cdd-agent 0.0.6

AI coding assistant with structured CDD workflows - LLM agnostic terminal agent

CDD Agent

LLM-agnostic AI coding assistant with structured workflows and background execution

---

What is CDD Agent?

CDD Agent is a terminal-based AI coding assistant that lets you use any LLM provider - Anthropic (Claude), OpenAI (GPT), or custom endpoints - without vendor lock-in. It implements structured Context-Driven Development workflows while giving you complete control over your tools and configuration.

Unlike vendor-specific tools (Claude Code, Cursor), CDD Agent gives you:

• 🔄 Freedom to choose and switch LLM providers
• 🎯 Structured workflows that maintain context
• 🏠 Local configuration and complete control
• 💰 No subscription lock-in
• ⚡ Background bash execution for long-running tasks
• 🔐 OAuth support for convenient authentication

---

✨ What's New in v0.0.4

🚀 Background Bash Execution

Run long-running commands without blocking your workflow:

• Start processes in background and continue chatting
• Monitor output in real-time
• Interrupt or check status anytime
• Perfect for tests, builds, servers, and watch processes

🔐 OAuth Authentication

Convenient OAuth 2.0 authentication flow:

• Browser-based authorization
• Create API keys through OAuth
• Automatic token management
• Note: Zero-cost access restricted to official Claude Code

⚡ Performance Improvements

• Lazy imports for faster startup
• Optimized tool execution
• Efficient context loading

🛠️ Enhanced Tools

• Background process management
• Improved error handling
• Better streaming output

---

Current Status: v0.0.4 (Beta)

This is a beta release with core features functional and production-ready for daily use.

✅ Core Features

Multi-Provider Architecture:

• Anthropic (Claude Haiku, Sonnet, Opus)
• OpenAI (GPT-4o, GPT-4, GPT-3.5)
• Custom endpoints (local models, proxies, alternative providers)

Beautiful TUI:

• Textual-based split-pane interface
• Token-by-token streaming responses
• Syntax-highlighted code blocks
• Interactive tool approval

Advanced Tools:

• File operations: read_file, write_file, edit_file
• Code search: glob_files (pattern matching), grep_files (regex)
• Shell execution: run_bash with output capture
• Background execution: Run long commands without blocking

Safety & Context:

• Tool approval system (paranoid/balanced/trusting modes)
• Hierarchical context loading (global → project)
• Security warnings for dangerous operations
• Context from CDD.md or CLAUDE.md files

Authentication:

• Interactive setup wizard
• OAuth 2.0 support for API key creation
• Environment variable overrides
• Multi-provider configuration

🔜 Coming Soon (v0.1.x)

• Further performance optimization (<200ms startup)
• Enhanced git integration with commit preview
• Specialized CDD workflow agents (Socrates, Planner, Executor)
• Plugin system for custom tools

See ROADMAP.md for the full development plan.

---

📦 Installation

pip install cdd-agent

Requirements:

• Python 3.10 or higher
• API keys for your chosen LLM provider(s)

From source:

git clone https://github.com/guilhermegouw/context-driven-development-agent.git
cd context-driven-development-agent
poetry install
poetry run cdd-agent --help

---

🚀 Quick Start

1. Configure your LLM provider

Interactive Setup (Recommended):

# Setup wizard with provider selection
cdd-agent auth setup

# OAuth authentication (creates API key via browser)
cdd-agent auth oauth

# Choose provider: anthropic, openai, or custom

Manual Configuration:

# Set environment variables
export ANTHROPIC_API_KEY="your-api-key"
export OPENAI_API_KEY="your-api-key"

2. Verify your configuration

# Check configured providers
cdd-agent auth status

# Test API credentials
cdd-agent auth test anthropic

3. Start coding

# Interactive TUI mode (default)
cdd-agent chat

# With custom approval mode
cdd-agent chat --approval paranoid    # Ask for every tool execution
cdd-agent chat --approval balanced    # Auto-approve safe tools (default)
cdd-agent chat --approval trusting    # Remember approvals per session

# Choose provider and model
cdd-agent chat --provider anthropic --model mid
cdd-agent chat --provider openai --model big

# Simple streaming mode (no TUI)
cdd-agent chat --simple

# Single-shot execution
cdd-agent chat "Explain this codebase" --no-stream

4. Set up context files (optional but recommended)

# Global preferences (applies to all projects)
mkdir -p ~/.cdd
cat > ~/.cdd/CDD.md << 'EOF'
# Global Context

## Coding Style
- Always use type hints in Python
- Prefer functional programming style
- Write comprehensive docstrings

## Testing
- Use pytest for all tests
- Aim for >80% code coverage
EOF

# Project context (specific to this project)
cat > CDD.md << 'EOF'
# Project Context

## Architecture
This is a Flask web application with PostgreSQL backend.

## Tech Stack
- Python 3.11
- Flask 3.0
- SQLAlchemy ORM
- PostgreSQL 15

## Development
- Use virtual environment
- Run migrations before testing
EOF

# The agent automatically loads this context
cdd-agent chat  # Context loaded from both files

---

🎯 Features

Multi-Provider Support

Switch between providers anytime without changing your workflow:

# Use Claude for complex reasoning
cdd-agent chat --provider anthropic --model big "Refactor this codebase"

# Use GPT for quick queries
cdd-agent chat --provider openai --model small "Fix syntax errors"

# Use custom endpoint for local models
cdd-agent auth setup  # Choose "custom" and enter your endpoint

Supported Providers:

• Anthropic: Claude 4.5 Haiku, Sonnet, Opus
• OpenAI: GPT-4o, GPT-4o-mini, GPT-3.5
• Custom: Any OpenAI-compatible API (Ollama, LM Studio, etc.)

Model Tier Abstraction

Configure models by purpose instead of specific versions:

{
  "models": {
    "small": "claude-3-5-haiku-20241022",      // Fast, cheap
    "mid": "claude-sonnet-4-5-20250929",       // Balanced
    "big": "claude-opus-4-20250514"            // Maximum capability
  }
}

Benefits:

• Update model versions in one place
• Switch providers without code changes
• Cost optimization based on task complexity

Background Bash Execution

Run long commands without blocking your workflow:

# In chat, the AI can run commands in background:
> "Run the test suite in background"

# AI executes:
# run_bash_background("pytest tests/")

# You can continue chatting while tests run
> "While tests run, explain the auth module"

# Check background status anytime:
> "Check if tests finished"

# AI executes:
# get_background_status(process_id)

Available background tools:

• run_bash_background - Start command in background
• get_background_status - Check if process is running
• get_background_output - Get process output
• interrupt_background_process - Stop a process
• list_background_processes - List all running processes

Perfect for:

• Test suites
• Build processes
• Development servers
• Watch processes
• Long-running scripts

OAuth Authentication

Convenient browser-based authentication:

# Start OAuth flow
cdd-agent auth oauth

# Choose mode:
# - "api-key": Create permanent API key (recommended)
# - "max": OAuth tokens (restricted to Claude Code)

# Browser opens for authorization
# Paste code back to CLI
# Tokens/keys saved automatically

Benefits:

• No need to manually create API keys in console
• Secure browser-based authorization
• Automatic token management
• OAuth 2.0 with PKCE security

Important: Zero-cost API access (Claude Pro/Max) only works with official Claude Code application. Third-party apps like cdd-agent use regular API pricing.

Hierarchical Context Loading

Load context from multiple sources with intelligent merging:

Context Sources (in order of priority):

1. Project context: CDD.md or CLAUDE.md at project root
2. Global context: ~/.cdd/CDD.md or ~/.claude/CLAUDE.md
3. System prompt: Built-in pair programming guidelines

Features:

• Automatic project root detection (.git, pyproject.toml, package.json)
• LLM-aware context merging (project overrides global)
• Caching for performance
• Disable with --no-context flag

Use cases:

• Global coding standards across all projects
• Project-specific architecture documentation
• Team conventions and preferences
• Technology stack information

Security & Approval System

Three approval modes to control tool execution:

Paranoid Mode:

cdd-agent chat --approval paranoid

• Ask for approval on every tool execution
• Perfect for learning or critical operations
• See exactly what the AI wants to do

Balanced Mode (Default):

cdd-agent chat --approval balanced

• Auto-approve safe read-only tools (read_file, glob_files)
• Ask for approval on write operations
• Best balance of safety and convenience

Trusting Mode:

cdd-agent chat --approval trusting

• Remember approvals within session
• Minimal interruptions
• For experienced users

Risk Classification:

• 🟢 SAFE: Read-only operations (read_file, grep_files)
• 🟡 MEDIUM: File modifications (write_file, edit_file)
• 🔴 HIGH: Dangerous operations (run_bash, system commands)

---

⚙️ Configuration

Configuration File

CDD Agent stores settings in ~/.cdd-agent/settings.json:

{
  "version": "1.0",
  "default_provider": "anthropic",
  "approval_mode": "balanced",
  "providers": {
    "anthropic": {
      "auth_token": "sk-ant-...",
      "base_url": "https://api.anthropic.com",
      "models": {
        "small": "claude-3-5-haiku-20241022",
        "mid": "claude-sonnet-4-5-20250929",
        "big": "claude-opus-4-20250514"
      },
      "default_model": "mid"
    },
    "openai": {
      "api_key": "sk-...",
      "base_url": "https://api.openai.com/v1",
      "models": {
        "small": "gpt-4o-mini",
        "mid": "gpt-4o",
        "big": "gpt-4o"
      }
    },
    "custom": {
      "auth_token": "optional",
      "base_url": "http://localhost:11434/v1",
      "provider_type": "anthropic",
      "models": {
        "small": "llama3",
        "mid": "codellama",
        "big": "mixtral"
      }
    }
  },
  "ui": {
    "stream_responses": true,
    "syntax_highlighting": true
  },
  "conversation": {
    "auto_save": true,
    "history_limit": 100
  }
}

Environment Variables

Override settings with environment variables (highest priority):

# Authentication
export CDD_AUTH_TOKEN="your-key"
export ANTHROPIC_AUTH_TOKEN="your-key"
export OPENAI_API_KEY="your-key"

# Provider settings
export CDD_BASE_URL="https://custom-endpoint.com"
export ANTHROPIC_BASE_URL="https://api.anthropic.com"

# Approval mode
export CDD_APPROVAL_MODE="paranoid"

Priority order:

1. CLI flags (--approval, --provider)
2. Environment variables (CDD_APPROVAL_MODE)
3. Settings file (~/.cdd-agent/settings.json)
4. Defaults (balanced approval, anthropic provider)

---

🛠️ Available Tools

CDD Agent provides Claude Code-level tool capabilities:

File Operations

• read_file(path) - Read file contents
• write_file(path, content) - Create or overwrite file
• write_file_lines(path, line_number, content) - Insert at specific line
• edit_file(path, old_text, new_text) - Replace text in file

Code Search

• glob_files(pattern, path) - Find files by pattern (.py, src//test_*.py)
• grep_files(pattern, file_pattern, path) - Search code with regex

Shell Execution

• run_bash(command) - Execute shell command and capture output
• run_bash_background(command) - Run command in background
• get_background_status(process_id) - Check background process
• get_background_output(process_id, lines) - Get process output
• interrupt_background_process(process_id) - Stop background process
• list_background_processes() - List all background processes

Git Operations

• git_status() - Show git status
• git_diff(file) - Show file diff

All tools include:

• Detailed descriptions for the LLM
• Type-safe argument validation
• Error handling and recovery
• Risk level classification
• User approval integration

---

📊 Comparison

Feature	CDD Agent	Claude Code	Cursor	Copilot
Multi-Provider	✅ Any LLM	❌ Claude only	❌ OpenAI only	❌ GitHub only
Local Control	✅ Full	⚠️ Limited	⚠️ Limited	❌ Cloud
Background Execution	✅ Yes	✅ Yes	❌ No	❌ No
Tool Approval	✅ 3 modes	⚠️ Basic	❌ None	❌ None
Context Loading	✅ Hierarchical	✅ Yes	✅ Yes	⚠️ Limited
OAuth Support	✅ Yes	✅ Yes	✅ Yes	✅ Yes
Custom Tools	🔜 v0.1.x	❌ No	❌ No	❌ No
Cost	💰 API usage	💰 Subscription	💰 Subscription	💰 Subscription
Open Source	✅ MIT	❌ Proprietary	❌ Proprietary	❌ Proprietary

---

🤝 Why Choose CDD Agent?

The Problem

Existing AI coding assistants create vendor lock-in:

• Claude Code → Anthropic only (can't use GPT)
• Cursor → OpenAI only (can't use Claude)
• GitHub Copilot → Microsoft ecosystem only

What happens when:

• Your preferred model improves on a different platform?
• API pricing changes?
• You want to use local models?
• The service has an outage?

The Solution

CDD Agent provides freedom and flexibility:

✅ Provider Independence

• Switch between Claude, GPT, or local models instantly
• No workflow changes when switching providers
• Use the best model for each task

✅ No Lock-In

• Your configuration is local and portable
• Open source - audit and modify the code
• No proprietary formats or cloud dependencies

✅ Cost Control

• Choose cheaper models for simple tasks
• Use local models for free inference
• Pay only for what you use

✅ Future-Proof

• Works with any OpenAI-compatible API
• Easy to add new providers
• Continuous improvement without breaking changes

Who Should Use CDD Agent?

Perfect for:

• Developers who want flexibility in LLM choice
• Teams with multiple LLM providers
• Users of local/custom models
• Projects requiring audit trails and approval
• Anyone who values open source and control

Not ideal for:

• Users who prefer integrated IDE experience
• Teams standardized on a single vendor tool
• Those who want zero configuration

---

🗺️ Roadmap

✅ Completed

v0.0.1 - Foundation

• Authentication system
• Multi-provider configuration
• CLI framework

v0.0.2 - Basic Agent

• Agentic loop with tool execution
• TUI with split-pane interface
• Streaming responses

v0.0.3 - Advanced Features

• Comprehensive tool suite
• Approval system
• Hierarchical context loading

v0.0.4 - Background & OAuth (Current)

• Background bash execution
• OAuth 2.0 authentication
• Performance improvements

🚧 In Progress

v0.1.0 - Performance & Polish

• Startup time <200ms (currently ~700ms)
• Enhanced git integration
• Improved error messages
• Plugin system foundation

🔜 Planned

v0.2.0 - CDD Workflows

• Socrates agent (requirements gathering)
• Planner agent (architecture design)
• Executor agent (implementation)
• Workflow orchestration

v0.3.0 - Team Features

• Shared context repositories
• Team configuration templates
• Collaborative workflows

v1.0.0 - Production Ready

• Full CDD methodology implementation
• Comprehensive documentation
• Extensive testing
• Performance optimization

See ROADMAP.md for detailed milestones and feature descriptions.

---

🧪 Development

Tech Stack

• Python 3.10+ - Modern Python features
• Click - CLI framework
• Rich - Terminal formatting
• Textual - TUI framework
• Pydantic - Configuration validation
• Poetry - Dependency management

Project Structure

cdd-agent-cli/
├── src/cdd_agent/
│   ├── cli.py              # CLI entry point
│   ├── agent.py            # Main agent loop
│   ├── tools.py            # Tool registry
│   ├── auth.py             # Authentication
│   ├── config.py           # Configuration management
│   ├── oauth.py            # OAuth implementation
│   ├── tui.py              # Textual TUI
│   ├── ui.py               # Streaming UI
│   ├── approval.py         # Tool approval system
│   ├── context.py          # Context loading
│   ├── background_executor.py  # Background processes
│   └── utils/              # Utilities
├── tests/                  # Test suite
├── docs/                   # Documentation
└── pyproject.toml          # Project metadata

Contributing

We welcome contributions! Here's how to get started:

1. Check the roadmap: See ROADMAP.md for planned features
2. Open an issue: Discuss your idea before implementing
3. Fork and code: Follow existing code style
4. Add tests: Ensure your changes are tested
5. Submit PR: Include description and test results

Good first issues:

• Add new tool implementations
• Improve error messages
• Add provider-specific optimizations
• Enhance documentation

Running Tests

# Install dev dependencies
poetry install

# Run full test suite
poetry run pytest

# Run with coverage
poetry run pytest --cov=cdd_agent

# Run specific test
poetry run pytest tests/test_tools.py -v

Code Quality

# Format code
poetry run black src/ tests/

# Type checking
poetry run mypy src/

# Linting
poetry run ruff check src/

---

📚 Documentation

• ROADMAP.md - Development roadmap and milestones
• OAUTH_IMPLEMENTATION.md - OAuth implementation details
• OAUTH_LIMITATION.md - OAuth restrictions and limitations
• OAUTH_TESTING.md - OAuth testing guide

---

🐛 Troubleshooting

Common Issues

"No configuration found"

# Run setup wizard
cdd-agent auth setup

"API error: 401 Unauthorized"

# Check API key
cdd-agent auth test

# Verify environment variables
echo $ANTHROPIC_API_KEY

"OAuth: This credential is only authorized for Claude Code"

• Choose "api-key" mode instead of "max" when running cdd-agent auth oauth
• Zero-cost access only works with official Claude Code application

Slow startup time

• Already improved from ~2s to ~700ms
• Further optimization planned for v0.1.0

Debug Mode

# Enable debug logging
export CDD_LOG_LEVEL=DEBUG
cdd-agent chat

# Check logs
cdd-agent logs show

Getting Help

1. Check documentation files in the repo
2. Search existing issues
3. Open a new issue with:
   • CDD Agent version (cdd-agent --version)
   • Python version (python --version)
   • Error message and logs
   • Steps to reproduce

---

📄 License

MIT License - see LICENSE for details.

In short: Use it, modify it, distribute it. Just keep the license notice.

---

🙏 Acknowledgments

• Inspired by Claude Code and Cursor
• Built on the Context-Driven Development methodology
• Powered by amazing open-source libraries (Click, Rich, Textual, Pydantic)

---

💬 Community

• Issues: GitHub Issues
• Discussions: GitHub Discussions
• Author: @guilhermegouw

---

Built by developers, for developers. No vendor lock-in. Your code, your choice.

CDD Agent - Because your tools should work for you, not lock you in.