claude-agent-toolkit 0.1.4

Python framework for building Claude Code agents with custom tools

Claude Code Agent Toolkit (claude-agent-toolkit)

A Python framework for building Claude Code agents with custom tools, designed to leverage Claude Code's advanced reasoning capabilities with your subscription token. The framework provides Docker-isolated environments where Claude Code can orchestrate custom MCP tools for production workflows.

Key Features

• Claude Code Integration - Leverage Claude Code's advanced reasoning with your existing subscription token
• Docker Isolation - Complete isolation of agent execution environment with Claude Code CLI
• Explicit Data Management - Users control their own data without automatic state management
• CPU-bound Operations - Support for CPU-intensive operations with process pools and parallel execution
• Multi-tool Coordination - Claude Code orchestrates multiple tools in complex workflows
• Production Ready - Build scalable agents using Claude Code's capabilities with custom tool integration

Architecture

Core Components

• Agent Framework (src/claude_agent_toolkit/agent/) - Docker-isolated Agent class that runs Claude Code with MCP tool support
• MCP Tool Framework (src/claude_agent_toolkit/tool/) - BaseTool class for creating custom MCP tools with explicit data management
• Example Tools (src/examples/) - Demonstration tools showing practical agent development patterns
• Docker Environment (src/docker/) - Isolated environment with Claude Code CLI and dependencies

Quick Start

Prerequisites

• Python 3.12+ with uv package manager
• Docker Desktop (must be running)
• Claude Code OAuth Token - Get from Claude Code

Installation

# Using pip
pip install claude-agent-toolkit

# Using uv
uv add claude-agent-toolkit

# Using poetry
poetry add claude-agent-toolkit

# Set your OAuth token
export CLAUDE_CODE_OAUTH_TOKEN='your-token-here'

Run the Demo

# Clone the repository for examples
git clone https://github.com/cheolwanpark/claude-agent-toolkit.git
cd claude-agent-toolkit

# Start Docker Desktop first, then run the examples
# Calculator example (now includes parallel operations):
cd src/examples/calculator && python main.py
# Weather example:
cd src/examples/weather && python main.py

This will run demonstration examples:

1. Calculator Demo - Shows stateful operations, parallel processing (factorial, fibonacci, prime checking), and mathematical problem solving
2. Weather Demo - Demonstrates external API integration with real-time data and async operations

Tool Development

Creating Custom Tools

Create tools by inheriting from BaseTool and using the @tool() decorator:

from claude_agent_toolkit import BaseTool, tool

class MyTool(BaseTool):
    def __init__(self):
        super().__init__()  # Server starts automatically
        # Explicit data management - no automatic state management
        self.counter = 0
        self.operations = []
    
    @tool(description="Increment counter and return new value")
    async def increment(self) -> dict:
        self.counter += 1
        return {"value": self.counter}
    
    @tool(description="Heavy computation with parallel processing", parallel=True, timeout_s=120)  
    def compute_heavy(self, data: str) -> dict:
        # CPU-intensive operation runs under ProcessPoolExecutor
        # Note: ProcessPoolExecutor creates new instance, self.counter won't persist
        import time
        time.sleep(2)  # Simulate heavy computation
        return {"processed": f"Heavy result for {data}", "parallel_execution": True}

Context Manager Support

For explicit resource management, use the context manager pattern:

# Single tool with guaranteed cleanup
with MyTool(workers=2) as tool:
    agent = Agent(tools=[tool])
    result = await agent.run("Process my data")
# Server automatically cleaned up here

# Multiple tools in one statement
with MyTool() as calc_tool, WeatherTool() as weather_tool:
    agent = Agent(tools=[calc_tool, weather_tool])
    result = await agent.run("Calculate something and check weather")
# Both tools cleaned up automatically

# Parameters can be passed to constructor
with MyTool(host="127.0.0.1", port=8080, workers=4, log_level="INFO") as tool:
    # Tool server starts immediately with specified configuration
    agent = Agent(tools=[tool])
    result = await agent.run("Heavy computation task")
# Guaranteed cleanup even if exceptions occur

Using Tools with Agents

from claude_agent_toolkit import Agent, ConnectionError, ExecutionError

try:
    # Create tool (server starts automatically)
    my_tool = MyTool(workers=2)

    # New pattern (recommended) - cleaner initialization
    agent = Agent(
        system_prompt="You are a helpful assistant specialized in calculations",
        tools=[my_tool]
    )

    # Traditional pattern - still supported
    # agent = Agent()
    # agent.connect(my_tool)

    # Run agent with prompt (verbose=True shows detailed message processing)
    result = await agent.run(
        "Please increment the counter twice and tell me the result",
        verbose=True  # Shows detailed Claude Code interaction logs
    )
    print(f"Success: {result['success']}")
    print(f"Response: {result['response']}")

    # Verify tool was actually called
    print(f"Counter value: {my_tool.counter}")
    print(f"Operations count: {len(my_tool.operations)}")
    
except ConnectionError as e:
    print(f"Connection issue: {e}")
    # Handle Docker, network, or port binding problems
    
except ExecutionError as e:
    print(f"Execution failed: {e}")
    # Handle agent execution or tool failures

Model Selection

Choose the right Claude model for your agent's needs:

Available Models

• "haiku" - Fast and efficient for simple tasks
• "sonnet" - Balanced performance (good default choice)
• "opus" - Most capable for complex reasoning

Usage Examples

from claude_agent_toolkit import Agent

# Use fast Haiku model for simple tasks
weather_agent = Agent(
    system_prompt="You are a weather assistant",
    tools=[weather_tool],
    model="haiku"  # Fast, efficient for simple weather queries
)

# Use Sonnet for general-purpose tasks
general_agent = Agent(
    system_prompt="You are a helpful assistant", 
    tools=[calculator_tool, weather_tool],
    model="sonnet"  # Balanced performance
)

# Use Opus for complex analysis
analysis_agent = Agent(
    system_prompt="You are a data analyst",
    tools=[analysis_tool],
    model="opus"  # Maximum reasoning capability
)

# Override model for specific queries
result = await weather_agent.run(
    "Complex weather pattern analysis for next month",
    model="opus"  # Use more capable model for this specific task
)

# Full model IDs also work
agent = Agent(model="claude-3-5-haiku-20241022")

When to Use Each Model

• Haiku: Simple queries, basic operations, fast responses needed
• Sonnet: General purpose tasks, good balance of speed and capability
• Opus: Complex reasoning, detailed analysis, maximum quality needed

Why Claude Code Agents?

Unlike generic agent frameworks, this toolkit specifically leverages Claude Code's unique capabilities:

1. Advanced Reasoning - Use Claude Code's sophisticated decision-making in your agents
2. Existing Subscription - Build production agents with your current Claude Code subscription
3. Stateful Workflows - Claude Code builds context across multiple tool interactions
4. Intelligent Orchestration - Claude Code decides which tools to use and when
5. Production Infrastructure - Leverage Claude's robust infrastructure for your agents

Example: Intelligent Workflow

# Claude Code analyzes data with one tool, then decides to process it with another
# The agent maintains context and makes intelligent decisions about tool usage
# Your tools provide capabilities, Claude Code provides the intelligence

API Reference

Agent Class

class Agent:
    def __init__(                                          # Initialize agent
        self,
        oauth_token: Optional[str] = None,                 # Your Claude Code token
        system_prompt: Optional[str] = None,               # Custom agent behavior
        tools: Optional[List[BaseTool]] = None,            # Tools to connect automatically
        model: Optional[Union[Literal["opus", "sonnet", "haiku"], str]] = None  # Model selection
    )
    def connect(self, tool: BaseTool) -> 'Agent'           # Connect custom tools  
    async def run(                                         # Run Claude Code with tools
        self,
        prompt: str,                                       # Instruction for Claude
        verbose: bool = False,                             # Show detailed processing logs
        model: Optional[Union[Literal["opus", "sonnet", "haiku"], str]] = None  # Override model
    ) -> Dict[str, Any]

BaseTool Class

class BaseTool:
    def __init__(self, host="127.0.0.1", port=None, *, workers=None, log_level="ERROR")
    @property def connection_url(self) -> str  # Always accessible after construction
    @property def health_url(self) -> str      # Always accessible after construction
    def __enter__(self) -> 'BaseTool'          # Context manager support
    def __exit__(self, exc_type, exc_val, exc_tb) -> bool
    def __del__(self)                          # Automatic cleanup on destruction

@tool() Decorator

@tool(
    name: Optional[str] = None,           # Tool method name
    description: str = "",               # Method description  
    parallel: bool = False,              # Use process pool
    timeout_s: int = 60,                 # Timeout for parallel operations
)

Exception Classes

Claude Agent Toolkit provides specific exception types for clear error handling:

# Import exception classes
from claude_agent_toolkit import (
    ClaudeAgentError,     # Base exception for all library errors
    ConfigurationError,   # Missing OAuth tokens, invalid configuration
    ConnectionError,      # Docker, network, port binding failures  
    ExecutionError,       # Agent execution, tool failures, timeouts
)

# Exception hierarchy
ClaudeAgentError
├── ConfigurationError    # Configuration issues
├── ConnectionError       # Network/service connectivity
└── ExecutionError       # Runtime execution failures

When to catch each exception:

• ConfigurationError: Handle setup issues, missing tokens, invalid configs
• ConnectionError: Handle Docker, network, and port binding failures
• ExecutionError: Handle runtime failures, timeouts, tool execution issues
• ClaudeAgentError: Catch all library errors with a single handler

Development Workflow

1. Start Docker Desktop

Required for agent execution - must be running before creating Claude Code agents.

2. Set OAuth Token

export CLAUDE_CODE_OAUTH_TOKEN='your-token-here'

3. Create Custom Tools

Inherit from BaseTool and implement @tool methods that extend Claude Code's capabilities.

4. Build Your Agent

Use the examples in src/examples/ to see demonstrations or create custom agent scripts.

5. Deploy to Production

Use your Claude Code subscription to run agents at scale with custom tool integration.

Dependencies

Runtime Dependencies

• docker>=7.1.0 - Docker container management

• fastmcp>=2.11.3 - MCP server framework

• httpx>=0.28.1 - HTTP client for health checks

• uvicorn>=0.35.0 - ASGI server for MCP HTTP endpoints

Docker Environment

• Python 3.11 with Claude Code SDK
• Node.js 20 with Claude Code CLI
• Non-root user execution for security

Error Handling

Claude Agent Toolkit uses specific exception types to help you handle errors gracefully:

from claude_agent_toolkit import (
    Agent, BaseTool, tool,
    ClaudeAgentError, ConfigurationError, ConnectionError,
    ExecutionError
)

# Handle specific error types
try:
    agent = Agent(
        oauth_token="your-token",
        tools=[MyTool()]
    )
    result = await agent.run("Process my request")
    
except ConfigurationError as e:
    print(f"Configuration issue: {e}")
    # Handle missing OAuth token, invalid tool config
    
except ConnectionError as e:
    print(f"Connection failed: {e}")
    # Handle Docker, network, port binding issues
    
except ExecutionError as e:
    print(f"Execution failed: {e}")
    # Handle agent execution, tool failures, timeouts
    
    
except ClaudeAgentError as e:
    print(f"Library error: {e}")
    # Catch all library errors

Troubleshooting

Common Issues

ConfigurationError: "OAuth token required"

# Set environment variable
export CLAUDE_CODE_OAUTH_TOKEN='your-token-here'

# Or pass directly to Agent
agent = Agent(oauth_token='your-token-here')

ConnectionError: "Cannot connect to Docker"

• Ensure Docker Desktop is running
• Check Docker daemon is accessible
• Linux: sudo systemctl start docker

ConnectionError: "Port binding failed"

# Let tools auto-select available ports
tool = MyTool()  # Auto-selects port

# Or specify different port
tool = MyTool(port=9000)

ConnectionError: "Tool server failed to start"

# Tool server starts automatically in constructor
tool = MyTool()  # Server starts immediately
url = tool.connection_url  # Always accessible after construction

ExecutionError: "Operation timed out"

# Increase timeout for parallel operations
@tool(parallel=True, timeout_s=300)  # 5 minute timeout
def heavy_computation(self, data: str) -> dict:
    # Parallel operations must be sync functions
    return {"result": "processed"}

Debug Mode

from claude_agent_toolkit import set_logging, LogLevel

# Enable detailed logging
set_logging(LogLevel.DEBUG, show_time=True, show_level=True)

# Run with verbose output
result = await agent.run("your prompt", verbose=True)

Contributing

1. Create custom tools for different Claude Code agent use cases
2. Add new agent development patterns and templates
3. Improve Docker image efficiency and security
4. Enhance state management and conflict resolution
5. Add support for additional MCP server types

License

This project is licensed under the MIT License - see the LICENSE file for details.

Related Projects

• Claude Code - Official Claude Code interface (required for this framework)
• MCP (Model Context Protocol) - Protocol for AI-tool integration
• FastMCP - Fast MCP server implementation