Python AI Kit

A production-ready framework for building AI agents that actually work in production.

🚀 Quick Start

First, you need to have copier installed. We suggest to do it with uv (https://docs.astral.sh/uv/getting-started/installation/#installation-methods):

uv tool install copier

[!TIP] You can also just use uvx before copier copy command if you don't want to install this tool.

To use this template, run:

copier copy https://github.com/the-momentum/python-ai-kit $TARGET_DIRECTORY --trust
# uvx copier copy https://github.com/the-momentum/python-ai-kit $TARGET_DIRECTORY --trust

Instead of $TARGET_DIRECTORY put directory name where you want to generate your project. You can use indirect path.

[!IMPORTANT]

• copier copy DOESN'T create new directory, it just copies files to $TARGET_DIRECTORY
• it will copy version from latest release
• if you want to fetch from latest unreleased commit, use flag -r HEAD (also accepts remote branches names)

If you've already created project and templates get updated, you can still update your project:

copier update --trust --defaults
# run this inside ypur project directory

[!IMPORTANT]

• your project HAS to be a git repo in order to use this command
• omit --defaults flag, if you want to update answers (like python_versions)
• you can also use flag -r here to update with template version from unreleased versions

🔧 Optional add-ons

Some setup options require extra tools to be installed beforehand. If you skip them, the template will fail when that option is selected.

DVC (Data Version Control)

• Requires: dvc installed (e.g. uv tool install dvc) and, when enabled, AWS CLI with a profile that has access to your DVC S3 bucket.
• During setup: You can enable DVC when prompted; you will then provide the full S3 path (starting with s3://) for the remote and the AWS region.
• After generation: Set your AWS profile locally in the project root (stays out of Git):
  dvc remote modify --local <remote-name> profile <your-aws-profile> (use the remote name you chose during setup)

Features

Core Framework

• Multi-agent orchestration - Built-in support for coordinating multiple specialized agents with defined workflows and error handling
• State management - Persistent memory across sessions using dedicated database storage and thread preservation
• Observability by default - Integrated logging, tracing, and monitoring with Pydantic Logfire or Opik
• Structured prompt management - Modular, versioned prompts with POML patterns instead of monolithic strings

Development & Testing

• Automated evaluation pipeline - Integrated Ragas/Opik evaluators with quantifiable metrics, not "by feel" testing
• Prompt versioning system - Track changes, rollback, and compare prompt iterations using database or JSON storage
• Built-in testing patterns - Standard test structures for agentic logic, not just code structure
• Code standards enforcement - Pre-configured linters and formatters that understand AI agent patterns

Production Readiness

• Security hardened - Fernet encryption for API keys, SOPS standard support, no exposed credentials
• Artifact management - Proper handling of RAG contents, static files, and model storage without cluttering repos
• Workflow control - Template-based routing with custom error handling and predictable execution paths
• MLOps pipeline - Deployment patterns for custom models and retraining workflows

Integration & Compatibility

• Curated tool ecosystem - Pre-integrated best-of-breed tools that actually work together
• Framework flexibility - Strong core with optional features, avoiding both barebones implementations and bloated abstractions
• Standard interfaces - Consistent APIs across components, minimal manual adjustments needed

Why Use Python AI Kit

You're tired of stitching tools together. Every AI agent project feels like forcing incompatible blocks to work. You spend more time debugging integrations than building features.

You can't review prompt changes with confidence. Merge requests for prompts are guesswork. Small changes cause unpredictable behavior. You have no baseline to compare against.

Your agents lose context between sessions. Users complain that the agent "forgets" previous conversations. You've bolted on hacky state management that breaks under load.

Testing is a manual nightmare. You ask your agent arbitrary questions and judge responses subjectively. There are no metrics. You can't prove your changes made things better or worse.

You can't explain why the agent did something. When things go wrong in production, you have no visibility into the decision chain. This is a non-starter in regulated industries.

You're rebuilding patterns from scratch every time. There's no standard way to structure agents. Every project starts at zero. Code reviews are inconsistent because there's no established patterns.

Security is an afterthought. API keys in environment variables. Secrets committed to repos. You know it's wrong but there's no easy alternative baked in.

This framework solves these problems by integrating proven solutions into a cohesive platform. Not another thin wrapper - battle-tested patterns for the entire development lifecycle.

📚 Documentation

• API Architecture - Learn about the database, repositories, services, and API design patterns
• Agents - Instructions for running and working with AI agents
• Develop your agent workflow - patterns you might use to develop your own workflow
• MCP Server - MCP Server template documentation

🎯 Project Types

This template generates projects optimized for:

• Microservice API - Lightweight, focused services with minimal dependencies
• Monolith Service API - Full-featured applications with comprehensive architecture layers
• MCP Server - Model Context Protocol servers for AI tool integration
• AI Agent - Intelligent agent systems with workflow and tool management

Each generated project includes modern Python tooling, comprehensive testing, and production-ready architecture patterns.

📁 Project Structure

The generated project includes:

• FastAPI-based API with proper architecture layers
• SQLAlchemy database models and repositories
• Service layer with error handling
• AI agent integration with Streamlit GUI
• Comprehensive testing setup
• Modern Python tooling (uv, ruff, etc.)

For detailed information about specific components, please refer to the linked documentation pages above.