marimo

Most notebooks lie about their state. marimo doesn't. That's the whole pitch, and it holds up. If you've ever spent an hour hunting down a stale variable in Jupyter, the reactive execution model here is a genuine relief — change a cell and every dependent cell updates automatically, with no hidden state. Delete a cell and its variables disappear. That determinism is what makes marimo's Git-friendly story work; notebooks are pure .py files you can diff, review, and version like any other code. We'd reach for this when reproducibility and collaboration matter more than the last bit of notebook flexibility. The recent PyCharm plugin and notebooks-as-slides show marimo is moving beyond a Jupyter alternative into a full development environment, and 'marimo pair' — letting AI agents work inside the kernel instead of through MCP wrappers — is a forward-looking take on agentic coding. But it's not for everyone. If you're married to Jupyter's extension ecosystem, you'll find marimo's plugin universe smaller. Real-time multi-user collaboration like Google Colab? Not here, yet. And if you want a no-code drag-and-drop tool, marimo is still code-first — it won't replace Retool. The closest alternative is Streamlit, but that's a different beast: Streamlit is a framework for building apps, while marimo is a notebook that can be an app. That means marimo's learning curve is gentler if you already think in notebooks, but you lose some of Streamlit's app-specific features like theming and session management. In practice, teams that adopt marimo often replace both Jupyter and a separate app tool, because one artifact can serve as a notebook, a script, and a shareable web app. Watch out for the GPU upgrade on molab — it's paid, but the free tier covers core notebook use. At 22.5k GitHub