dockyter 0.2.2

Run heavy tools in Docker or via an API from Jupyter notebooks

Dockyter

Dockyter is an IPython extension that adds:

• a %%docker cell magic to run whole cells inside Docker containers,
• an optional ! shell redirection so that !cmd runs inside Docker,
• and a pluggable backend system:
  • local Docker daemon (default),
  • or a remote HTTP API backend.

The goal is to run heavy CLI tools packaged as Docker images from notebooks, while keeping the base Python environment light and reproducible.

Typical use cases:

• running ML frameworks, data validation tools, or internal CLIs from Docker images,
• keeping notebook kernels small and simple,
• using the same Dockerised tools across local Jupyter, JupyterHub, and Binder-like deployments,
• delegating container execution to a remote HTTP API instead of the local Docker daemon.

---

Installation

pip install dockyter

Then in a notebook:

%load_ext dockyter

By default, Dockyter expects:

• a working docker CLI on PATH, and
• access to a Docker-compatible daemon (or rootless runtime)

in order to run containers with the Docker backend.

If you use the API backend, the Docker daemon can live on a separate machine: Dockyter just talks HTTP to your API.

---

Backends: Docker daemon vs HTTP API

Dockyter has two backends:

• Docker backend (default) Runs containers by calling the local docker CLI:

  docker run --rm [ARGS] IMAGE bash -lc "cmd"
  

• API backend Sends commands to an HTTP API that you implement and manage.

You can switch backend at runtime:

# Use local Docker daemon (default)
%docker_backend docker

# Use HTTP API backend
%docker_backend api http://127.0.0.1:8000

The current backend and status can be inspected with:

%docker_status

This prints:

• which backend is active (Docker or API),
• whether it appears available,
• the current Docker arguments (image, volumes, etc.),
• whether ! redirection is enabled.

---

API backend contract

The API backend is intentionally small and simple. Dockyter only assumes two endpoints:

1. Health check

   GET /health
   

   • Must return a 2xx status code if the backend is available.
   • Response body is ignored by Dockyter.

2. Command execution

   POST /execute
   Content-Type: application/json
   

   Request body:

   {
     "cmd": "echo hello",
     "args": "ubuntu:22.04 -v /host:/data"
   }
   

   Response body (JSON):

   {
     "stdout": "hello",
     "stderr": ""
   }
   

Dockyter:

• passes the entire cell or ! command as cmd,
• passes the raw %docker / %%docker arguments as args,
• prints stdout to the notebook,
• prints stderr in red if not empty.

What happens inside the API is entirely up to you. A typical implementation:

• receives cmd and args,
• constructs a docker run ... command on the server,
• captures stdout / stderr,
• returns them in JSON.

But the API does not have to use Docker internally; it could use Kubernetes, a job queue, or anything else — Dockyter only cares about the HTTP contract above.

Security responsibility

Dockyter does not implement any security for the API backend.

• Authentication, authorisation, rate limiting, logging, etc. are entirely the responsibility of the API owner.

• The example API server in this repository is not intended for exposure on the public internet. It is a minimal reference implementation for local / trusted environments.

• A real deployment must:

  • protect the API (auth, HTTPS),
  • control which images and arguments are allowed,
  • run on a hardened host.

Dockyter only provides a convenient client for this API from inside notebooks; it is not a security boundary.

---

Basic usage

Cell magic: %%docker (recommended)

%%docker myorg/tool:latest
echo "Hello from inside the container"
pwd

The entire cell is sent to bash -lc inside a single container. All lines share the same shell state:

• cd persists for the rest of the cell,
• environment variables set in one line are visible to the others,
• multi-line scripts, if/for, heredocs, etc. work as expected.

This is the recommended way to run anything non-trivial in Docker from a notebook.

The same syntax works with both backends:

• Docker backend → runs docker run ... locally.
• API backend → sends cmd and args to POST /execute.

---

Line magic: %docker + ! redirection

%docker -v /host/path:/data myorg/tool:latest

Then:

!tool --input /data/file.txt

Here %docker configures Dockyter:

• Docker arguments and image are stored,

• subsequent !cmd calls in that notebook are rerouted to the active backend:

  • Docker backend → docker run --rm [ARGS] IMAGE bash -lc "cmd"
  • API backend → POST /execute with cmd="cmd" and args="[ARGS] IMAGE"

Important behaviour:

• each !cmd runs in a fresh container (or fresh backend execution),

• shell state is not shared between ! calls:

  %docker myimage:latest
  !cd /data
  !pwd   # runs in a new container. Not in /data
  

For anything that relies on cd, multi-line shell logic, or persistent state, prefer %%docker. %docker + ! is best for simple one-shot commands.

---

Commands

• %%docker [DOCKER ARGS...] IMAGE[:TAG] Run the cell content in a single Docker container with the given image/arguments, using the currently selected backend.

• %docker [DOCKER ARGS...] IMAGE[:TAG] Configure “Docker mode” for ! so that each !cmd is executed inside a container via the currently selected backend. (%docker_on is effectively activated.)

• %docker_off Restore the original ! behaviour (no Docker redirection).

• %docker_on Activate Docker mode for ! again, using the last configured image/arguments.

• %docker_status Show the current backend type, its availability, whether ! redirection is enabled, and which image/arguments are currently configured.

• %docker_backend docker Use the local Docker daemon backend. (call %docker_status automatically after)

• %docker_backend api <URL> Use an HTTP API backend at the given base URL (for example http://127.0.0.1:8000). (call %docker_status automatically after)

---

Binder / JupyterHub integration (high-level)

In BinderHub / JupyterHub, Dockyter can be used in a few ways:

1. Extension-only (safest default)

• Install dockyter in the image (e.g. via requirements.txt).
• Users do %load_ext dockyter in notebooks.
• If docker is not available in the container, Dockyter just reports it and does not crash.

This is the right choice for public / untrusted notebook environments.

2. Direct Docker daemon access (trusted only)

You can expose a Docker runtime inside user containers and use the Docker backend.

This is very dangerous for public notebooks:

• Users can bypass Dockyter and run !docker ... directly,
• including flags like --privileged or --network=host.

Only consider this if users are trusted and the platform is carefully locked down.

3. API backend (recommended for untrusted users)

A safer option for public or multi-tenant setups:

• Do not expose docker in user containers.
• Run a separate, hardened Dockyter-compatible API backend.
• In notebooks, use:

  %docker_backend api https://your-secure-api.example.com

The API is then responsible for all security (auth, allowed images/flags, rate limiting, etc.).

Dockyter is not a security boundary; it only provides convenience and light guardrails. Real isolation must come from the surrounding platform or the API implementation.

---

Examples

This repository includes several example notebooks and an example API server:

• docs/exemples/01_local_cli.ipynb Run simple commands in a local Docker image (%%docker basics).

• docs/exemples/02_ml_tool_in_docker.ipynb Use a real ML framework (e.g. PyTorch) inside Docker, keeping the notebook kernel light.

• docs/exemples/03_api_backend.ipynb Use Dockyter with the API backend, switching with %docker_backend api and running commands via the HTTP API instead of the local Docker daemon.

• docs/api_example/server.py Minimal example of a Dockyter-compatible API implemented with FastAPI + Uvicorn. This is a reference implementation for local / trusted environments only.

---

Tests

Dockyter has:

• Unit tests for:

  • DockerBackend and APIBackend (using monkeypatch for subprocess / requests),
  • the IPython magics layer.

• Integration tests that:

  • execute the example notebooks via nbconvert,
  • start the example API server for API backend tests,
  • fail if Dockyter prints error messages in red.

You can run them locally with:

# Unit tests only
uv run pytest -m "not integration"

# Full test suite (including notebook + API integration)
uv run pytest