Author with a coding agent

A coding agent can inspect a notebook, create a view, edit its source, and verify the result in a browser. Studio ships the instructions and Python API with the package, so the agent works from your installed version.

Ask for a view

Paste one 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.

Agent Plugins prints the instructions a Python package ships for coding agents, and uv supplies uvx. Studio's briefing tells the agent to start a notebook with Studio, write its cells, then create, build, and show each view in Preview beside the notebook.

Name an existing notebook to build on its analysis:

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

The agent runs Python in the notebook kernel through marimo pair, a marimo command for working in a live notebook session. It finds local notebooks started with --no-token and starts a stopped notebook that way. For a notebook you started with a token, give the agent its URL and a file that holds the token. The agent then inspects the notebook and available starters, reads each view's AGENTS.md, and edits the view source.

In marimo's AI sidebar, the agent already runs in the kernel. Switch to Code Mode (beta) and ask for the view directly.

Write a good request

Give the agent an audience, a task, and a result 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. Verify the view at desktop and phone widths and check that changing a control updates its dependent results.

Review the result in Preview. Keep shared calculations in the notebook and audience-specific layout and wording in the view. Record lasting visual decisions in the project's DESIGN.md.

Inspect and verify

For agent authors and integrations, the live API begins with the current workspace. From the notebook's Python environment, help(marimo_studio.agent) prints the briefing with the Python API. Run this in one code-mode execution:

import marimo_studio

workspace = marimo_studio.agent.current_workspace()
notebook = await workspace.inspect_notebook()
for cell in notebook.cells:
    print(cell.name, cell.definitions, cell.has_output_expression)
for starter in await workspace.starters():
    print(starter.id, starter.title)

inspect_notebook() reads saved source. Its optional runtime inspection runs a separate process. Live kernel values, the published build, and the displayed browser result are separate evidence.

After authoring and building briefing, show it in a new code-mode execution:

import marimo_studio

view = marimo_studio.agent.current_workspace().view("briefing")
await view.show()
print(await view.preview_url(runtime="server"))

Reacquire workspace and view handles in each execution. Finish the execution before an external browser tool waits for notebook results. After changing Python, run the changed cells in the live notebook.

Check source and build freshness with view.inspect(). A failed build retains the last working artifact, so a visible page alone does not prove the edit was published. In the browser, wait for html[data-marimo-studio-state="ready"], inspect wide and narrow layouts, and exercise controls through to their dependent results. Verify exports in the runtime visitors will use.

The Python API defines methods and records. The installed skill supplies the complete authoring workflow and references for source edits, projections, verification, and delivery:

import marimo_studio

print(marimo_studio.agent.skill().file("references/verification.md").read_text())

Select results with Lens

Lens lets you select a rendered result and attach a note. The agent receives the image and the notebook context behind that selection. Install Lens in the notebook's Python environment:

uv pip install "marimo-studio[lens]"

The lens extra installs a Lens release compatible with the installed Studio. For sandboxed notebooks, also declare marimo-studio[lens] in the script's dependencies. Restart a running notebook after installing or upgrading Lens.

When the notebook imports no Lens of its own, marimo mounts one in the notebook. The development preview reuses that same Lens, so the Notebook pane and the preview each show a dock, and selections from either one reach the agent together. Select a result, add a note, and ask:

Use Lens to address my current selection in this Studio view.

Native projections carry their producing context automatically. For custom charts or authored page regions, connect the rendered result to its inputs. Lens's agent guide owns the selection and feedback workflow.