Does my notebook leave my browser?

The notebook is read in your browser before anything is sent, and stored outputs never leave it. What the metered lanes receive is a cell-structured digest: the code of each cell plus a one-line description of what that cell produced, rather than its base64 images. Credential-shaped literals are found and masked in the browser first, and the page tells you how many were masked and how many characters were clipped before you run.

What are the four lanes?

Reproducibility audit asks whether the notebook runs top to bottom on a clean kernel. Port to marimo turns it into a marimo notebook that loads, resolving the top-level redefinitions marimo refuses. WASM check says whether it can run as a WebAssembly notebook in a browser and what has to be substituted first. Scheduled batch run turns it into a parameterised, unattended job with a runner command and failure handling. All four run over the same document.

Can I see the output without signing in or spending a credit?

Yes. Two example notebooks are bundled - a tangled research notebook and a tidy pipeline notebook - and both ship with saved model runs for every lane, so all four outputs including the generated marimo notebook are reachable without signing in and without spending a credit.

What does the free prescan find before I spend anything?

It reads every cell with a real Python lexer that strips comments and every string form before counting an identifier, then reports each cell's top-level bindings, names used in one cell but defined only in a later one, execution counts that run backwards, and every top-level import mapped against Pyodide availability. It also downloads the reordered notebook as a valid .ipynb, with cells sorted into dependency order, outputs cleared and every execution count reset.

What happens to a prescan flag the model does not answer?

Every flag sent with the run appears in the prescan reconciliation table with what the run did with it. A flag the reply never mentioned is named there explicitly and described as unreviewed rather than cleared, so a silently dropped flag cannot pass as a resolved one.

Which skills is NB Shift built on?

NB Shift is a derived work built on four published SkillSafe skills, credited in full on the page: @openai/jupyter-notebook for the reproducibility audit lane, @marimo-team/jupyter-to-marimo for the port lane, @marimo-team/wasm-compatibility for the WASM lane, and @marimo-team/marimo-batch for the scheduled batch lane. It is not affiliated with, endorsed by, or operated by OpenAI, Project Jupyter or marimo.

Your notebook, read cell by cell — then audited, ported, or scheduled

Drop the .ipynb. Before anything is sent, this page parses it in your browser: every cell's top-level bindings, every name used before it is defined, every execution count that runs backwards, every import that cannot load under WebAssembly. Then pick a lane and the model has to answer for each of those findings.

Both examples ship with saved model runs for every lane, so you can see all four outputs — including the generated marimo notebook — without signing in and without spending a credit.

Your notebook nothing pasted yet

Drag a notebook in, or Everything is read in your browser. Stored outputs never leave it.

Depth

Written for

Target Python

Anything the notebook does not say (optional)

Paste a notebook to price the run.

Verdict—

Findings—

Blockers—

Cells read—

Document—

Artifact

The model replied in an unexpected shape

Nothing is hidden from you — here is the raw reply.

Your notebook, read cell by cell — then audited, ported, or scheduled

Checks

Findings

Artifact

Do this, in order

Prescan reconciliation

Assumptions and open questions

The model replied in an unexpected shape

Your runs

What this run actually did