Make kernel-local HTTP and WebSocket servers reachable from the notebook browser. Zero user config.

If your library runs a web server inside a Jupyter kernel (a tile server, a trame app, Bokeh, Dash, a custom debug UI, anything bound to 127.0.0.1:<port>), your users hit the same wall every time. Works locally. Breaks on JupyterHub, MyBinder, VS Code Remote, Colab, Shiny. The usual fix is a README paragraph telling users to install jupyter-server-proxy, export LIBRARY_CLIENT_PREFIX='proxy/{port}', and prepend $JUPYTERHUB_SERVICE_PREFIX on Hub. Most users skip it, get a broken notebook, and file an issue.

jupyter-loopback replaces that paragraph. Libraries register once. End users configure nothing.

Who this is for

Library authors who spin up an HTTP or WebSocket server inside a Jupyter kernel and need the browser to reach it without asking every user to configure a proxy. If you've ever written "make sure to install jupyter-server-proxy and set FOO_CLIENT_PREFIX='proxy/{port}'" in your README, this is for you.

Install

pip install jupyter-loopback            # HTTP/WS proxy
pip install jupyter-loopback[comm]      # + anywidget comm bridge fallback

The 30-second demo

docker build -t jupyter-loopback-demo .
docker run --rm -it -p 8888:8888 jupyter-loopback-demo

Open the printed token URL, run example.ipynb. You'll see:

• A JSON response fetched through the proxy.
• A red square rendered inline (binary-body correctness check).
• A live WebSocket echo box you can type into.

All of it flowing through <base_url>/loopback-demo-proxy/<port>/… with no config.

For library authors

Suppose your library is mylib and it spins up a server at 127.0.0.1:<port> inside the kernel. Three files wire it up.

1. Server-side: register the proxy

# mylib/_jupyter/__init__.py
from jupyter_loopback import setup_proxy_handler

def _jupyter_server_extension_points():
    return [{"module": "mylib._jupyter"}]

def _load_jupyter_server_extension(server_app):
    setup_proxy_handler(server_app.web_app, namespace="mylib")

2. Auto-enable the extension

Ship jupyter-config/jupyter_server_config.d/mylib.json:

{
  "ServerApp": {
    "jpserver_extensions": {
      "mylib._jupyter": true
    }
  }
}

And wire it into pyproject.toml:

[tool.setuptools.data-files]
"etc/jupyter/jupyter_server_config.d" = [
  "jupyter-config/jupyter_server_config.d/mylib.json",
]

3. Kernel-side: build browser-reachable URLs

from jupyter_loopback import autodetect_prefix

def browser_url(port: int, path: str) -> str:
    prefix = autodetect_prefix("mylib")  # None outside Jupyter
    if prefix is None:
        return f"http://127.0.0.1:{port}/{path.lstrip('/')}"
    return f"{prefix.format(port=port)}/{path.lstrip('/')}"

That's it. In a local Python REPL, autodetect_prefix returns None and you hit 127.0.0.1 directly. In JupyterLab, Hub, Binder, or any jupyter-server environment, it returns mylib-proxy/{port} (with any per-user Hub prefix already attached) and the browser loads through the proxy.

For users on VS Code Remote, Colab, Shiny, Solara, marimo

These frontends don't run a jupyter-server, so the HTTP proxy above isn't available. They do have kernel comms (the WebSocket the notebook widgets use). jupyter-loopback ships an anywidget that tunnels request/response pairs over that comm channel.

Users enable it once at the top of a notebook:

import jupyter_loopback
jupyter_loopback.enable_comm_bridge()

Library authors register handlers:

from jupyter_loopback import on_request

@on_request("mylib", "get_tile")
def _(data, buffers):
    z, x, y = data["z"], data["x"], data["y"]
    return {"ok": True}, [make_tile(z, x, y)]   # (json, buffers)

Frontend code calls through window.__jupyter_loopback__:

const { status, data, buffers } = await window.__jupyter_loopback__.request(
  "mylib",
  "get_tile",
  { z: 8, x: 71, y: 110 },
);
if (status === "ok") {
  const blob = new Blob([buffers[0]], { type: "image/png" });
  imgElement.src = URL.createObjectURL(blob);
}

The bridge carries JSON plus binary buffers. Use it for anything request/response shaped. Streaming and server-push are out of scope; use the WS proxy for that.

What works where

Environment	Path	User does
Local notebook	direct loopback	nothing
JupyterLab / Notebook 7+	HTTP/WS proxy	nothing
JupyterHub / MyBinder	HTTP/WS proxy	nothing
VS Code Remote	comm bridge	enable_comm_bridge()
Google Colab	comm bridge	enable_comm_bridge()
Shiny for Python / Solara	comm bridge	enable_comm_bridge()
marimo	comm bridge	enable_comm_bridge()

Relationship to jupyter-server-proxy

jupyter-server-proxy proxies arbitrary subprocesses. It handles HTTP/WS wire formatting, subprocess lifecycle, URL rewriting, auth.

jupyter-loopback does less, on purpose:

• Proxies loopback only. No cross-host surface.
• No subprocess management. You bring your own server on any port.
• Autodetects the URL prefix from Jupyter's own env vars.
• Ships a comm-based fallback for frontends without a jupyter-server.

The two can coexist. Set LIBRARY_CLIENT_PREFIX explicitly and jupyter-loopback's autodetect steps out of the way.

Status

Extracted from localtileserver after the same pattern solved its long tail of remote-Jupyter issues. Generalized so other libraries can adopt it without reinventing the wheel.

See ARCHITECTURE.md for the design walkthrough. MIT licensed.