nuv

Scaffold opinionated uv Python projects — tests passing, linting green, out of the box.

nuv new my-tool

That's it. You get a working project with click, logging, test coverage, ruff, and ty — all green from commit zero.

Install

uv tool install nuv

Run without installing

uvx nuv new my-tool

Usage

nuv new <name>                              # creates ./<name>/, syncs deps, prints tool install command
nuv new <name> --at <path>                  # creates at an explicit path
nuv new <name> --archetype spark            # PySpark 4 project with notebooks
nuv new <name> --archetype fastapi          # FastAPI + Granian + Docker
nuv new <name> --archetype polars           # Polars + DuckDB + Delta Lake for local data work
nuv new <name> --archetype ds               # Data science: thin core + commented DS/ML/LLM catalog
nuv new <name> --python-version 3.13        # override default Python version
nuv new <name> --install none               # scaffold + sync, skip tool install
nuv new <name> --install command-only       # log install command, do not execute (default)
nuv new <name> --keep-on-failure            # keep generated files if sync/install fails

Archetypes

script (default)

A single-file CLI tool with click and logging.

nuv new my-tool

my-tool/
├── main.py          # click + logging
├── _logging.py
├── pyproject.toml   # pytest, ruff, ty, uv
├── README.md
└── tests/
    ├── __init__.py
    └── test_main.py

spark

A PySpark 4 project with src-layout, chispa testing, and dual notebooks (Jupyter + marimo).

nuv new my-spark-app --archetype spark

my-spark-app/
├── main.py
├── pyproject.toml
├── README.md
├── src/my_spark_app/
│   ├── __init__.py
│   ├── _logging.py
│   ├── config.py
│   ├── session.py
│   └── jobs/
│       ├── __init__.py
│       └── example.py
├── tests/
│   ├── __init__.py
│   ├── conftest.py
│   └── test_example.py
└── notebooks/
    ├── explore.ipynb
    └── explore_marimo.py

Default Python version: 3.13 (PySpark 4 compatibility).

uv run pytest          # 8 tests, passing
uv run ruff check .    # clean
uv run ty check        # clean

Notebooks are an optional dependency group:

uv sync --group notebooks
uv run jupyter lab notebooks/
uv run marimo run notebooks/explore_marimo.py

fastapi

A production-ready FastAPI project with Granian ASGI server, Pydantic settings, and a multi-stage Dockerfile.

nuv new my-api --archetype fastapi

my-api/
├── main.py                         # Granian server entry point
├── pyproject.toml
├── README.md
├── Dockerfile                      # multi-stage build
├── .dockerignore
├── src/my_api/
│   ├── __init__.py
│   ├── app.py                      # FastAPI factory with lifespan
│   ├── config.py                   # Pydantic settings from env vars
│   ├── _logging.py
│   ├── dependencies.py             # shared FastAPI deps
│   └── routes/
│       ├── __init__.py
│       └── health.py               # /healthz endpoint
└── tests/
    ├── __init__.py
    ├── conftest.py                  # async httpx client fixture
    └── test_health.py

Default Python version: 3.13 (Delta Lake wheel compatibility).

uv run pytest          # passing
uv run ruff check .    # clean
uv run ty check        # clean

Run locally:

uv run python main.py
uv run python main.py --host 0.0.0.0 --port 8000

Docker:

docker build -t my-api .
docker run -p 8000:8000 my-api

polars

A single-node data project with Polars, DuckDB, and Delta Lake. For local analysis, ETL, and feature engineering when Spark is overkill.

nuv new my-pipeline --archetype polars

my-pipeline/
├── main.py                          # Click CLI entry point
├── pyproject.toml
├── README.md
├── data/
│   ├── raw/                         # input datasets (gitignored)
│   └── features/                    # derived datasets (gitignored)
├── src/my_pipeline/
│   ├── __init__.py
│   ├── _logging.py
│   ├── _io.py                       # read/write CSV, Parquet, JSON, Delta + show/glimpse helpers
│   ├── _db.py                       # DuckDB SQL → Polars DataFrame
│   ├── config.py                    # Pydantic settings
│   └── main.py
├── tests/
│   ├── __init__.py
│   ├── conftest.py
│   └── test_io.py                   # roundtrip tests for every supported format
└── notebooks/
    └── explore.py                   # marimo

Default Python version: 3.14.

uv run pytest          # passing, ≥90% branch coverage enforced
uv run ruff check .    # clean
uv run ty check        # clean

Run locally:

uv run python main.py --help
uv run python main.py --log-level INFO

Notebooks are an optional dependency group:

uv sync --group notebooks
uv run marimo edit notebooks/explore.py

Use this when:

• You want fast single-node dataframes without the JVM (Polars instead of PySpark).
• You want to mix Polars expressions with SQL on the same data — _db.sql("...") returns a Polars DataFrame via DuckDB.
• You're reading or writing Parquet, CSV, JSON, or Delta Lake tables.
• You want a marimo notebook for exploration with the I/O helpers pre-wired.

Pick spark instead when the dataset doesn't fit on one machine, or you need a long-running cluster.

ds

A batteries-included data science project. Manifesto: all of what you need or want, but off by default.

nuv new my-ds-project --archetype ds

pyproject.toml ships a deliberately thin active core — numpy, pandas, Arrow, a Click CLI, and typed config — alongside a large, curated, commented-out catalog of the rest of the modern stack: classical ML (scikit-learn, scipy, statsmodels, XGBoost/LightGBM/CatBoost), deep learning & neural nets (PyTorch, TensorFlow, Keras, JAX/Flax/Optax), LLMs & NLP (transformers, datasets, accelerate, peft, vLLM, OpenAI/Anthropic clients, LangChain, LlamaIndex, spaCy), vector stores, viz, and experiment tracking (MLflow, Weights & Biases, DVC, Hydra, Ray). Uncomment a line, run uv sync, and uv resolves it against everything already locked.

my-ds-project/
├── main.py                          # Click CLI entry point
├── pyproject.toml                   # thin core + commented catalog
├── src/my_ds_project/
│   ├── __init__.py
│   ├── _logging.py
│   ├── config.py                    # Pydantic settings (paths, seed, ...)
│   ├── data.py                      # CSV/Parquet/JSON I/O + quick-look helpers
│   └── main.py
├── tests/
│   ├── __init__.py
│   ├── conftest.py
│   └── test_data.py
├── notebooks/
│   ├── explore.ipynb                # Jupyter / IPython
│   └── explore_marimo.py            # marimo
├── data/
│   ├── raw/                         # inputs (gitignored)
│   └── processed/                   # derived data (gitignored)
└── models/                          # trained artifacts (gitignored)

Default Python version: 3.13 (broadest compatibility across the DS/ML ecosystem).

uv run pytest          # passing, 100% coverage
uv run ruff check .    # clean
uv run ty check        # clean

Notebooks ship in two flavors and are an optional dependency group (Jupyter/IPython and marimo):

uv sync --group notebooks
uv run jupyter lab notebooks/                  # Jupyter / IPython
uv run marimo edit notebooks/explore_marimo.py # marimo

The generated .gitignore is comprehensive for AI/ML work: datasets, model weights and checkpoints (*.safetensors, *.ckpt, *.gguf, ...), experiment-tracking dirs (mlruns/, wandb/, lightning_logs/), notebook and library caches, and .env secrets.

Quality out of the box

Every generated project ships with these tools configured and green:

By default, nuv new logs the command you can run to install the generated project as a tool:

uv tool install --editable <project-path>

Why

uv init produces a stub. The gap between that and "actually writing code" is annoying when you create projects frequently. nuv closes it.

Publishing to PyPI

This project uses trusted publishing from GitHub Actions.

1. Bump [project].version in pyproject.toml.
2. Push a matching tag: vX.Y.Z.
3. GitHub Actions builds and publishes to PyPI.

After release:

uv tool install nuv