The reactive Python notebook that stopped pretending a JSON scratchpad was a program - and rebuilt the whole thing as real code.
Here is a thing that is true about the notebook every data scientist opens each morning: it lies. Not maliciously. It just quietly forgets. You run cell three, then cell one, then delete cell two, and the notebook keeps humming along as if the variable you removed still exists, because it does - somewhere, in hidden memory, in a state no one can see and no git diff can catch. This is the reason a striking number of published results cannot be reproduced by the people who published them. marimo, a San Francisco company founded in 2022, looked at this arrangement and reached an unusual conclusion: the problem is not the user. The problem is the tool.
marimo's fix is almost aggressively simple to describe and quite hard to build. It treats a notebook as what it secretly always was - a dataflow graph. When you define a variable in one cell and use it in another, marimo notices, and it draws an invisible edge between them. Change the first cell and the second re-runs automatically. Delete a cell and its variables vanish from memory, everywhere, at once. There is no "Run All," no stale output, no cell that works only if you run it in the exact order you happened to run it last Tuesday. The notebook is, in the strict sense, reproducible: run it top to bottom and you get the same thing every time, because that is the only way it can run.
The second trick is where it gets interesting for anyone who has tried to put a notebook into production. A marimo notebook is not stored as JSON. It is stored as a pure Python file - a real .py you can read, diff in a pull request, and version with git like any other code. Which means the same file is also a runnable script. And, with a flag, a deployable web app, complete with sliders and dropdowns and SQL cells. And, thanks to WebAssembly, it can run entirely inside a browser tab with no Python installed at all. One file, several lives. That is the pitch, and the reason people describe first encountering it, per the company's own users, as "like magic becoming real."
"Python developers deserved dramatically better tools for working with data."
- Akshay Agrawal, Co-Founder & CEOChange Cell A and marimo re-runs only B and C - the cells that actually depend on it. Nothing hidden, nothing stale, nothing run out of order.
ONE FILE, FOUR LIVES
An open-source reactive notebook stored as a pure Python file. The reactive engine tracks dependencies between cells and re-runs only what changed - killing hidden state and making notebooks reproducible and git-friendly.
The same file runs as a standalone Python script or deploys as an interactive web app with sliders, dropdowns, and SQL cells. No rewrite, no separate framework.
Powered by Pyodide, marimo runs entirely in the browser - no Python install, no backend server, no network round-trips to a remote runner.
A free cloud-hosted marimo environment - 4 CPUs and 32GB RAM by default, with an optional NVIDIA RTX Pro 6000 Blackwell GPU (96GB VRAM), now backed by CoreWeave infrastructure.
Former Google Brain engineer who worked on TensorFlow, with a PhD from Stanford and published research at NeurIPS. The idea for marimo grew from his own frustration, as a PhD student, that Jupyter couldn't show how a notebook's cells related to one another.
Former engineer at Palantir and CloudKitchens. Co-founded marimo in 2022 and leads the engineering behind the reactive runtime, the WebAssembly build, and the molab cloud environment.
"A reinvention of the Python notebook as a reproducible, interactive, and shareable Python program, instead of an error-prone JSON scratchpad."
- How marimo describes its own productThe roughly eight-person team joined the AI cloud company to scale molab with GPU compute - while keeping the notebook, in the company's words, "free, open-source, and permissively licensed."
Led by AIX Ventures (Anthony Goldbloom and Shyam Mani).
Angels: Jeff Dean (Google), Clement Delangue (Hugging Face), Wes McKinney (pandas / Posit), Charlie Marsh (Astral), Jordan Tigani (MotherDuck), Paige Bailey, Lukas Biewald, Ben Hamner, Shane Barratt.
Akshay Agrawal and Myles Scolnick begin building; early development supported by Stanford's SLAC National Accelerator Laboratory.
You can learn a lot about a company from who agrees to fund it, and marimo's seed round reads less like a cap table than a who's-who of the exact people who built the tools it is trying to improve on. Wes McKinney created pandas, the library that made Python the language of data. Jeff Dean runs much of Google's AI. Clement Delangue runs Hugging Face. Charlie Marsh built the Rust-based Python tooling that the ecosystem now runs on. These are not generalist growth investors chasing a category; they are practitioners who have spent careers inside the notebook and, evidently, wanted this particular thing to exist.
The other tell is the CoreWeave arrangement. In 2025 marimo joined CoreWeave, a company whose entire business is renting out GPUs, and the natural cynical read is that the open-source project becomes a funnel for cloud compute. But the terms are unusually explicit: the notebook stays free and permissively licensed, the roadmap stays open, the team keeps shipping. The commercial engine is molab and the compute behind it. It is a bet that you can give away the tool and still build a business - which is either naive or exactly how the most durable developer tools have always worked. marimo is wagering on the latter.
"marimo is, and always will be, free, open-source, and permissively licensed."
marimo, on joining CoreWeave"The tools we use shape the way we think - better tools, for better minds."
marimo, company valuesUsers describe first using it as "finally, someone is doing it" and "like magic becoming real."
From the seed announcementThe default for a decade. Powerful and ubiquitous - and the source of the hidden-state and reproducibility problems marimo set out to fix.
Collaborative, cloud-first notebooks aimed at teams. marimo's angle is different: open-source, local-first, and stored as plain code.
Reactive notebooks that inspired the model - one for JavaScript, one for Julia. marimo brings the reactive idea, done natively, to Python.