pydantic-ai-claude-code 0.8.0

Use local Claude Code CLI as a Pydantic AI model provider with full support for structured responses, tools, and streaming

t# Pydantic AI Claude Code

Use your local Claude Code CLI as a Pydantic AI model provider.

This package provides a Pydantic AI-compatible model implementation that wraps the local Claude CLI, enabling you to use Claude locally with all Pydantic AI features including structured responses, tool calling, streaming, and multi-turn conversations.

Features

• Full Pydantic AI Compatibility: Drop-in replacement for any Pydantic AI model
• Structured Responses: Get validated, typed responses using Pydantic models
• Custom Tool Calling: Use your own Python functions as tools
• True Streaming: Real-time response streaming via Claude CLI's stream-json mode
• Local Execution: All processing happens locally on your machine
• Session Persistence: Maintain conversation context across multiple requests
• Additional Files: Provide local files for Claude to read and analyze
• Automatic Response Saving: Raw prompts and responses saved for debugging
• Configurable: Fine-tune permissions, working directories, and tool access

Installation

# Using uv (recommended)
uv add pydantic-ai-claude-code

# Using pip
pip install pydantic-ai-claude-code

Prerequisites:

• Claude Code CLI must be installed and authenticated
• Sandbox Runtime (required): npm install -g @anthropic-ai/sandbox-runtime

Sandbox Mode (Enabled by Default)

New in v0.8.0: All Claude Code execution now runs in a secure sandbox by default using Anthropic's sandbox-runtime.

What This Means

• OS-Level Isolation: Uses bubblewrap (Linux) or sandbox-exec (macOS) for kernel-level security
• Filesystem Access: Full read/write access to /tmp only - everything else is read-only
• Network Access: Restricted to api.anthropic.com for Claude API calls
• Autonomous Execution: IS_SANDBOX=1 environment variable enables Claude to run without permission prompts
• Zero Configuration: Works automatically after installing sandbox-runtime

Installation

# Install sandbox-runtime globally via npm
npm install -g @anthropic-ai/sandbox-runtime

# Verify installation
srt --version

Disabling Sandbox (Not Recommended)

If you need to disable sandbox mode for debugging:

from pydantic_ai_claude_code import ClaudeCodeProvider
from pydantic_ai import Agent

provider = ClaudeCodeProvider({"use_sandbox_runtime": False})
agent = Agent('claude-code:sonnet')

⚠️ Warning: Disabling sandbox removes security isolation. Only do this in trusted environments.

Quick Start

Basic Usage (String Format - Simplest!)

import pydantic_ai_claude_code  # Register the provider

from pydantic_ai import Agent

# Just use the string format - easiest way!
agent = Agent('claude-code:sonnet')

# Run a simple query
result = agent.run_sync("What is 2+2?")
print(result.output)  # Output: 4

Structured Responses

import pydantic_ai_claude_code

from pydantic import BaseModel
from pydantic_ai import Agent

class Analysis(BaseModel):
    complexity: int  # 1-10
    maintainability: str
    suggestions: list[str]

# String format works with all Pydantic AI features
agent = Agent('claude-code:sonnet', output_type=Analysis)

result = agent.run_sync("Analyze this code: def foo(): pass")
print(f"Complexity: {result.output.complexity}")

Custom Tools

import pydantic_ai_claude_code

from pydantic_ai import Agent

def get_weather(city: str) -> str:
    """Get weather for a city."""
    return f"Weather in {city}: Sunny, 22°C"

agent = Agent(
    'claude-code:sonnet',
    tools=[get_weather],
)

result = agent.run_sync("What's the weather in Paris?")
print(result.output)

Streaming

import pydantic_ai_claude_code

from pydantic_ai import Agent

agent = Agent('claude-code:sonnet')

async with agent.run_stream('Write a haiku about code') as result:
    async for text in result.stream_text():
        print(text, end='', flush=True)

Advanced Configuration

When you need fine-grained control, use the explicit model and provider:

from pydantic_ai import Agent
from pydantic_ai_claude_code import ClaudeCodeModel, ClaudeCodeProvider

# Custom provider with specific settings
provider = ClaudeCodeProvider(
    working_directory="/path/to/project",
    allowed_tools=["Read", "Edit", "Grep"],  # Restrict tool access
    permission_mode="acceptEdits",
    use_temp_workspace=False,  # Use specific directory instead of /tmp
)

model = ClaudeCodeModel("sonnet", provider=provider)
agent = Agent(model)

result = agent.run_sync("Analyze the codebase structure")

Temporary Workspace

from pydantic_ai_claude_code import ClaudeCodeModel, ClaudeCodeProvider

# Create isolated workspace that auto-cleans
with ClaudeCodeProvider(use_temp_workspace=True) as provider:
    model = ClaudeCodeModel("sonnet", provider=provider)
    agent = Agent(model)
    result = agent.run_sync("Create a test file")
# Workspace automatically cleaned up

Providing Additional Files

Provide local files for Claude to read and analyze:

from pathlib import Path
from pydantic_ai import Agent

agent = Agent('claude-code:sonnet')

result = agent.run_sync(
    "Read utils.py and config.json. Summarize what they configure.",
    model_settings={
        "additional_files": {
            "utils.py": Path("src/utils.py"),           # Copy single file
            "config.json": Path("config/prod.json"),    # From different location
            "docs/spec.md": Path("specs/feature.md"),   # Into subdirectory
        }
    }
)

Files are copied into the working directory before execution, and Claude can reference them directly:

• "Read utils.py"
• "Read config.json"
• "Read docs/spec.md"

Each execution gets its own numbered subdirectory with isolated file copies.

Binary Content (Standard Interface)

Use Pydantic AI's standard BinaryContent interface to attach images, PDFs, and other binary files. This works identically across all model providers:

from pathlib import Path
from pydantic_ai import Agent, BinaryContent

agent = Agent('claude-code:sonnet')

# Send an image from file
image_data = Path('photo.jpg').read_bytes()
result = agent.run_sync(
    [
        'What is in this image?',
        BinaryContent(data=image_data, media_type='image/jpeg'),
    ]
)

# Send a PDF document
pdf_data = Path('document.pdf').read_bytes()
result = agent.run_sync(
    [
        'Summarize this document:',
        BinaryContent(data=pdf_data, media_type='application/pdf'),
    ]
)

# Compare multiple images
result = agent.run_sync(
    [
        'Compare these images:',
        BinaryContent(data=image1_data, media_type='image/png'),
        'and',
        BinaryContent(data=image2_data, media_type='image/jpeg'),
    ]
)

How it works:

• Binary content is automatically written to files in the working directory
• Files are referenced in the prompt (e.g., [Image: filename.png])
• Claude Code CLI can then read the files directly
• No Claude Code-specific code needed - this is standard Pydantic AI!

Supported formats:

• Images: PNG, JPEG, GIF, WebP, etc.
• Documents: PDF, TXT, JSON, etc.
• Audio: MP3, WAV, etc.
• Video: MP4, WebM, etc.

Benefits:

• Portable code - works with OpenAI, Anthropic, Google, and Claude Code
• Standard interface - same BinaryContent class for all providers
• No provider lock-in - switch between cloud and local with one line change

Error Handling

OAuth Token Expiration

For long-running processes (>7 hours), OAuth tokens may expire. Handle gracefully with ClaudeOAuthError:

from pydantic_ai import Agent
from pydantic_ai_claude_code import ClaudeOAuthError

agent = Agent('claude-code:sonnet')

try:
    result = agent.run_sync("Process large dataset")
except ClaudeOAuthError as e:
    print(f"Authentication expired: {e}")
    print(f"Please run: {e.reauth_instruction}")  # "Please run /login"
    # Prompt user to re-authenticate, then retry
except RuntimeError as e:
    # Handle other CLI errors
    print(f"CLI error: {e}")

For batch processing with automatic retry:

from pydantic_ai_claude_code import ClaudeOAuthError
import time

def process_batch_with_retry(items, max_auth_retries=3):
    """Process items with OAuth re-authentication support."""
    results = []

    for item in items:
        auth_retries = 0
        while auth_retries < max_auth_retries:
            try:
                result = agent.run_sync(f"Process: {item}")
                results.append(result.output)
                break  # Success

            except ClaudeOAuthError as e:
                auth_retries += 1
                print(f"\n{'='*60}")
                print(f"OAuth token expired: {e.reauth_instruction}")
                print(f"{'='*60}\n")

                if auth_retries >= max_auth_retries:
                    raise  # Give up after max retries

                input("Press Enter after running /login to continue...")
                time.sleep(2)  # Brief pause before retry

    return results

Logging

The package uses Python's standard logging module. To enable debug logging in your application:

import logging

# Enable debug logging for pydantic-ai-claude-code
logging.basicConfig(level=logging.DEBUG)
logging.getLogger('pydantic_ai_claude_code').setLevel(logging.DEBUG)

# Or just for specific components
logging.getLogger('pydantic_ai_claude_code.model').setLevel(logging.DEBUG)
logging.getLogger('pydantic_ai_claude_code.utils').setLevel(logging.INFO)

This will log:

• Model initialization and configuration
• CLI command execution and responses
• Message formatting and conversion
• Tool call parsing and execution
• Structured output handling
• Streaming events and completion

Available Models

• claude-code:sonnet - Claude 3.5 Sonnet (default, recommended)
• claude-code:opus - Claude 3 Opus (most capable)
• claude-code:haiku - Claude 3.5 Haiku (fastest)

Or use full model names like claude-code:claude-sonnet-4-5-20250929

Integration with Existing Projects

Replace your current LLM calls with Claude Code:

Before:

agent = Agent('openai:gpt-4o')
# or
agent = Agent('anthropic:claude-3-5-sonnet-latest')

After:

import pydantic_ai_claude_code  # Add this import

agent = Agent('claude-code:sonnet')  # Change this line

Everything else stays the same! All your tools, structured outputs, dependencies, and streaming code works identically.

Key Differences from Cloud Providers

Aspect	Cloud Providers	Claude Code Local
Execution	Remote API calls	Local on your machine
Cost	Per-token pricing	Uses Claude desktop subscription
Data Privacy	Data sent to cloud	Data stays local
Speed	Network latency	Local execution
API Key	Required	Not needed (uses local auth)

Examples

See the examples/ directory for more demonstrations:

• basic_example.py - Simple queries and usage tracking
• structured_example.py - Structured output with Pydantic models
• async_example.py - Async/await usage patterns
• advanced_example.py - Custom provider configurations
• tools_and_streaming.py - Custom tools and streaming responses
• sandbox_example.py - Sandbox-runtime integration for production security
• binary_content_example.py - Standard interface for images, PDFs, and binary files
• additional_files_example.py - Providing local files for analysis (Claude Code-specific)

License

MIT License