We’re Akshay Agrawal and Myles Scolnick, and we’re the developers of marimo, an open-source reactive notebook for Python. In this post, we’ll describe how we ported marimo to work via Pyodide, and why we’re incredibly excited about the potential of Pyodide to make computing more accessible.

What is marimo?

marimo is an open-source reactive notebook for Python built from the ground up to solve well-known problems with traditional notebooks, including reproducibility, interactivity, maintainability, and shareability.

A reactive notebook with no hidden state. In marimo, a notebook’s code, outputs, and program state are always consistent. Run a cell and marimo reacts by automatically running the cells that reference its declared variables. Delete a cell and marimo scrubs its variables from program memory, eliminating hidden state. Try this for yourself!

Our reactive runtime is based on static analysis, so it’s performant. If you’re worried about accidentally triggering expensive computations, you can disable specific cells from auto-running.

UI elements are automatically synchronized with Python. marimo comes with UI elements like sliders, a dataframe transformer, and interactive plots. Interact with an element and the cells that use it are automatically re-run with its latest value. Reactivity makes these UI elements more useful and ergonomic than Jupyter’s IPyWidgets.

Deploy as an app, run as a script. Every marimo notebook can be served as an interactive read-only web app (with outputs concatenated and code hidden). Additionally, marimo notebooks can be run as scripts from the command line, with cells executed in a topologically sorted order. For example, here’s a simple educational app:

Use cases. marimo can be used in many ways. It can be used as a next-gen replacement for Jupyter notebooks, allowing you to explore data rapidly — change code or interact with a UI element, and see outputs change instantly — and run experiments in an environment that prevents bugs before they happen. It can also be used to make powerful apps, dashboards, and internal tools.

A Pyodide-powered playground. All the embedded marimo notebooks you’ve seen in this blog post are powered by our Pyodide-marimo playground: you can try marimo in your browser by navigating to https://marimo.new.

If you want to use marimo locally, get started with

pip install marimo && marimo tutorial intro

at your command line.

Using Pyodide to make marimo accessible to everyone

We decided to port marimo to Pyodide for three main reasons:

  • to make scientific computing accessible to everyone by eliminating the need to install and manage Python;
  • to make it easy to share executable notebooks via a url by eliminating the financial and technical burden of deploying backend infrastructure; and
  • to provide a snappy development experience by eliminating network requests to a remote Python runner.

Our motivation for combining marimo with Pyodide echoes the vision articulated by the developers of Iodide, the experimental in-browser notebook for which Pyodide was originally created. While Iodide is no longer developed, we’re extremely grateful that it ushered Pyodide into existence.

The Pyodide-marimo playground. We have developed a Pyodide-powered marimo playground that makes it extremely easy for anyone in the world to tinker with marimo notebooks. With just one click, playground notebooks can be shared with others, and conversely every local notebook can be shared via our playground.

Here are some use cases that our Pyodide-powered notebooks enable:

  • Rapidly experiment with code and models.
  • Author blog posts, tutorials, and educational materials.
  • Embed interactive notebooks in library documentation or other static web pages.
  • Build and share tools like data labelers, embedding visualizers, model comparison tools, or anything else you can imagine.
  • Create and share bug reproductions.

You can learn more about the playground, including how to embed marimo notebooks in static HTML pages, at our docs.

We’ve created a few example notebooks to jump-start your imagination. There’s one that trains a tiny neural network, another that visualizes Bayes’ Theorem, and another that plots attractors of dynamical systems. One of our community members even created a QR code generator.

When you take a moment to think about all this, it’s really quite magical ✨. Thanks to Pyodide, you can use any pure Python package, as well as a number of packages for scientific computing, including Pandas, NumPy, SciPy, matplotlib, and scikit-learn, entirely in the sandbox of your web browser — without installing Python on your machine and without paying for infrastructure.

Implementation: Porting marimo to Pyodide

Porting marimo to work with Pyodide was relatively straightforward, a testament to the latter’s robustness and thorough documentation.

marimo’s original implementation has three main parts:

  1. a Python codebase implementing a kernel that runs Python code;
  2. a TypeScript codebase that issues control commands for the Python kernel;
  3. a server that connects the TypeScript and Python codebases.

All that was required were minimal changes to our TypeScript and Python codebases, and replacing the server with a lightweight bridge.

TypeScript

marimo uses the same TypeScript build when running with Pyodide or native Python. This helps maintain consistency between the two experiences and avoid feature drift. The code path splits based on feature flags; when running under Pyodide, we download the marimo wheel from PyPI and initialize a single-threaded, Pyodide-compatible marimo kernel. Passing only a filename and a message callback function, we initialize the kernel as an async, never-ending Python process.

We interact with this kernel through a lightweight RPC bridge that sends requests to the kernel and receives responses through the callback passed. This allows for an asynchronous-like feel to the execution. To support interruption of Python code, we simply followed the Pyodide documentation on using its SharedArrayBuffer-based mechanism.

For performance, we cache our assets and run Pyodide in a web worker to avoid blocking the main browser thread. This creates some extra complexity with type-safety and mimicking blocking RPCs.

We heavily leverage the Emscripten filesystem. Code is passed from the URL hash (for ease of sharing), to the main frame, to the worker, and finally to the Emscripten filesystem. Emscripten supports many filesystem implementations such as IndexDB which we use to persist the user’s files. We are working on our own implementation of the Emscripten filesystem built on top of any S3-compatible bucket. This means you’ll be able to list, read, and write files from your S3 bucket just by interacting with ”os” filesystem all in Python.

In order to make the experience smooth in any browser environment, we do our best to auto-install packages. Whenever a cell is run, we try to install it through Pyodide or otherwise fallback to micropip.

Python

Similar to our TypeScript codebase, our Python codebase uses the same implementation for Pyodide and native Python. We introduced just a few branches based on whether the kernel is running under Pyodide, since some Python features, such as shared memory, threading, and multiprocessing, are not yet available in Pyodide. As one example, we replaced multiprocessing queues with asyncio queues.

We added a new entrypoint for creating the marimo kernel when running under Pyodide, which receives messages from the RPC bridge in an asynchronous control loop. This entrypoint also sets up stream objects (e.g., for standard output) to use the message callback provided to it by the frontend.

In order to support package installation with micropip, we had to modify the kernel implementation to support top-level await. Typically, it is a syntax error to use await outside a function, but micropip.install is an async function and must be awaited. So we configured code compilation to allow top-level await, and wrapped the marimo kernel in an asyncio event loop. As a bonus, this change made it possible to use top-level await when running via native Python, a feature that had been requested by our users for some time.

What’s next?

We believe that marimo used with Pyodide holds enormous potential for making computing more accessible:

  • Anyone can create and share live notebooks, without having to pay for infrastructure.
  • Students can get started with Python instantly, without having to set up a development environment.
  • Library developers can embed reactive coding environments and demos in documentation — we’ve done this throughout our own API docs.
  • Authors can publish interactive blog posts and computational tools as part of static web pages.

We have ideas on how to make Pyodide-powered marimo notebooks more useful — e.g., we’d like to make it easier to work with auxiliary code and data files, and to save multiple notebooks at our playground. We’re also interested in helping the Pyodide team support more packages.

Most of all, though, we’d like to learn what you want: come chat with us in our Discord or Github, or send us an email at contact@marimo.io.