Claude Agent SDK — Python Guide

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

Naming: The Claude Code SDK was renamed to the Claude Agent SDK (v0.1.0+). Package: pip install claude-agent-sdk · Import: from claude_agent_sdk import ...

Message Types

All message types for type checking and isinstance() checks:

from claude_agent_sdk import (
    AssistantMessage,     # Claude's text/tool responses
    ResultMessage,        # Final result with subtype (success/error/cancelled)
    SystemMessage,        # System events (session_id, etc.)
    UserMessage,          # User prompts
    ToolUseMessage,       # Tool invocation requests
    ToolResultMessage,    # Tool execution results
)

Quick Reference — Two Interaction Modes

1. query() — Stateless, One-Shot

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

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

async def main():
    async for message in query(
        prompt="Review utils.py for bugs. Fix any issues you find.",
        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 hasattr(block, "name"):
                    print(f"Tool: {block.name}")
        elif isinstance(message, ResultMessage):
            print(f"Done: {message.subtype}")

asyncio.run(main())

2. ClaudeSDKClient — Stateful, Multi-Turn

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

from claude_agent_sdk import ClaudeSDKClient, ClaudeAgentOptions

async with ClaudeSDKClient(
    options=ClaudeAgentOptions(
        allowed_tools=["Read", "Write", "Bash"],
        permission_mode="acceptEdits",
    )
) as client:
    await client.query("Analyze the codebase structure")
    async for msg in client.receive_messages():
        print(msg)
    # Continue the conversation with context preserved
    await client.query("Now refactor the largest file you found")
    async for msg in client.receive_messages():
        print(msg)

ClaudeAgentOptions — Complete Configuration

All options are optional. Key fields:

Field	Type	Description
allowed_tools	list[str] | None	Tools Claude can use. See Built-in Tools below.
disallowed_tools	list[str] | None	Explicitly block specific tools.
permission_mode	"default" | "acceptEdits" | "bypassPermissions"	Permission strategy.
system_prompt	str | dict | None	Custom instructions. Use {"type": "preset", "preset": "claude_code"} for CC default.
model	str | None	e.g. "sonnet", "opus", "haiku", or full model string.
cwd	str | Path | None	Working directory for the agent.
max_turns	int | None	Maximum agentic loop iterations.
setting_sources	list[Literal["user", "project"]] | None	["user", "project"] to load Skills/CLAUDE.md from filesystem.
mcp_servers	dict | None	MCP server configurations.
agents	dict[str, AgentDefinition] | None	Named subagent definitions.
hooks	dict | None	Lifecycle hook callbacks.

Built-in Tools

Tool names for allowed_tools:

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

Custom Tools via SDK MCP Server

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

from claude_agent_sdk import tool, create_sdk_mcp_server, ClaudeAgentOptions

@tool("search_orders", "Search orders by customer ID", {"customer_id": str, "status": str})
async def search_orders(args):
    results = await db.query_orders(args["customer_id"], args.get("status"))
    return {"content": [{"type": "text", "text": json.dumps(results)}]}

@tool("send_email", "Send an email notification", {"to": str, "subject": str, "body": str})
async def send_email(args):
    await email_service.send(args["to"], args["subject"], args["body"])
    return {"content": [{"type": "text", "text": f"Email sent to {args['to']}"}]}

server = create_sdk_mcp_server(name="business-tools", tools=[search_orders, send_email])

options = ClaudeAgentOptions(
    mcp_servers={"biz": server},
    allowed_tools=["mcp__biz__search_orders", "mcp__biz__send_email"],
)

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:

from claude_agent_sdk import query, ClaudeAgentOptions, AgentDefinition

async for message in query(
    prompt="Review auth module for security issues, then write tests",
    options=ClaudeAgentOptions(
        allowed_tools=["Read", "Grep", "Glob", "Task"],  # Task is required
        agents={
            "security-reviewer": AgentDefinition(
                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": AgentDefinition(
                description="Test specialist. Use to generate test suites.",
                prompt="You are a testing expert. Write comprehensive unit tests...",
                tools=["Read", "Write", "Bash"],
                model="sonnet",
            ),
        },
    ),
):
    print(message)

Factory pattern for dynamic agents:

def create_reviewer(language: str) -> AgentDefinition:
    return AgentDefinition(
        description=f"{language} code review specialist",
        prompt=f"You are an expert {language} developer...",
        tools=["Read", "Grep", "Glob"],
        model="opus" if language in ["rust", "c++"] else "sonnet",
    )

Hooks — Lifecycle Callbacks

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

from claude_agent_sdk import query, ClaudeAgentOptions, HookMatcher

async def block_dangerous_commands(input_data, tool_use_id, context):
    """Block destructive bash commands."""
    if input_data.get("tool_name") == "Bash":
        cmd = input_data.get("tool_input", {}).get("command", "")
        if any(danger in cmd for danger in ["rm -rf /", "DROP TABLE", "mkfs"]):
            return {
                "hookSpecificOutput": {
                    "hookEventName": "PreToolUse",
                    "permissionDecision": "deny",
                    "permissionDecisionReason": f"Blocked dangerous command: {cmd}",
                }
            }
    return {}

async def audit_log(input_data, tool_use_id, context):
    """Log all tool usage for auditing."""
    with open("audit.log", "a") as f:
        f.write(f"{input_data.get('tool_name')}: {input_data.get('tool_input')}\n")
    return {}

options = ClaudeAgentOptions(
    hooks={
        "PreToolUse": [
            HookMatcher(matcher="Bash", hooks=[block_dangerous_commands]),
            HookMatcher(matcher=".*", hooks=[audit_log]),
        ]
    }
)

MCP Integration (External Servers)

options = ClaudeAgentOptions(
    mcp_servers={
        "github": {
            "type": "stdio",
            "command": "npx",
            "args": ["-y", "@modelcontextprotocol/server-github"],
            "env": {"GITHUB_TOKEN": os.getenv("GITHUB_TOKEN")},
        },
        "postgres": {
            "type": "stdio",
            "command": "docker",
            "args": ["run", "mcp-postgres-server"],
            "env": {"DATABASE_URL": os.getenv("DATABASE_URL")},
        },
    },
    allowed_tools=["mcp__github", "mcp__postgres"],
)

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:

options = ClaudeAgentOptions(
    cwd="/path/to/project",
    setting_sources=["user", "project"],  # REQUIRED — loads Skills from filesystem
    allowed_tools=["Skill", "Read", "Write", "Bash"],
)

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

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 allowed_tools in ClaudeAgentOptions to control tool access.

Sessions and Conversation Management

from claude_agent_sdk import query, ClaudeAgentOptions

session_id = None

# First interaction — capture session_id
async for message in query(
    prompt="Review this codebase and identify the top 3 issues",
    options=ClaudeAgentOptions(allowed_tools=["Read", "Glob", "Grep"]),
):
    if message.type == "system" and hasattr(message, "session_id"):
        session_id = message.session_id
    print(message)

# Resume with context
async for message in query(
    prompt="Now fix issue #1 that you found",
    options=ClaudeAgentOptions(
        session_id=session_id,
        allowed_tools=["Read", "Edit", "Bash"],
        permission_mode="acceptEdits",
    ),
):
    print(message)

System Prompt Configuration

# 1. Custom system prompt (v0.1.0+ default: minimal prompt)
options = ClaudeAgentOptions(system_prompt="You are a senior Python engineer...")

# 2. Claude Code's full system prompt (opt-in)
options = ClaudeAgentOptions(
    system_prompt={"type": "preset", "preset": "claude_code"}
)

# 3. No system prompt — SDK default (minimal)
options = ClaudeAgentOptions()  # 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)

import asyncio

async def process_file(filepath: str):
    async for msg in query(
        prompt=f"Review {filepath} for security issues",
        options=ClaudeAgentOptions(
            allowed_tools=["Read", "Grep"],
            max_turns=50,
        ),
    ):
        if isinstance(msg, ResultMessage):
            return msg.result

results = await asyncio.gather(
    process_file("auth.py"),
    process_file("payments.py"),
    process_file("users.py"),
)

Structured Output Collection

messages = []
async for msg in query(prompt="Analyze this codebase", options=options):
    messages.append(msg)

# Extract final result
result = next((m for m in reversed(messages) if hasattr(m, "result")), None)

Error Handling

Available exception types:

• CLINotFoundError — Claude Code CLI binary not found in PATH
• CLIConnectionError — Failed to connect to CLI process
• Standard Python exceptions (e.g., asyncio.TimeoutError) may also be raised

from claude_agent_sdk import CLINotFoundError, CLIConnectionError

try:
    async for msg in query(prompt="...", options=options):
        print(msg)
except CLINotFoundError:
    print("Claude Code CLI not found. Install: curl -fsSL https://claude.ai/install.sh | bash")
except CLIConnectionError as e:
    print(f"Connection error: {e}")

Migration from Claude Code SDK (< v0.1.0)

Key changes:

1. ClaudeCodeOptions → ClaudeAgentOptions
2. System prompt no longer loads Claude Code's prompt by default
3. setting_sources must be explicitly set (was auto-loaded before)
4. Import path unchanged (claude_agent_sdk), but class names changed

Best Practices

1. Principle of least privilege: Only grant tools the agent actually needs.
2. Use permission_mode="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 max_turns to prevent runaway agents in production environments.
7. Use cwd to scope the agent to a specific directory.
8. Capture session_id from system messages if you need conversation continuity.

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

Official Resources

• Overview: https://platform.claude.com/docs/en/agent-sdk/overview
• Quickstart: https://platform.claude.com/docs/en/agent-sdk/quickstart
• Python reference: https://platform.claude.com/docs/en/agent-sdk/python
• Migration guide: https://platform.claude.com/docs/en/agent-sdk/migration-guide
• Python SDK repo: https://github.com/anthropics/claude-agent-sdk-python
• Demo agents: https://github.com/anthropics/claude-agent-sdk-demos
• Cookbook: https://platform.claude.com/cookbook