The best AI agent for Jupyter notebooks

Jupyter predates coding agents by a decade. The .ipynb format was built to store the outputs of cells together with the inputs. That’s cool if you want to render a notebook, but terrible for agents that read, edit, run, and verify code. So maybe if you’re looking for a better agent experience in Jupyter, you’re actually looking for a better agentic notebook.

That notebook is marimo: an open-source, AI-native Python notebook stored as pure Python, not JSON, with reactive, reproducible execution. AI-features were designed into the notebook from the outset, not bolted on with a plugin.

Why Jupyter fights AI agents

Several properties of Jupyter notebooks make them hard for agents to work in:

• The file format isn’t plaintext. Jupyter notebooks are stored as JSON with base64-encoded outputs. LLMs do their best work generating and editing code, not surgically patching JSON blobs interleaved with binary outputs. Edits are fragile and diffs are noisy.
• Hidden state breaks reproducibility. Cells can run in any order, mutating a shared workspace. This is why over 75% of Jupyter notebooks on GitHub don’t run top-to-bottom, and 96% don’t reproduce. Hidden state trips up humans and it trips up agents even harder.
• Agents can’t easily verify their work. Good agentic loops depend on running code and inspecting results. A .ipynb is not a well-formed program an agent can simply execute and trust.

Why marimo is the best notebook for AI agents

marimo was designed from the ground up for the way agents work:

1. Stored as pure Python. A marimo notebook is a regular .py file, so agents read and edit it the same way they edit any source file. No JSON wrangling just clean git diffs.
2. Self-checking. marimo ships with marimo check, a built-in linter that statically validates notebooks. This gives agents a fast, deterministic signal to catch and fix their own mistakes.
3. Runnable and introspectable. You can start a marimo notebook with a --watch flag which will update the browser view whenever changes are made. Any changes that an agent makes will be visible to your immediately. Keeping you in the loop.
4. Agent scratchpads. marimo offers skills to agents that give access to a virtual scratchpad so that agents gain access to everything the notebook has access to.

Drop your agent inside the notebook with marimo pair

marimo pair is an agent skill that drops your coding agent of choice inside a running marimo notebook session, giving it access to everything you see. The agent can run code in an ephemeral scratchpad, check data schemas, add and delete cells, install packages, and inspect the notebook’s in-memory variables and state. The notebook becomes a shared canvas for you and the agent — and a reactive REPL that eliminates hidden state and guarantees a reproducible Python program.

Because it’s implemented as a skill, marimo pair works with any agentic harness that can use skills, which includes Claude Code, Codex, and OpenCode.

Prefer to drive an agent from inside the editor? marimo’s built-in Agent sidebar connects ACP-compatible agents directly to your notebook so they can write and modify cells as you steer.

Need real compute? marimo pair can connect to molab, marimo’s free cloud notebooks, which run in sandboxed environments with free GPUs. Your agent gets a clean, reproducible workspace with the horsepower to finetune models or crunch large datasets.

A linter built for agent feedback loops

marimo ships with marimo check, a built-in linter that statically validates notebooks. It catches the kinds of marimo-specific mistakes agents introduce before they become runtime failures. This gives an agent a fast, deterministic feedback signal it can act on and self-correct against, something a .ipynb file does not provide.

How to use marimo

marimo is entirely free and open source. Just pip install marimo or uv add marimo to get started.

marimo also offers a free, cloud-hosted notebook; molab. Check out our gallery for inspiration and examples from our community.