Jupyter containers that feel local

JovyKit

A venv-like command line for project-local JupyterLab environments: readable generated files, uv-locked dependencies, layered notebook images, and Docker Compose kept politely out of your way.

View on GitHub Read the wiki

$ jovy init .jovy --image base && jovy add pandas scikit-learn && jovy run

Local by default

A project gets jovy.toml, .jovy/, and work/. The shape is easy to inspect and easy to throw away.

Reproducible overlays

Direct package specs live in TOML, uv compiles a lockfile, and the generated Containerfile installs exactly that environment.

Comfortable operations

Run attached, start detached, follow logs, open a shell, or use the full-screen terminal dashboard when you want a cockpit.

Workflow

From empty folder to working notebook.

Init jovy init .jovy --image base --gpus auto

Add jovy add pandas plotly scikit-learn

Lock jovy install --upgrade

Run jovy run

Why it exists

Docker power without Docker-shaped daily work.

Research projects often start as notebooks, then grow into packages, scripts, and repeatable environments. JovyKit keeps that path grounded in normal files instead of one-off container state.

You still get Docker Compose, GPU reservations, custom volumes, extra apt packages, and a generated overlay image. You just operate them through commands that read like the project workflow.

Image levels

Start small. Scale the stack when the work asks.

minimal

Jupyter runtime and core scientific Python.

base

Everyday data science, statistics, ML, and local data access.

extended

NLP, time series, distributed compute, and API tooling.

full

Heavier AI, graph, geospatial, big-data, and research tooling.

jovy status --json
{
  "environment": ".jovy",
  "image": "jovykit-research:local",
  "jupyter_url": "http://127.0.0.1:8888/lab",
  "running": true
}

Scriptable when it matters

Clean commands for humans and CI.

Status output can be JSON. Logs can be tailed with timestamps. Shell commands run inside the notebook service. The dashboard uses the same command model, so interactive work and automation stay in step.