marimo explained: reactive Python notebooks stored as testable scripts

marimo is a reactive Python notebook stored as a normal `.py` file, with cells connected through variable definitions and references. At capture time it had 21,925 stars, about 171 default-branch commits in 30 days, and release 0.23.14. Its important claim is not a different canvas; it addresses hidden state, stale out-of-order results, and opaque JSON diffs in conventional notebooks.

The project supports data exploration, SQL analysis, model experiments, teaching material, interactive reports, and internal tools. Built-in UI elements pass values from sliders, tables, and forms into Python. SQL cells return data frames. One file can be edited as a notebook, executed as a script, or served as a read-only application. AI assistance is optional and is not required for the reactive runtime.

marimo statically examines names defined and read by each cell, then builds a directed acyclic dependency graph. A changed variable causes only affected downstream cells to run in topological order; deleting a cell removes its definitions. The system rejects multiple cells that define the same variable, reducing ambiguity. Lazy execution can protect expensive work, while in-place mutation of objects may remain invisible to name-based analysis.

The backend and notebook programs are primarily Python, while the browser interface uses TypeScript and React. A notebook is an importable, testable Python program with readable Git diffs. App mode sends interface values to backend computation. A WebAssembly build uses Pyodide to execute Python in the browser, which helps serverless demonstrations but constrains available packages, browser resources, networking, and durable storage.