marimo Notebook

marimo is a reactive Python notebook — cells form a dataflow graph, and changing one cell automatically re-runs its dependents. This skill encodes conventions for how to run marimo and how to build a canonical scatter-plot-with-selected-rows-table exploration UI.

Running marimo

Run marimo listening on all interfaces, with no token, in sandbox mode:

marimo edit --host 0.0.0.0 --no-token --sandbox notebook.py
# view-only (app mode):
marimo run  --host 0.0.0.0 --no-token --sandbox notebook.py

Why each flag:

Flag	Purpose
--host 0.0.0.0	Bind to every interface so a phone, tablet, or another laptop on the LAN can hit http://<host-ip>:<port>. 127.0.0.1 (the default) only reaches localhost.
--no-token	Skip the URL token, so the link is paste-able. Also: servers started with --no-token register in marimo's local server registry, which is what tooling like marimo-pair relies on to auto-discover sessions.
--sandbox	Run the notebook in an isolated uv-managed venv driven by inline # /// script dependency metadata at the top of the .py file. Keeps per-notebook deps out of the system/project env.

Use --no-token only on trusted networks (home LAN, local dev). On anything shared, leave the token on and use MARIMO_TOKEN to pass it to tooling.

Inline dependencies for sandbox mode

With --sandbox, put PEP 723 script metadata at the top of the notebook so uv knows what to install:

# /// script
# requires-python = ">=3.11"
# dependencies = [
#     "marimo",
#     "polars",
#     "jupyter-scatter",
#     "pandas",
# ]
# ///

import marimo
app = marimo.App()

Edit this block when you add imports — sandbox mode will refuse to import anything that isn't declared here.

Scatter plots: always use jscatter

Default to jupyter-scatter (import jscatter) for scatter plots. It's GPU-accelerated, handles millions of points, has a built-in lasso selection tool, and bridges cleanly into marimo's reactive graph through traitlets. Reach for matplotlib / plotly / altair only for small, non-interactive cases.

Long-press activates the lasso

Explicitly enable long-press as the lasso initiator:

scatter.lasso(on_long_press=True)

on_long_press=True is actually jupyter-scatter's default, but set it explicitly — the keyboard-modifier initiator (shift+drag) is unreliable on some machines/keyboards/OSes, so readers of the notebook should be able to see at a glance that long-press is the supported gesture and not have to guess at a keyboard shortcut.

Always show a table of the selected rows below the plot

Bridge the jscatter widget's selection trait into marimo's reactive graph with mo.state + .observe(...), then render a mo.ui.table in the next cell that filters the dataframe by the current selection. The table updates live as the user lassos.

@app.cell
def _(df, mo):
    import jscatter

    # jupyter-scatter wants pandas
    pdf = df.to_pandas() if hasattr(df, "to_pandas") else df

    scatter = (
        jscatter.Scatter(x="x", y="y", data=pdf)
        .height(500)
        .color(by="category")
        .legend(True)
        .tooltip(True)
    )
    scatter.lasso(on_long_press=True)

    # bridge jscatter's `selection` traitlet into marimo's reactive graph
    get_selection, set_selection = mo.state(scatter.widget.selection)
    scatter.widget.observe(
        lambda _: set_selection(scatter.widget.selection),
        names=["selection"],
    )
    scatter.widget
    return get_selection, pdf

@app.cell
def _(get_selection, mo, pdf):
    sel = get_selection()
    mo.ui.table(
        pdf.iloc[sel] if len(sel) else pdf.head(0),
        page_size=25,
    )
    return

Notes:

• Use scatter.widget, not scatter, for the observe/display. Scatter is a configuration object; scatter.widget is the anywidget instance that owns the selection traitlet.
• Return get_selection (the getter), not the state value. Returning the scalar freezes the downstream cell at construction time; returning the getter lets the downstream cell re-read on every reactive tick.
• Empty-selection handling: show pdf.head(0) (empty frame with headers) rather than the full table when nothing is selected — the point of the table is to reflect the lasso, not to be a second data browser.
• mo.ui.table vs quak.Widget: mo.ui.table is the default. Reach for mo.ui.anywidget(quak.Widget(df)) when you want SQL-style filtering, faceting, and column stats inside the table itself.

Notebook skeleton

# /// script
# requires-python = ">=3.11"
# dependencies = ["marimo", "polars", "jupyter-scatter", "pandas"]
# ///

import marimo
app = marimo.App(width="medium")


@app.cell
def _():
    import marimo as mo
    import polars as pl
    return mo, pl


@app.cell
def _(pl):
    df = pl.read_parquet("data.parquet")
    return (df,)


# ... scatter cell + selected-rows table cell from above ...


if __name__ == "__main__":
    app.run()

Gotchas

• Don't edit the .py file while marimo edit is running on it. The kernel owns the file; writes from the outside will be clobbered or ignored. Use the marimo UI, or stop the server first.
• Sandbox mode is per-notebook. Each --sandbox invocation resolves its own venv from the inline script metadata. Don't try to share a venv across notebooks via --sandbox; just run them with the project's regular env.
• --no-token leaks your notebook to the LAN. On coffee-shop wifi or a shared office network, either drop --host 0.0.0.0 or keep the token.
• Widget traitlets live outside the reactive graph. Setting scatter.widget.selection = [...] directly from Python works, but won't flow into marimo state unless you've wired an .observe(...) bridge like the one above.