Claude Agent Sdk

Claude Agent SDK

Build AI agents that autonomously read files, run commands, search the web, edit code, and more. The Agent SDK gives you the same tools and agent loop that power Claude Code, programmable in Python and TypeScript.

Renamed: The Claude Code SDK is now the Claude Agent SDK. Package names changed to claude-agent-sdk (Python) and @anthropic-ai/claude-agent-sdk (TypeScript).

Quick Start

Install

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

# Python (uv)
uv init && uv add claude-agent-sdk

# Python (pip)
pip install claude-agent-sdk

The TypeScript SDK bundles a native Claude Code binary — no separate install needed.

Set API Key

export ANTHROPIC_API_KEY=your-api-key

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

First Agent

# Python
import asyncio
from claude_agent_sdk import query, ClaudeAgentOptions, AssistantMessage, ResultMessage

async def main():
    async for message in query(
        prompt="Find and fix bugs in utils.py",
        options=ClaudeAgentOptions(
            allowed_tools=["Read", "Edit", "Glob"],
            permission_mode="acceptEdits",
        ),
    ):
        if isinstance(message, AssistantMessage):
            for block in message.content:
                if hasattr(block, "text"):
                    print(block.text)
        elif isinstance(message, ResultMessage):
            print(f"Done: {message.subtype}")

asyncio.run(main())

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

for await (const message of query({
  prompt: "Find and fix bugs in utils.ts",
  options: {
    allowedTools: ["Read", "Edit", "Glob"],
    permissionMode: "acceptEdits"
  }
})) {
  if (message.type === "assistant") {
    for (const block of message.message.content) {
      if ("text" in block) console.log(block.text);
    }
  } else if (message.type === "result") {
    console.log(`Done: ${message.subtype}`);
  }
}

Core Concepts

The agent loop: Receive prompt → Evaluate → Call tools → Observe results → Repeat → Return result.

Each round-trip (Claude calls tools, SDK executes, results feed back) is one turn. The loop ends when Claude responds with text only (no tool calls), yielding a ResultMessage.

Choosing an Entry Point

TypeScript uses query() with continue: true for multi-turn. For a session-holding client, see the TypeScript V2 preview.

# Python multi-turn with ClaudeSDKClient
async with ClaudeSDKClient(options=options) as client:
    await client.query("Analyze the auth module")
    async for msg in client.receive_response():
        print_response(msg)

    await client.query("Now refactor it to use JWT")
    async for msg in client.receive_response():
        print_response(msg)

Configuration

Key options for ClaudeAgentOptions (Python) / Options (TypeScript):

For full type definitions, read references/typescript-api.md or references/python-api.md.

Message Handling

The SDK streams messages as the agent works. Key types:

Checking ResultMessage

Always check subtype before reading result:

# Python
if isinstance(message, ResultMessage):
    if message.subtype == "success":
        print(message.result)
        print(f"Cost: ${message.total_cost_usd:.4f}")
    elif message.subtype == "error_max_turns":
        print(f"Hit turn limit. Resume: {message.session_id}")
    elif message.subtype == "error_max_budget_usd":
        print("Hit budget limit.")

// TypeScript
if (message.type === "result") {
  if (message.subtype === "success") {
    console.log(message.result);
    console.log(`Cost: $${message.total_cost_usd.toFixed(4)}`);
  } else if (message.subtype === "error_max_turns") {
    console.log(`Hit turn limit. Resume: ${message.session_id}`);
  }
}

Result subtypes: success, error_max_turns, error_max_budget_usd, error_during_execution, error_max_structured_output_retries.

Permission Modes

Evaluation order: Hooks → Deny rules → Permission mode → Allow rules → canUseTool callback.

allowedTools auto-approves listed tools — it does NOT prevent Claude from calling other tools. Tools not listed fall through to permissionMode and canUseTool. To prevent a tool from running at all (including in bypassPermissions), use disallowedTools.

Built-in Tools

Hooks

Hooks are callback functions that intercept agent behavior at key execution points. Register them in options.hooks as a dict/object mapping event names to arrays of matchers.

Hook Events

Hook Pattern

# Python
async def protect_env(input_data, tool_use_id, context):
    file_path = input_data["tool_input"].get("file_path", "")
    if file_path.endswith(".env"):
        return {
            "hookSpecificOutput": {
                "hookEventName": input_data["hook_event_name"],
                "permissionDecision": "deny",
                "permissionDecisionReason": "Cannot modify .env files",
            }
        }
    return {}

options = ClaudeAgentOptions(
    hooks={"PreToolUse": [HookMatcher(matcher="Write|Edit", hooks=[protect_env])]}
)

// TypeScript
const protectEnv: HookCallback = async (input, toolUseID, { signal }) => {
  const preInput = input as PreToolUseHookInput;
  const filePath = (preInput.tool_input as any)?.file_path as string;
  if (filePath?.endsWith(".env")) {
    return {
      hookSpecificOutput: {
        hookEventName: preInput.hook_event_name,
        permissionDecision: "deny",
        permissionDecisionReason: "Cannot modify .env files"
      }
    };
  }
  return {};
};

const options = {
  hooks: { PreToolUse: [{ matcher: "Write|Edit", hooks: [protectEnv] }] }
};

Hooks can: block tools (permissionDecision: "deny"), modify inputs (updatedInput), inject context (systemMessage), auto-approve (permissionDecision: "allow"), or run async side effects ({ async: true }).

For full hook API and recipes, read references/hooks-guide.md.

Subagents

Spawn specialized agents with isolated context windows. Define them programmatically via the agents option:

# Python
options = ClaudeAgentOptions(
    allowed_tools=["Read", "Grep", "Glob", "Agent"],
    agents={
        "code-reviewer": AgentDefinition(
            description="Expert code reviewer for security and quality.",
            prompt="Analyze code quality. Be thorough but concise.",
            tools=["Read", "Grep", "Glob"],
            model="sonnet",
        ),
    },
)

// TypeScript
const options = {
  allowedTools: ["Read", "Grep", "Glob", "Agent"],
  agents: {
    "code-reviewer": {
      description: "Expert code reviewer for security and quality.",
      prompt: "Analyze code quality. Be thorough but concise.",
      tools: ["Read", "Grep", "Glob"],
      model: "sonnet"
    }
  }
};

Key AgentDefinition fields: description (when to use), prompt (system prompt), tools (restrict tools), model ("sonnet" / "opus" / "haiku" / "fable" / "inherit" / full model ID), skills (preload named skills), maxTurns, background, effort, permissionMode.

As of Claude Code v2.1.172, subagents can spawn their own subagents up to 5 levels deep. The 5th-level subagent cannot spawn further. Include Agent in a subagent's tools only when nesting is intentional.

Include Agent in the parent's allowedTools. Claude auto-delegates based on descriptions, or use explicit prompting: "Use the code-reviewer agent to..."

Tracking: Messages from inside a subagent include parent_tool_use_id, letting you identify which subagent produced each message.

Resuming: Subagents can be resumed by capturing the session_id and agentId from the first run, then passing resume=session_id with the agent ID in the follow-up prompt. Subagent transcripts persist independently and survive main conversation compaction.

Custom Tools

Define your own functions Claude can call using in-process MCP servers:

# Python
from claude_agent_sdk import tool, create_sdk_mcp_server

@tool("get_temperature", "Get current temperature", {"latitude": float, "longitude": float})
async def get_temperature(args):
    # fetch weather data...
    return {"content": [{"type": "text", "text": f"Temperature: {temp}°F"}]}

weather = create_sdk_mcp_server(name="weather", tools=[get_temperature])
options = ClaudeAgentOptions(
    mcp_servers={"weather": weather},
    allowed_tools=["mcp__weather__get_temperature"],
)

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

const getTemp = tool("get_temperature", "Get current temperature", {
  latitude: z.number(), longitude: z.number()
}, async (args) => {
  // fetch weather data...
  return { content: [{ type: "text", text: `Temperature: ${temp}°F` }] };
});

const weather = createSdkMcpServer({ name: "weather", tools: [getTemp] });
const options = {
  mcpServers: { weather },
  allowedTools: ["mcp__weather__get_temperature"]
};

Tool naming: mcp__{server_name}__{tool_name}. Use mcp__server__* wildcards in allowedTools.

For schemas, error handling, annotations, and images, read references/custom-tools.md.

MCP Servers (External)

Connect to external tools via MCP — databases, browsers, APIs:

# stdio (local process)
mcp_servers={"github": {"command": "npx", "args": ["-y", "@modelcontextprotocol/server-github"], "env": {"GITHUB_TOKEN": token}}}

# HTTP/SSE (remote)
mcp_servers={"api": {"type": "sse", "url": "https://api.example.com/mcp/sse", "headers": {"Authorization": f"Bearer {token}"}}}

Transport types: stdio (local), sse (streaming remote), http (non-streaming remote), sdk (in-process).

Always add MCP tools to allowedTools — they're available but not auto-approved by default. permissionMode: "acceptEdits" does NOT auto-approve MCP tools.

Sessions

Capture session ID from ResultMessage.session_id. Sessions persist conversation, not filesystem — use file checkpointing for file snapshots.

Sessions are stored at ~/.claude/projects/<encoded-cwd>/<session-id>.jsonl. The cwd must match when resuming.

User Input

Tool Approval (canUseTool)

Handle permission prompts programmatically:

async def can_use_tool(tool_name, input_data, context):
    if approved:
        return PermissionResultAllow(updated_input=input_data)
    return PermissionResultDeny(message="User rejected")

options = ClaudeAgentOptions(can_use_tool=can_use_tool)

Clarifying Questions (AskUserQuestion)

Claude can ask multiple-choice questions. Check tool_name == "AskUserQuestion" in your canUseTool callback, display questions, collect answers, return them:

return PermissionResultAllow(updated_input={
    "questions": input_data["questions"],
    "answers": {"How should I format output?": "Summary"},
})

Include AskUserQuestion in your tools array if you restrict tools.

Streaming Input

Pass an AsyncIterable as prompt (instead of a string) for interactive apps:

# Python
async def messages():
    yield {"type": "user", "message": {"role": "user", "content": "Analyze auth.py"}}
    yield {"type": "user", "message": {"role": "user", "content": [
        {"type": "text", "text": "Check this diagram"},
        {"type": "image", "source": {"type": "base64", "media_type": "image/png", "data": b64}},
    ]}}

async with ClaudeSDKClient(options) as client:
    await client.query(messages())
    async for msg in client.receive_response(): print(msg)

// TypeScript
async function* messages() {
  yield { type: "user" as const, message: { role: "user" as const, content: "Analyze auth.ts" } };
}
for await (const msg of query({ prompt: messages(), options })) { /* ... */ }

Streaming input enables: image attachments, queued messages, mid-task interruption, and natural multi-turn. Single message mode (string prompt) doesn't support images or interruption.

Hosting & Deployment

Session Patterns

Requirements

• Python 3.10+ or Node.js 18+
• 1 GiB RAM, 5 GiB disk, 1 CPU (minimum)
• Outbound HTTPS to api.anthropic.com

Cross-Host Session Persistence

Use the sessionStore option to mirror transcripts to S3, Redis, or Postgres, so sessions survive container restarts. Required for the Hybrid pattern. SessionStore mirrors transcripts only — not CLAUDE.md or working-directory artifacts.

Multi-tenant Isolation

In shared containers, pass settingSources: [], set CLAUDE_CODE_DISABLE_AUTO_MEMORY=1, use a per-tenant CLAUDE_CONFIG_DIR, and pass an explicit cwd per tenant. See references/patterns.md for the full code pattern.

Sandbox Providers

Modal, Cloudflare Sandboxes, Daytona, E2B, Fly Machines, Vercel Sandbox.

TypeScript serverless note: The SDK bundles a ~230 MB native binary. It is generally not viable for AWS Lambda / Vercel Functions (250 MB limit). Use a long-running container. See references/bun-runtime.md.

For production patterns, cost tracking, and security hardening, read references/patterns.md.

Advisor Tool (Beta — Messages API)

Pair a fast executor model with a higher-intelligence advisor that consults mid-generation. The executor decides when to call the advisor; Anthropic runs the sub-inference server-side at no extra round-trip cost.

Best for: long-horizon agentic workloads (coding, computer use, multi-step research) where most turns are mechanical but a good plan is crucial.

Platform: Claude API and Claude Platform on AWS only. Not available on Bedrock, Google Cloud, or Foundry.

response = client.beta.messages.create(
    model="claude-sonnet-5",         # executor
    betas=["advisor-tool-2026-03-01"],
    tools=[{
        "type": "advisor_20260301",
        "name": "advisor",
        "model": "claude-opus-4-8",  # or "claude-fable-5" for encrypted advice
    }],
    messages=[...],
    max_tokens=4096,
)

Key facts:

• Tool type advisor_20260301, name must be "advisor"
• claude-fable-5 / claude-mythos-5 advisors return encrypted advisor_redacted_result (not ZDR eligible, 30-day retention). Use claude-opus-4-8 for plaintext.
• Advisor must be at least as capable as the executor
• Multi-turn: always include advisor_tool_result blocks in subsequent requests

Agent SDK query() compatibility: not confirmed as of July 2026 — spike before assuming it works via the SDK.

For parameters, result variants, model compatibility table, billing, and system prompt recipes, read references/advisor-tool.md.

Structured Output

Force Claude to return JSON matching a schema:

options = ClaudeAgentOptions(
    output_format={"type": "json_schema", "schema": {"type": "object", "properties": {"summary": {"type": "string"}}, "required": ["summary"]}},
)
# Access via message.structured_output on ResultMessage

File Checkpointing

Enable enable_file_checkpointing / enableFileCheckpointing to track file changes. Use rewindFiles(userMessageId) to undo changes since a specific message.

Pre-warming (startup() — TypeScript only)

const warm = await startup({ options: { maxTurns: 3 } });
const q = warm.query("What files are here?");
for await (const msg of q) { /* ... */ }
warm.close();

Critical Defaults (v0.1.0+)

Since v0.1.0, the SDK no longer behaves like the Claude Code CLI out of the box:

No system prompt by default. To get Claude Code's full system prompt:

• Python: system_prompt={"type": "preset", "preset": "claude_code"}
• TypeScript: systemPrompt: { type: "preset", preset: "claude_code" }

No settings loaded by default. CLAUDE.md, skills, slash commands, and settings.json are NOT read unless you opt in:

• Python: setting_sources=["user", "project", "local"]
• TypeScript: settingSources: ["user", "project", "local"]

Pass [] or omit entirely for fully isolated behavior — ideal for CI/CD and multi-tenant deployments.

Troubleshooting

Reference Files

Read these for deeper API details — they're loaded on-demand: