aw_agents 0.1.2

AI agents for agentic workflows deployable to multiple chatbot platforms

aw-agents

AI Agents for Agentic Workflows - Easily deploy AI agents to multiple chatbot platforms (Claude, ChatGPT, and more).

Overview

aw-agents provides a unified framework for building AI agents and deploying them to different chatbot platforms with minimal boilerplate. Write your agent once, deploy everywhere.

Key Features

• 🎯 Write Once, Deploy Everywhere: Single agent implementation works with Claude, ChatGPT, and future platforms
• 🔌 Pluggable Adapters: MCP for Claude, OpenAPI for ChatGPT
• 📦 Batteries Included: Built-in download agent with smart features
• 🛠️ Easy to Extend: Simple base class for creating new agents
• 🎨 Clean Architecture: Separation of agent logic and platform adapters
• 🚀 Run as Script: Each agent can run directly as a server with built-in setup instructions

Documentation

• 📚 Quick Usage Guide - Simplest way to run agents
• 📖 This README - Comprehensive documentation
• 💻 Examples - Code examples

Installation

# From source (recommended for development)
git clone https://github.com/thorwhalen/aw_agents.git
cd aw_agents
pip install -e .

# With Claude Desktop support (MCP)
pip install -e .[mcp]

# With ChatGPT Custom GPT support (FastAPI)
pip install -e .[api]

# With all features
pip install -e .[all]

# For development
pip install -e .[dev]

Verify Installation

# Test the installation
python -c "from aw_agents.agents.download import DownloadAgent; agent = DownloadAgent(); print('✓ Installation successful!')"

Quick Start

Easiest Way: Run Agent as Server

Each agent can be run directly as a server with built-in instructions:

# For Claude Desktop (MCP server)
python -m aw_agents.agents.download.agent --mcp

# For ChatGPT (FastAPI server)
python -m aw_agents.agents.download.agent --api

# With ngrok for remote access
python -m aw_agents.agents.download.agent --api --server-url https://your-ngrok-url.ngrok.io

The agent will print:

• ✅ Configuration snippets (copy-paste ready)
• 📍 Exact file locations for configs
• 📋 Step-by-step setup instructions
• 🚀 Everything you need to get started
• 🌐 Proper OpenAPI schema with server URLs (fixes ChatGPT import issues)

Using the Download Agent

The package includes a smart download agent out of the box:

from aw_agents.agents.download import DownloadAgent

agent = DownloadAgent()

# Get available tools
tools = agent.get_tools()
print([t['name'] for t in tools])
# ['download_content', 'download_multiple', 'list_downloads']

# Execute a tool
result = agent.execute_tool('download_content', {
    'url': 'https://arxiv.org/pdf/2103.00020.pdf',
    'context': 'Important ML Paper'
})

print(result['data']['path'])  # ~/Downloads/Important_ML_Paper.pdf

Deploy to Claude Desktop (MCP)

Quick Start - Run the agent directly:

# The agent will automatically configure Claude Desktop for you!
python -m aw_agents.agents.download.agent --mcp

# Or run directly:
python aw_agents/agents/download/agent.py --mcp

The agent will:

• Automatically add itself to your Claude Desktop config
• Start the MCP server
• Give you step-by-step instructions
• You just need to restart Claude Desktop!

Managing Claude Desktop Configuration:

The agent uses claude_desktop_config() from the aw package to automatically manage your Claude Desktop configuration:

from aw.util import claude_desktop_config

# View all configured MCP servers
mcp = claude_desktop_config()
list(mcp)  # ['download', 'other_server', ...]

# Add a new server manually
mcp['my_agent'] = {
    'command': 'python',
    'args': ['/path/to/my_agent.py']
}

# Remove a server
del mcp['old_server']

# View a server's config
print(mcp['download'])

Manual Setup (Alternative):

If you prefer to create your own script and manage configuration manually:

Step 1: Create MCP Server Script

from aw_agents.agents.download import DownloadAgent
from aw_agents.adapters import MCPAdapter

agent = DownloadAgent()
adapter = MCPAdapter(agent, "download-agent")

# This starts an MCP server for Claude Desktop
if __name__ == "__main__":
    adapter.run_sync()

Save this as mcp_download.py or use the deployment script:

python scripts/deploy_mcp.py DownloadAgent --output mcp_download.py

Step 2: Configure Claude Desktop

Using claude_desktop_config():

from aw.util import claude_desktop_config

mcp = claude_desktop_config()
mcp['download'] = {
    'command': 'python',
    'args': ['/absolute/path/to/mcp_download.py']
}

Or manually edit the config file:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json
• Linux: ~/.config/Claude/claude_desktop_config.json

Important: Use absolute paths, not ~/ or relative paths.

Step 3: Restart Claude Desktop

Completely quit and restart Claude Desktop.

Step 4: Test It

Ask Claude:

Download this paper for me: https://arxiv.org/pdf/2103.00020.pdf
Context: "Attention is All You Need"

Troubleshooting:

• Check logs: ~/Library/Logs/Claude/ (macOS)
• Verify Python path in config
• Ensure dependencies are installed in the correct Python environment

Deploy to ChatGPT Custom GPT (FastAPI)

Quick Start - Run the agent directly:

# The agent will print setup instructions for you
python -m aw_agents.agents.download.agent --api

# Or specify a custom port:
python aw_agents/agents/download/agent.py --api --port 8080

The agent will:

• Start the FastAPI server
• Print the server URLs (API, docs, OpenAPI schema)
• Show you exactly how to configure ChatGPT Custom GPT
• Provide ngrok instructions for remote access

Manual Setup (Alternative):

Step 1: Create and Start API Server

from aw_agents.agents.download import DownloadAgent
from aw_agents.adapters import OpenAPIAdapter

agent = DownloadAgent()
adapter = OpenAPIAdapter(agent, title="Download Agent API")

# This starts a FastAPI server
if __name__ == "__main__":
    adapter.run(port=8000)

Save this as api_download.py or use the deployment script:

python scripts/deploy_api.py DownloadAgent --output api_download.py
python api_download.py  # Start the server

The server will be available at http://localhost:8000

Step 2: Create Custom GPT

1. Go to https://chat.openai.com/gpts/editor
2. Click "Create a GPT"
3. Name it (e.g., "Download Assistant")
4. Add instructions:

   You help users download files from URLs. Use the download_content 
   action when users provide URLs. Always ask for context to generate 
   good filenames. You can download multiple files at once and list 
   previously downloaded files.
   

Step 3: Add Actions

1. In the GPT editor, go to "Configure" → "Actions"
2. Click "Create new action"
3. Import schema from: http://localhost:8000/openapi.json
4. Click "Add Action"

Step 4: Test Your GPT

Ask your Custom GPT:

Download this dataset: https://github.com/user/repo/blob/main/data.csv
Context: "Sales Data Q4"

Important Notes:

• Local Testing: ChatGPT servers can't access localhost from OpenAI's infrastructure
• For Remote Access:
  • Use ngrok: ngrok http 8000 (provides a public URL)
  • Deploy to cloud service (AWS, GCP, Azure, Heroku)
  • Use local network IP if on the same network

Production Deployment:

1. Deploy to cloud service with HTTPS
2. Add authentication (API keys, OAuth)
3. Update Custom GPT actions with public URL

Architecture

aw_agents/
├── base.py              # AgentBase, ToolExecutionResult
├── adapters/
│   ├── mcp.py          # MCPAdapter for Claude
│   └── openapi.py      # OpenAPIAdapter for ChatGPT
└── download/           # Download agent implementation
    ├── agent.py        # DownloadAgent (AgentBase subclass)
    └── download_core.py # Core download logic

Creating Your Own Agent

from aw_agents.base import AgentBase, ToolExecutionResult, create_json_schema

class MyAgent(AgentBase):
    """My custom agent."""
    
    def get_tools(self):
        return [
            {
                'name': 'my_tool',
                'description': 'Does something useful',
                'parameters': create_json_schema(
                    properties={
                        'input': {
                            'type': 'string',
                            'description': 'Input text'
                        }
                    },
                    required=['input']
                )
            }
        ]
    
    def execute_tool(self, name, arguments):
        if name == 'my_tool':
            # Your logic here
            result = self._process(arguments['input'])
            
            return ToolExecutionResult.success_result(
                data={'output': result},
                message="Processed successfully"
            ).to_dict()
        
        return ToolExecutionResult.error_result(
            message=f"Unknown tool: {name}"
        ).to_dict()
    
    def _process(self, input_text):
        # Your implementation
        return input_text.upper()

Then deploy it:

# For Claude
from aw_agents.adapters import MCPAdapter
adapter = MCPAdapter(MyAgent(), "my-agent")
adapter.run_sync()

# For ChatGPT
from aw_agents.adapters import OpenAPIAdapter
adapter = OpenAPIAdapter(MyAgent())
adapter.run(port=8001)

Included Agents

Download Agent

Smart content downloader with:

• Landing page detection: Automatically finds actual download links
• Context-aware file naming: Uses conversation context for meaningful filenames
• Special handling for:
  • GitHub (converts blob URLs to raw)
  • HuggingFace (handles dataset downloads)
  • Kaggle datasets
  • Academic papers (arXiv, etc.)
• Support for: PDFs, datasets, CSV, JSON, and more

Tools:

• download_content - Download a single URL with smart handling
• download_multiple - Download multiple URLs at once
• list_downloads - List downloaded files

Examples:

from aw_agents.agents.download import DownloadAgent

agent = DownloadAgent()

# Download with context-aware naming
result = agent.execute_tool('download_content', {
    'url': 'https://github.com/user/repo/blob/main/data.csv',
    'context': 'Dataset for analysis'
})
# Automatically converts to raw.githubusercontent.com
# Saves as: Dataset_for_analysis.csv

# Download multiple files
result = agent.execute_tool('download_multiple', {
    'urls': [
        'https://arxiv.org/pdf/2103.00020.pdf',
        'https://arxiv.org/pdf/1706.03762.pdf'
    ],
    'contexts': [
        'Attention Paper',
        'Transformer Original'
    ]
})

# List downloaded PDFs
result = agent.execute_tool('list_downloads', {
    'pattern': '*.pdf'
})

Smart Features:

1. Intelligent Extension Detection:

   # Priority-based routing handles edge cases correctly
   agent.execute_tool('download_content', {
       'url': 'https://arxiv.org/pdf/2103.00020.pdf',
       'context': 'ML Paper'
   })
   # Result: ML_Paper.pdf ✓ (not .html despite Content-Type header)
   
   # Explicit extension override
   agent.execute_tool('download_content', {
       'url': 'https://example.com/ambiguous_file',
       'context': 'Data',
       'explicit_extension': 'csv'
   })
   # Result: Data.csv
   

2. Landing Page Detection:

   # Input: Landing page URL
   agent.execute_tool('download_content', {
       'url': 'https://papers.nips.cc/paper/2017/hash/xxx-Abstract.html'
   })
   # Agent detects HTML, finds PDF link, downloads PDF, warns about redirect
   

3. GitHub URL Handling:

   # Input: GitHub blob URL
   'https://github.com/user/repo/blob/main/data.csv'
   # Converts to: 
   'https://raw.githubusercontent.com/user/repo/main/data.csv'
   

4. Context-Aware Naming:

   # Without context:
   # Filename: 2103.00020.pdf
   
   # With context:
   agent.execute_tool('download_content', {
       'url': 'https://arxiv.org/pdf/2103.00020.pdf',
       'context': 'Attention is All You Need - Transformer Paper'
   })
   # Filename: Attention_is_All_You_Need_Transformer_Paper.pdf
   

Customizing Extension Detection

The download agent uses configurable routing from the aw.routing module:

from aw.routing import ExtensionRouter
from aw_agents.agents.download import DownloadEngine

# Inspect default mappings
router = ExtensionRouter()
print(router.content_type_map)
print(router.priority_extensions)

# Customize mappings
router.content_type_map['text/x-log'] = '.log'

# Add custom detection strategy
def detect_special(ctx):
    if 'special' in ctx.url:
        return '.special'
    return None

custom_router = router.with_prepended_strategy(detect_special)

# Use in agent
engine = DownloadEngine(extension_router=custom_router)

See EXTENSION_ROUTING_SUMMARY.md for details.

Integration Guides

Claude Desktop Integration

Easiest Way - Automatic Configuration:

# Run the agent - it will automatically configure Claude Desktop!
python -m aw_agents.agents.download.agent --mcp

The agent automatically adds itself to your Claude Desktop config. Just restart Claude Desktop and you're done!

Alternative - Manual Configuration:

1. Create MCP server script:

   python scripts/deploy_mcp.py DownloadAgent --output ~/mcp_download.py
   

2. Add to Claude Desktop config using claude_desktop_config():

   from aw.util import claude_desktop_config
   
   mcp = claude_desktop_config()
   mcp['download'] = {
       'command': 'python',
       'args': ['/absolute/path/to/mcp_download.py']
   }
   

3. Restart Claude Desktop

Managing Servers:

from aw.util import claude_desktop_config

mcp = claude_desktop_config()

# List all servers
list(mcp)  # ['download', 'other_server', ...]

# View a server's config
print(mcp['download'])

# Remove a server
del mcp['old_server']

Config file locations:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json
• Linux: ~/.config/Claude/claude_desktop_config.json

Usage Examples in Claude:

Download this paper: https://arxiv.org/pdf/2103.00020.pdf
Context: "Attention is All You Need paper"

I need these papers:
1. https://arxiv.org/pdf/2103.00020.pdf - "Attention Paper"
2. https://arxiv.org/pdf/1706.03762.pdf - "Transformer Original"

Show me what PDFs I've downloaded

ChatGPT Custom GPT Integration

1. Create API server script:

   python scripts/deploy_api.py DownloadAgent --output ~/api_download.py
   python ~/api_download.py
   

2. Create Custom GPT at https://chat.openai.com/gpts/editor

3. Add actions using schema from: http://localhost:8000/openapi.json

4. For remote access, deploy to cloud or use ngrok:

   ngrok http 8000
   

Usage Examples in ChatGPT:

Download this dataset: https://github.com/user/repo/blob/main/data.csv
Context: "Sales Data Q4 2024"

Can you download these research papers for me?
[List of URLs with contexts]

Troubleshooting

Installation Issues

Problem: Module not found errors

Solution:

# Update pip and reinstall
pip install --upgrade pip
pip uninstall aw-agents
pip install -e .[all]

Problem: Missing dependencies

Solution:

# Install specific feature sets
pip install -e .[mcp]    # For Claude
pip install -e .[api]    # For ChatGPT
pip install -e .[all]    # Everything

Claude Desktop Issues

Problem: Claude doesn't see the download tools

Solution:

1. Verify config file location (see paths above)
2. Use absolute path in config (not ~/ or ./)
3. Check logs: ~/Library/Logs/Claude/ (macOS)
4. Restart Claude Desktop completely (Quit, not just close window)
5. Verify Python environment: which python

Problem: "Module not found" when Claude tries to use agent

Solution:

# Install in the Python environment Claude uses
which python  # Check which Python Claude is using
# Use that Python to install dependencies
/path/to/python -m pip install mcp requests beautifulsoup4 graze

ChatGPT Custom GPT Issues

Problem: Actions can't connect to localhost

Solution:

• ChatGPT servers can't access localhost from OpenAI's infrastructure
• Options:
  1. For testing: Use ngrok to expose local server

     ngrok http 8000
     # Use the ngrok URL in your Custom GPT actions
     

  2. For production: Deploy to cloud service (AWS, GCP, Azure, Heroku)
  3. Same network: Use your local IP address

Problem: Schema import fails

Solution:

1. Ensure server is running: curl http://localhost:8000/openapi.json
2. Check for CORS issues (FastAPI handles this by default)
3. Try manual schema paste instead of URL import

General Issues

Problem: Downloads failing

Solution:

1. Check internet connection
2. Verify URL is accessible: curl -I [URL]
3. Some sites may block automated downloads
4. Check download directory permissions

Problem: Import errors in Python

Solution:

# Verify installation
python -c "from aw_agents.agents.download import DownloadAgent; print('✓ Success')"

# Reinstall if needed
pip install -e .[all]

Development

Running Tests

pytest

Code Quality

# Format code
black aw_agents tests

# Lint
ruff aw_agents tests

# Type check
mypy aw_agents

Adding a New Agent

1. Create a new directory in aw_agents/:

   aw_agents/myagent/
   ├── __init__.py
   ├── agent.py
   └── core.py (if needed)
   

2. Implement AgentBase in agent.py

3. Add to aw_agents/__init__.py:

   from aw_agents.myagent import MyAgent
   __all__.append('MyAgent')
   

4. Create tests in tests/test_myagent.py

5. Update documentation

Roadmap

• Additional built-in agents (research, data prep, etc.)
• Support for more platforms (Slack, Discord, etc.)
• Agent composition and chaining
• Persistent state management
• Authentication and authorization
• Web UI for agent management
• CLI tools for deployment

Contributing

Contributions welcome! Please:

1. Fork the repo
2. Create a feature branch
3. Add tests for new features
4. Ensure all tests pass
5. Submit a pull request

License

MIT License - see LICENSE file for details

Credits

Built with:

• dol - Storage abstractions
• graze - Download and caching
• MCP - Model Context Protocol
• FastAPI - Modern web framework

Part of the thorwhalen ecosystem.