Author with a coding agent

A coding agent can turn a request such as "a slide deck and a scrollytelling report explaining calculus" into a notebook and its views. Studio ships the instructions the agent follows inside the marimo-studio package, and gives the agent a Python API in the live notebook to create, build, and show views. You review each result in Preview and point at what should change with Lens.

How an agent works with Studio

1. It reads Studio's instructions.Agent Plugins prints the instructions that ship inside the marimo-studio package.
2. It connects to the live notebook. A terminal agent pairs with the running notebook through marimo pair. An agent in marimo's AI sidebar already runs there in Code Mode (beta). Either way, the agent runs Python in the notebook's sessionSessionA running copy of the notebook. marimo runs the notebook's cells in a session and keeps their results, and a page reads those results by name..
3. It writes Python in the notebook and presentation in the view. Data and computation go into notebook cells. Layout, wording, and page code go into the view projectView projectThe folder that holds one view's source, such as studio/analysis/dashboard/. It is an ordinary project in the view's format, such as an HTML page, a React app, or a Typst document, so any editor, tool, or coding agent can change it., following its AGENTS.md, and its DESIGN.md when the view has one.
4. It builds and shows the view. The agent builds the view, checks the names it places, and selects the view in your Studio tab, so Preview shows what it made.
5. It takes your feedback. You review the view in Preview and select the parts that should change with Lens. When a selection shows a notebook result, the agent receives it with the notebook cells behind it.

Ask for a view

Paste a request into a terminal agent such as Claude Code or Codex:

Run `uvx --with marimo-studio agent-plugins read marimo-studio` and create
a scrollytelling report and a slide deck explaining calculus basics.

The first sentence makes the agent read the instructions, and uv supplies uvx. The agent then starts a notebook with Studio, writes the analysis in Python cells, creates and builds each view, and shows it in Preview beside the notebook.

Name a notebook you already have to build on its analysis:

Run `uvx --with marimo-studio agent-plugins read marimo-studio` and build
a Typst report of analysis.py that leads with the headline results.

marimo pair finds local notebooks started with --no-token, and the agent starts a stopped notebook that way. For a notebook started with a token, give the agent its URL and a file that holds the token. In marimo's AI sidebar, switch to Code Mode (beta) and ask for the view directly.

Write a good request

Name the audience, the task, and something to check:

Create a view named briefing for a quarterly review. Use the notebook's existing measures and controls. Lead with the headline results, then show the evidence behind them. Check the view at desktop and phone widths, and check that changing a control updates its results.

Review the result in Preview. Python belongs in the notebook, and layout, wording, and page code belong in the view, as Python in the notebook, presentation in the view describes. Ask the agent to record lasting decisions, such as audience and visual direction, in the view's DESIGN.md.

The agent reads each view's AGENTS.md before it edits the view. The Python API lists the methods it uses to inspect the notebook, build views, and show them.

Point at what to change

Words like "the dip in that chart" leave the agent guessing which cell and which value you mean. Select results with Lens lets you select the dip itself, in a page or inside a PDF, and add a note. The agent receives the selection with the cell that computed it and the place in the view that shows it, so it changes the notebook when the data is wrong and the view when the presentation is.