marimo

Reactive Python notebooks stored as pure .py files — cells form a dependency DAG, so changing one cell automatically reruns its dependents (or marks them stale under the lazy runtime) and hidden state

SKILL.md

GitHub

marimo — reactive Python notebooks

Overview

marimo is a reactive notebook: cells form a DAG based on which variables each cell defines and reads. Run a cell and every dependent cell reruns (or is marked stale) — no hidden state, no out-of-order execution bugs. Notebooks are pure Python files that diff cleanly in git and double as apps and scripts. Pairs with [[pandas]]/[[polars]] for data work and [[matplotlib]] for plots; project tooling in [[modern-python]].

uv add marimo            # or: uv tool install marimo
marimo tutorial intro    # interactive intro
marimo edit nb.py        # create/edit a notebook (browser editor)
marimo run nb.py         # serve as a read-only web app (code hidden by default)
marimo convert nb.ipynb -o nb.py   # migrate from Jupyter
marimo check nb.py       # lint/format notebook files

Verified against marimo 0.23.16 (Python >= 3.10).

File format — a notebook is a Python file

import marimo

app = marimo.App()

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

@app.cell
def _(mo):
    n = mo.ui.slider(1, 100, value=10, label="n")
    n                      # last expression = cell output
    return (n,)

@app.cell
def _(n):
    total = sum(i**2 for i in range(n.value))
    print(total)
    return (total,)

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

Each cell is a function: parameters are the variables it reads, the return tuple is what it defines. The editor maintains this wiring for you — hand-edit freely, marimo re-derives the DAG from the code. python nb.py executes it top-of-DAG-down like a script.

Reactivity rules

  • A variable may be defined in only one cell. Duplicate definitions are a MultipleDefinitionError — the file won't run. Prefix throwaway names with _ to make them cell-local (_x, _fig).
  • Reading a variable creates the edge; mutations are not tracked — lst.append(...) in another cell won't retrigger dependents. Create new objects instead of mutating.
  • Cell order in the file doesn't matter; execution order comes from the DAG.
  • mo.stop(condition, mo.md("waiting...")) short-circuits a cell; mo.md(f"...") renders markdown with interpolated values.

UI elements (mo.ui)

n = mo.ui.slider(1, 100, value=10)       # assign to a GLOBAL, display it
text = mo.ui.text(placeholder="query")
pick = mo.ui.dropdown(["a", "b"], value="a")
run = mo.ui.run_button()                  # gate expensive cells
mo.vstack([n, text, pick, run])           # last expression: renders the elements

Interacting with an element updates element.value and reruns cells that read it — no callbacks. Two rules: the element must be assigned to a global variable and displayed (as cell output or inside mo.md/layout), and .value cannot be read in the cell that creates the element (a cell can't depend on itself). Others: mo.ui.table, mo.ui.dataframe, mo.ui.file, mo.ui.form, mo.ui.altair_chart, mo.ui.checkbox, mo.ui.radio, mo.ui.date, mo.ui.chat, mo.ui.anywidget.

SQL cells

df = mo.sql(f"SELECT city, count(*) FROM users_df GROUP BY city")

SQL cells (uv add "marimo[sql]" for DuckDB + friends) query dataframes, CSVs, or attached databases in place; the result is a dataframe, and {python_expr} values interpolate into the query. Reactivity crosses the language boundary — the SQL cell reruns when the cell defining users_df reruns (rebinding, not in-place mutation). See [[duckdb-docs]] for SQL specifics.

Apps, scripts, exports

marimo run nb.py --host 0.0.0.0 -p 8080   # deploy as interactive app
python nb.py                               # run as a plain script
marimo run app.py -- --arg value           # pass CLI args through
marimo export html-wasm nb.py -o out.html  # self-contained, runs in browser (Pyodide)
marimo export ipynb nb.py -o nb.ipynb      # also: html, md, pdf, script

Sandboxed notebooks (inline dependencies)

marimo edit --sandbox nb.py    # also: run --sandbox, new --sandbox (requires uv)

--sandbox runs the notebook in an isolated venv and records imports as PEP 723 inline metadata in the file header, so a single .py file carries its environment — uv run nb.py recreates the environment anywhere:

# /// script
# requires-python = ">=3.11"
# dependencies = ["pandas==3.0.5", "altair==6.2.2"]
# ///

marimo vs Jupyter

Prefer marimo when you want git-diffable notebooks, guaranteed execution order (reproducibility), notebooks that ship as apps/scripts without rewriting, or interactive UI without widget callback plumbing. Prefer Jupyter when you depend on its ecosystem (nbconvert pipelines, JupyterHub, kernels for other languages) — or convert with marimo convert.

Gotchas

  • MultipleDefinitionError: same name defined in two cells. Rename, or use a _-prefixed local.
  • UI element seems inert: it wasn't assigned to a global, wasn't displayed, or you read .value in its defining cell.
  • Dependents don't rerun after mutation: marimo tracks definitions, not mutations. Rebind (df = df.assign(...)) instead of mutating in place.
  • Expensive cells rerun too eagerly: gate them behind mo.ui.run_button(), use mo.stop(), or switch runtime to lazy in settings.
  • marimo run hides code by default — pass --include-code to show it.

Related

Data wrangling with [[pandas]] / [[polars]]; plotting with [[matplotlib]]; SQL via [[duckdb-docs]]; environments and PEP 723 scripts in [[modern-python]].