AgentWerkstatt 🤖

Agents for Minimalists.

A minimalistic agentic framework for building AI agents with tool calling capabilities. Why do we need another agentic framework? I felt that all other frameworks were too complex and had too many dependencies, I wanted to build a framework that was easy to understand and use, and that was also easy to extend. The main goal here isn't on production scenarios (yet), but on understanding and prototyping agentic systems. Therefore, the name "AgentWerkstatt" is a play on the term Agent and the German word "Werkstatt" (workshop).

Overview

AgentWerkstatt is a lightweight, extensible framework for creating AI agents. It is powered by Claude (Anthropic), but it is highly extensible and can be used with other LLMs. It features a modular architecture with pluggable LLM providers and tools, making it easy to build conversational agents with access to external capabilities like web search. More LLMs will be supported in the future.

Features

• 🧠 Modular LLM Support - Built with extensible LLM abstraction (currently supports Claude)
• 🔧 Tool System - Pluggable tool architecture with automatic tool discovery
• 💬 Conversation Management - Built-in conversation history and context management
• 🧮 Persistent Memory - Optional mem0 integration for long-term memory and context retention
• 🌐 Web Search - Integrated Tavily API for real-time web information retrieval
• 📊 Observability - Optional Langfuse integration for comprehensive tracing and analytics
• 🖥️ CLI Interface - Ready-to-use command-line interface
• 🐳 3rd Party Services - Docker Compose stack with PostgreSQL, Neo4j, and other services
• ⚡ Lightweight - Minimal dependencies and clean architecture

Quick Start

Prerequisites

• Python 3.10 or higher
• An Anthropic API key for Claude
• (Optional) A Tavily API key for web search
• (Optional) An OpenAI API key for mem0 memory system

Installation

1. Clone the repository:

   git clone https://github.com/hanneshapke/agentwerkstatt.git
   cd agentwerkstatt

2. Install dependencies:

   # Basic installation
   uv sync
   
   # With optional features
   uv sync --extra tracing  # Langfuse tracing support
   uv sync --extra memory   # mem0 memory support
   uv sync --all-extras     # All optional features

3. Set up environment variables:

   # Create a .env file
   echo "ANTHROPIC_API_KEY=your_anthropic_api_key_here" >> .env
   echo "TAVILY_API_KEY=your_tavily_api_key_here" >> .env          # Optional for web search
   echo "OPENAI_API_KEY=your_openai_api_key_here" >> .env          # Optional for mem0 memory

3rd Party Services (Optional)

AgentWerkstatt includes a Docker Compose stack with integrated services:

• mem0 - AI memory management system for persistent context
• Langfuse - Observability and tracing platform
• PostgreSQL - Database with pgvector for embeddings
• Neo4j - Graph database for memory relationships
• Redis - Caching and session storage
• MinIO - S3-compatible object storage

To start the services:

# Start all services
docker compose -f third_party/docker-compose.yaml up -d

# Or start specific services
docker compose -f third_party/docker-compose.yaml up -d mem0 neo4j postgres

For detailed setup instructions, see:

• MEM0_SETUP.md - Memory system setup
• LANGFUSE_SETUP.md - Observability setup
• LANGFUSE_INTEGRATION.md - Integration guide

API Keys Setup

Anthropic API Key (Required)

1. Sign up at console.anthropic.com
2. Generate an API key
3. Add it to your .env file as ANTHROPIC_API_KEY

Tavily API Key (Optional, for web search)

1. Sign up at app.tavily.com
2. Get your API key (1,000 free searches/month)
3. Add it to your .env file as TAVILY_API_KEY

OpenAI API Key (Optional, for mem0 memory)

1. Sign up at platform.openai.com
2. Generate an API key
3. Add it to your .env file as OPENAI_API_KEY

Usage

Command Line Interface

Run the interactive CLI:

# Using default configuration (config.yaml)
agentwerkstatt

# Using a custom configuration file
agentwerkstatt --config /path/to/your/config.yaml

Example conversation:

🤖 AgentWerkstatt
==================================================
Loading config from: config.yaml

I'm an example AgentWerkstatt assistant with web search capabilities!
Ask me to search the web for information.
Commands: 'quit'/'exit' to quit, 'clear' to reset, 'status' to check conversation state.

You: What's the latest news about AI developments?
🤔 Agent is thinking...

🤖 Agent: I'll search for the latest AI developments for you.

[Search results and AI summary will be displayed here]

You: clear  # Clears conversation history
🧹 Conversation history cleared!

You: quit
👋 Goodbye!

Programmatic Usage

from agentwerkstatt import Agent, AgentConfig

# Initialize with default config
config = AgentConfig.from_yaml("config.yaml")
agent = Agent(config)

# Or customize the configuration
config = AgentConfig(
    llm={"provider": "claude", "model": "claude-sonnet-4-20250514"},
    tools_dir="./tools",
    verbose=True,
)
agent = Agent(config)

# Process a request
response = agent.process_request("Search for recent Python releases")
print(response)

# Clear conversation history
agent.llm.clear_history()

Command Line Options

The CLI supports the following command line arguments:

• --config - Path to the agent configuration file (default: config.yaml)
• --help - Show help message and available options

Examples:

# Use default configuration
agentwerkstatt

# Use custom configuration file
agentwerkstatt --config my_custom_config.yaml

# Show help
agentwerkstatt --help

Documentation

For comprehensive documentation, please visit our documentation directory:

• 📚 Complete Documentation - Main documentation hub
• 🚀 Getting Started - Installation and quick start guide
• 🏗️ Architecture - Framework design and components
• ⚙️ Configuration - Environment setup and configuration options
• 🛠️ Development Guide - Contributing and extending the framework
• 📖 API Reference - Detailed API documentation

Quick Configuration Reference

Basic configuration in config.yaml:

# LLM Model Configuration
llm:
  provider: "claude"
  model: "claude-sonnet-4-20250514"

# Tools Configuration
tools_dir: "./tools"

# Logging Configuration
verbose: true

# Persona Configuration
personas:
  - id: databot
    name: "DataBot"
    description: "A persona for data analysis and visualization."
    file: "personas/databot.md"

default_persona: "databot"

Environment Variables:

• ANTHROPIC_API_KEY - Required for Claude API access
• TAVILY_API_KEY - Optional, for web search functionality

For complete configuration options, see the Configuration Guide.

Roadmap

Check out our ROADMAP.md to see what's planned for future releases, including:

• 🧠 Multi-LLM Support - OpenAI, Google AI, and local model integration
• ✅ Memory & Persistence - mem0 integration (✅ COMPLETED)
• ✅ 3rd Party Integrations - Observability tools and database services (✅ COMPLETED)
• 🛠️ Advanced Tools - API discovery, file operations, and code execution
• 🤖 Agent Intelligence - Self-reflection, planning, and reasoning capabilities

We welcome feedback and contributions to help shape the project's direction!

Contributing

1. Fork the repository
2. Create a feature branch (git checkout -b feature/amazing-feature)
3. Make your changes
4. Run the quality checks:

   uv run ruff check --fix
   uv run ruff format
   uv run mypy .
   uv run pytest

5. Commit your changes (git commit -m 'Add amazing feature')
6. Push to the branch (git push origin feature/amazing-feature)
7. Open a Pull Request

See CONTRIBUTING.md for detailed guidelines.

License

The license is still under development.

Acknowledgments

• Anthropic for the Claude API
• Tavily for web search capabilities
• mem0 for AI memory management
• Langfuse for observability and tracing
• The open-source community for inspiration and tools

Support

• 📚 Documentation
• 🐛 Bug Reports
• 💬 Discussions
• 🔧 MEM0 Setup Guide
• 📊 Langfuse Integration Guide

---

AgentWerkstatt - Building intelligent agents, one tool at a time. 🚀