Email sent to ${args.to}` }] };
  }
);

const server = createSdkMcpServer({
  name: "business-tools",
  tools: [searchOrders, sendEmail],
});

for await (const msg of query({
  prompt: "Find recent orders for customer C-123",
  options: {
    mcpServers: { biz: server },
    allowedTools: ["mcp__biz__search_orders", "mcp__biz__send_email"],
  },
})) {
  console.log(msg);
}

Tool naming convention: MCP tools are accessed as mcp__<server-name>__<tool-name>.

Subagents

Delegate specialized tasks to isolated agents with their own context and tool permissions:

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

for await (const message of query({
  prompt: "Review auth module for security issues, then write tests",
  options: {
    allowedTools: ["Read", "Grep", "Glob", "Task"], // Task is required
    agents: {
      "security-reviewer": {
        description: "Security specialist. Use for vulnerability analysis.",
        prompt: "You are a security expert. Analyze code for OWASP Top 10...",
        tools: ["Read", "Grep", "Glob"],
        model: "opus",
      },
      "test-writer": {
        description: "Test specialist. Use to generate test suites.",
        prompt: "You are a testing expert. Write comprehensive unit tests...",
        tools: ["Read", "Write", "Bash"],
        model: "sonnet",
      },
    },
  },
})) {
  if (message.type === "result") console.log(message.result);
}

Factory pattern for dynamic agents:

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

function createReviewer(language: string): AgentDefinition {
  return {
    description: `${language} code review specialist`,
    prompt: `You are an expert ${language} developer...`,
    tools: ["Read", "Grep", "Glob"],
    model: ["rust", "c++"].includes(language) ? "opus" : "sonnet",
  };
}

Hooks — Lifecycle Callbacks

Available events: PreToolUse, PostToolUse, Stop, SessionStart, SessionEnd, UserPromptSubmit.

import { query, type HookCallback } from "@anthropic-ai/claude-agent-sdk";
import { appendFileSync } from "node:fs";

const blockDangerousCommands: HookCallback = async (input) => {
  if (input.tool_name === "Bash") {
    const cmd = input.tool_input?.command ?? "";
    const dangers = ["rm -rf /", "DROP TABLE", "mkfs"];
    if (dangers.some((d) => cmd.includes(d))) {
      return {
        hookSpecificOutput: {
          hookEventName: "PreToolUse",
          permissionDecision: "deny",
          permissionDecisionReason: `Blocked dangerous command: ${cmd}`,
        },
      };
    }
  }
  return {};
};

const auditLog: HookCallback = async (input) => {
  appendFileSync(
    "audit.log",
    `${new Date().toISOString()}: ${input.tool_name}: ${JSON.stringify(input.tool_input)}\n`
  );
  return {};
};

for await (const msg of query({
  prompt: "Refactor utils.ts",
  options: {
    permissionMode: "acceptEdits",
    hooks: {
      PreToolUse: [
        { matcher: "Bash", hooks: [blockDangerousCommands] },
        { matcher: ".*", hooks: [auditLog] },
      ],
    },
  },
})) {
  if (message.type === "result") console.log(msg.result);
}

MCP Integration (External Servers)

for await (const msg of query({
  prompt: "List open issues in the repo",
  options: {
    mcpServers: {
      github: {
        type: "stdio",
        command: "npx",
        args: ["-y", "@modelcontextprotocol/server-github"],
        env: { GITHUB_TOKEN: process.env.GITHUB_TOKEN! },
      },
      postgres: {
        type: "stdio",
        command: "docker",
        args: ["run", "mcp-postgres-server"],
        env: { DATABASE_URL: process.env.DATABASE_URL! },
      },
    },
    allowedTools: ["mcp__github", "mcp__postgres"],
  },
})) {
  console.log(msg);
}

You can mix SDK MCP servers (in-process) and external MCP servers in the same config.

Using Skills in the SDK

Skills are filesystem-based and must be explicitly enabled:

for await (const msg of query({
  prompt: "Help me process this PDF",
  options: {
    cwd: "/path/to/project",
    settingSources: ["user", "project"], // REQUIRED — loads Skills from filesystem
    allowedTools: ["Skill", "Read", "Write", "Bash"],
  },
})) {
  console.log(msg);
}

Common mistake: forgetting settingSources. Without it, Skills won't be discovered even if "Skill" is in allowedTools.

Skill locations:

  • Project: .claude/skills/*/SKILL.md (shared via git)
  • User: ~/.claude/skills/*/SKILL.md (personal, cross-project)

Note: The allowed-tools field in SKILL.md frontmatter only works in Claude Code CLI, not in the SDK. Use allowedTools in options to control tool access.

Sessions and Conversation Management

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

let sessionId: string | undefined;

// First interaction — capture sessionId
for await (const msg of query({
  prompt: "Review this codebase and identify the top 3 issues",
  options: { allowedTools: ["Read", "Glob", "Grep"] },
})) {
  if (msg.type === "system" && "session_id" in msg) {
    sessionId = msg.session_id;
  }
  console.log(msg);
}

// Resume with context
for await (const msg of query({
  prompt: "Now fix issue #1 that you found",
  options: {
    sessionId,
    allowedTools: ["Read", "Edit", "Bash"],
    permissionMode: "acceptEdits",
  },
})) {
  console.log(msg);
}

System Prompt Configuration

// 1. Custom system prompt (v0.1.0+ default: minimal prompt)
const options = { systemPrompt: "You are a senior TypeScript engineer..." };

// 2. Claude Code's full system prompt (opt-in)
const options = {
  systemPrompt: { type: "preset", preset: "claude_code" },
};

// 3. No system prompt — SDK default (minimal)
const options = {}; // uses minimal built-in prompt

Breaking change in v0.1.0: The SDK no longer loads Claude Code's system prompt by default. If you need the old behavior, explicitly set preset: "claude_code".

Authentication

# Direct API (default)
export ANTHROPIC_API_KEY=your-api-key

# Amazon Bedrock
export CLAUDE_CODE_USE_BEDROCK=1
# + configure AWS credentials

# Google Vertex AI
export CLAUDE_CODE_USE_VERTEX=1
# + configure GCP credentials

# Microsoft Azure AI Foundry
export CLAUDE_CODE_USE_FOUNDRY=1
# + configure Azure credentials

Common Patterns

Batch Processing (Parallel Agents)

async function processFile(filepath: string): Promise<string | undefined> {
  for await (const msg of query({
    prompt: `Review ${filepath} for security issues`,
    options: {
      allowedTools: ["Read", "Grep"],
      maxTurns: 50,
    },
  })) {
    if (msg.type === "result") return msg.result;
  }
}

const results = await Promise.all([
  processFile("auth.ts"),
  processFile("payments.ts"),
  processFile("users.ts"),
]);

Structured Output Collection

const messages: Message[] = [];
for await (const msg of query({ prompt: "Analyze this codebase", options })) {
  messages.push(msg);
}

// Extract final result
const result = messages.findLast((m) => "result" in m)?.result;

Error Handling

import { CLINotFoundError, CLIConnectionError } from "@anthropic-ai/claude-agent-sdk";

try {
  for await (const msg of query({ prompt: "...", options })) {
    console.log(msg);
  }
} catch (error) {
  if (error instanceof CLINotFoundError) {
    console.error("Claude Code CLI not found. Install: curl -fsSL https://claude.ai/install.sh | bash");
  } else if (error instanceof CLIConnectionError) {
    console.error(`Connection error: ${error.message}`);
  } else {
    throw error;
  }
}

Migration from Claude Code SDK (< v0.1.0)

Key changes:

  1. @anthropic-ai/claude-code → @anthropic-ai/claude-agent-sdk
  2. ClaudeCodeOptions → ClaudeAgentOptions (type name)
  3. System prompt no longer loads Claude Code's prompt by default
  4. settingSources must be explicitly set (was auto-loaded before)

Best Practices

  1. Principle of least privilege: Only grant tools the agent actually needs.
  2. Use permissionMode: "acceptEdits" for automation, "default" for interactive use.
  3. Prefer SDK MCP servers over external ones for custom tools — less overhead, easier debugging.
  4. Use subagents for specialized tasks — isolate context and apply the right model per task.
  5. Add hooks for safety — block dangerous commands and audit tool usage in production.
  6. Set maxTurns to prevent runaway agents in production environments.
  7. Use cwd to scope the agent to a specific directory.
  8. Capture sessionId from system messages if you need conversation continuity.

For troubleshooting common issues, see references/troubleshooting.md.

Official Resources

Weekly Installs
5
Repository
waltersumbon/cl…dk-skill
GitHub Stars
5
First Seen
Feb 24, 2026
Security Audits
Gen Agent Trust HubPass
SocketFail
SnykWarn
Installed on
gemini-cli5
claude-code5
github-copilot5
codex5
kimi-cli5
cursor5