Python Template

Generate production-ready Python and AI stacks by combining frameworks, data engines, tooling, and cloud deployment.

Start with a simple library, CLI, or API. Or compose an agent, RAG system, training workspace, inference service, data layer, UI, quality stack, and deploy target without assembling the project conventions yourself.

Quick start

Install nothing globally beyond uv:

uvx copier copy --trust gh:leynier/python-template my-project
cd my-project
uv run pytest

Copier presents 12 editable recipes plus a custom layer-by-layer path. It initializes Git and installs the selected dependencies after rendering.

Want the small version? These remain first-class choices with no AI dependency:

Python Library     typed, buildable package
Typer CLI          tested command-line application
FastAPI API        production-shaped JSON service

Compose the stack you need

The generator resolves each choice as a layer, so infrastructure can change without replacing the application framework and model providers can change without coupling them to the embedding provider.

Layer	Examples
Workload	library, CLI, API, web, TUI, MCP, agent, RAG, inference, training, hybrid
Framework	FastAPI, Flask, FastMCP, Pydantic AI, LangGraph, LlamaIndex, Lingo, Transformers
Interface	Streamlit, Gradio, Chainlit, Textual, NiceGUI, FastHTML, Violetear, JupyterLab
Model provider	OpenAI, Anthropic, Gemini, Bedrock, Azure OpenAI, Ollama, OpenRouter and more
Embeddings	hosted provider or Sentence Transformers, selected independently
Data	SQL, document, vector, graph, and cache roles with one engine per role
Auth	API key, OAuth/OIDC, or Supabase Auth
Training	Lightning, Datasets, Accelerate, PEFT, TRL, Optuna
Serving	BentoML, LiteLLM, vLLM, Ollama, Ray Serve
MLOps and quality	Prefect, Dagster, DVC, MLflow, Ragas, DeepEval, OpenTelemetry and more
Deploy	Docker plus 14 managed or cloud targets
IaC	none, Pulumi, or Terraform for supported cloud targets

The catalog currently contains 126 components with explicit workload, Python, and support-tier metadata. See the complete component reference.

The 12 presets

Presets are useful starting points, not locked bundles. Every answer remains editable during generation.

Preset	Starting stack
python-library	Typed publishable package
typer-cli	Typer command-line app
fastapi-api	FastAPI JSON service
fastmcp-server	FastMCP tool server
pydantic-ai-openai	Pydantic AI + Harness + OpenAI
google-adk-gemini	Google ADK + Gemini
strands-bedrock	Strands Agents + Bedrock
langgraph-anthropic-api	LangGraph + Anthropic + FastAPI
llamaindex-rag	LlamaIndex + OpenAI + Pinecone + Gradio
local-lingo-app	Lingo + Ollama + Beaver + Violetear
litellm-gateway	LiteLLM + Redis + Docker
hf-finetuning	Transformers + PEFT/TRL + MLflow + BentoML

Preselect one while keeping the remaining questions interactive:

uvx copier copy --trust \
  -d preset=fastmcp-server \
  gh:leynier/python-template my-tools

Deploy without rebuilding the project

Every deployment choice keeps a portable Docker base. Choose one target among Docker, Render, Fly.io, Vercel, Railway, Hugging Face Spaces, Modal, RunPod, BentoCloud, AWS ECS, SageMaker, Cloud Run, Vertex AI, Azure Container Apps, or Azure ML. The six cloud targets can additionally generate Pulumi or Terraform.

Managed inference variants expose consistent health and prediction contracts, while Modal, RunPod, and BentoCloud receive native SDK adapters.

Skills for AI coding agents

This repository is directly discoverable by the open Agent Skills CLI:

npx skills add leynier/python-template --list
npx skills add leynier/python-template --skill compose-python-stack

Repository skills help an agent compose, maintain, and validate stacks. Every generated project also includes a common workflow skill and conditionally adds AI and deployment skills matching its selected layers.

What every generated project gets

• A src/ layout, typed package marker, bounded dependencies, and committed uv.lock.
• Ruff, pytest, deptry, advisory ty, pre-commit, coverage, and Poe tasks.
• CI across Linux, macOS, and Windows, with least-privilege permissions and pinned actions.
• CodeQL, zizmor, Dependabot, issue forms, security policy, contribution guide, and changelog.
• Optional Zensical docs, PyPI Trusted Publishing, Docker, devcontainer, and agent instructions.
• A saved .copier-answers.yml so later template releases can be applied with uvx copier update.

Support model

Catalog entries use three intentionally visible tiers:

• stable: open-source component exercised without external credentials.
• platform: hosted or cloud integration whose generated contract is tested offline; real deployment still requires the user's account and secrets.
• experimental: useful but evolving integration with a narrower compatibility promise.

The catalog favors maintained projects that add a distinct layer or a clear end-to-end recipe.

Develop the template

uv sync --all-groups
uv run python scripts/compile_catalog.py --check
uv run ruff check .
uv run ruff format --check .
uv run pytest -n auto
uv run pytest tests/test_compile_catalog.py --cov=scripts --cov-branch
uv run --group docs zensical serve

The test suite renders compatible combinations, runs real generated toolchains, parses deployment artifacts, checks copier update, and exercises representative AI/ML vertical slices. Hosted CI repeats the generated-project tests on Linux, macOS, and Windows. CI enforces 100% branch coverage for the repository's catalog compiler; generated projects enforce at least 95% coverage for their own src/ package through uv run poe cov.

Read the documentation, the contribution guide, or the security policy.

License

Python Template is collaborative open source under the MIT license.