Claude Agent SDK Python

Build production-ready agent applications with the Claude Agent SDK for Python.

Quick Reference

Installation

pip install claude-agent-sdk

Requires Python 3.10+ and bundled Claude CLI v2.0.60+.

Core APIs

API	Use Case
query()	One-shot or simple multi-turn queries
ClaudeSDKClient	Interactive streaming, bidirectional control
@tool + create_sdk_mcp_server()	Custom in-process tools
can_use_tool callback	Fine-grained permission control
hooks	Lifecycle event interception

Basic Patterns

One-Shot Query

import anyio
from claude_agent_sdk import query, AssistantMessage, TextBlock, ResultMessage

async def main():
    async for msg in query(prompt="Explain Python decorators"):
        if isinstance(msg, AssistantMessage):
            for block in msg.content:
                if isinstance(block, TextBlock):
                    print(block.text)
        elif isinstance(msg, ResultMessage):
            print(f"Cost: ${msg.total_cost_usd:.4f}")

anyio.run(main)

Interactive Client

import anyio
from claude_agent_sdk import (
    ClaudeSDKClient,
    ClaudeAgentOptions,
    AssistantMessage,
    TextBlock,
    ResultMessage,
)

async def main():
    options = ClaudeAgentOptions(
        system_prompt="You are a helpful assistant",
        allowed_tools=["Read", "Write", "Bash"],
        permission_mode="acceptEdits",
        max_turns=10,
    )

    async with ClaudeSDKClient(options=options) as client:
        await client.query("Analyze main.py")

        async for msg in client.receive_response():
            if isinstance(msg, AssistantMessage):
                for block in msg.content:
                    if isinstance(block, TextBlock):
                        print(block.text)
            elif isinstance(msg, ResultMessage):
                print(f"Done. Cost: ${msg.total_cost_usd:.4f}")

anyio.run(main)

Custom Tools

from typing import Any
from claude_agent_sdk import tool, create_sdk_mcp_server, ClaudeAgentOptions

@tool("add", "Add two numbers", {"a": float, "b": float})
async def add(args: dict[str, Any]) -> dict[str, Any]:
    result = args["a"] + args["b"]
    return {"content": [{"type": "text", "text": f"Result: {result}"}]}

@tool("divide", "Divide two numbers", {"a": float, "b": float})
async def divide(args: dict[str, Any]) -> dict[str, Any]:
    if args["b"] == 0:
        return {
            "content": [{"type": "text", "text": "Error: Division by zero"}],
            "is_error": True
        }
    return {"content": [{"type": "text", "text": f"Result: {args['a'] / args['b']}"}]}

calculator = create_sdk_mcp_server(
    name="calculator",
    version="1.0.0",
    tools=[add, divide]
)

options = ClaudeAgentOptions(
    mcp_servers={"calc": calculator},
    allowed_tools=["mcp__calc__add", "mcp__calc__divide"],
    permission_mode="acceptEdits",
)

Configuration Reference

ClaudeAgentOptions

ClaudeAgentOptions(
    # Model
    model="claude-sonnet-4-20250514",
    fallback_model=None,

    # Prompts
    system_prompt="You are a helpful assistant",

    # Limits
    max_turns=10,
    max_budget_usd=1.00,
    max_thinking_tokens=8000,

    # Environment
    cwd="/path/to/project",
    env={"CUSTOM_VAR": "value"},

    # Tools
    tools=["Read", "Write", "Bash"],  # or {"type": "preset", "preset": "claude_code"}
    allowed_tools=["Read", "Write"],  # whitelist
    disallowed_tools=["WebFetch"],    # blacklist
    mcp_servers={"name": config},     # custom tools

    # Permissions
    permission_mode="acceptEdits",  # default, acceptEdits, plan, bypassPermissions
    can_use_tool=my_permission_callback,

    # Sessions
    continue_conversation=False,
    resume="session-id",
    fork_session=False,

    # Hooks
    hooks={"PreToolUse": [HookMatcher(matcher="*", hooks=[my_hook])]},
)

Permission Modes

Mode	Behavior
"default"	Prompt for all actions
"acceptEdits"	Auto-approve file edits
"plan"	Planning mode (no execution)
"bypassPermissions"	Auto-approve all (use with caution)

Advanced Patterns

For detailed examples of:

• Permission callbacks: See references/permissions.md
• Lifecycle hooks: See references/hooks.md
• Testing patterns: See references/testing.md
• Production patterns: See references/production.md

Independent App Testing

Use scripts/run_agent_app.py to test agent applications in isolation:

# Run a simple agent app
python scripts/run_agent_app.py path/to/app.py

# Run with custom options
python scripts/run_agent_app.py path/to/app.py --cwd /project --max-turns 5

Use scripts/create_agent_project.py to scaffold new agent projects:

# Create a basic agent project
python scripts/create_agent_project.py my_agent --template basic

# Create with custom tools
python scripts/create_agent_project.py my_agent --template custom-tools

# Create interactive chat app
python scripts/create_agent_project.py my_agent --template chat

Message Types

Type	Description
AssistantMessage	Claude's responses (text, thinking, tool use)
UserMessage	User input or tool results
ResultMessage	Session summary (always last)
SystemMessage	System events

Content Blocks

from claude_agent_sdk import TextBlock, ThinkingBlock, ToolUseBlock, ToolResultBlock

# In AssistantMessage.content
for block in msg.content:
    if isinstance(block, TextBlock):
        print(block.text)
    elif isinstance(block, ThinkingBlock):
        print(f"Thinking: {block.thinking}")
    elif isinstance(block, ToolUseBlock):
        print(f"Tool: {block.name}, Input: {block.input}")

Error Handling

from claude_agent_sdk import (
    ClaudeSDKError,
    CLINotFoundError,
    ProcessError,
    SDKJSONDecodeError,
)

try:
    async with ClaudeSDKClient(options) as client:
        await client.query("Do something")
        async for msg in client.receive_response():
            process(msg)
except CLINotFoundError:
    print("Claude CLI not found. Install from https://claude.ai")
except ProcessError as e:
    print(f"CLI failed with exit code: {e.exit_code}")
except SDKJSONDecodeError:
    print("Failed to parse CLI response")

Best Practices

1. Always use context managers for ClaudeSDKClient to ensure cleanup
2. Set max_budget_usd to prevent runaway costs
3. Use permission_mode="acceptEdits" for file operations without prompts
4. Return is_error: True from custom tools on failure
5. Use structured outputs for predictable JSON responses
6. Test tools independently before integrating with full agent