Jupyter notebook viewer: read an .ipynb without Jupyter

What it shows

• A summary first: the kernel and language, the nbformat version, how many code, Markdown and raw cells, how many cells have outputs, and the five largest outputs by size, each linked to its cell.
• Whether the cells were run out of order. If In [7] sits above In [6], the outputs you see came from a different order than top to bottom, and re-running the notebook may not reproduce them. The page says where, and which execution numbers are missing.
• Markdown cells rendered: headings, lists, tables, links, code blocks and pictures attached to the notebook. Maths written as $…$, $$…$$ or \begin{equation} is typeset by KaTeX, an MIT-licensed typesetter served from this site.
• Code cells with syntax highlighting for Python, R, Julia, JavaScript and SQL kernels, and the execution count Jupyter showed beside each.
• The saved outputs: printed text, results, PNG, JPEG and SVG charts (matplotlib, seaborn), pandas DataFrames as tables, LaTeX results, and error tracebacks in their original terminal colours.
• Three saves: the whole notebook as one HTML file that opens in any browser; the code as a .py (or .r, .jl) with a # %% marker before each cell and the Markdown as comments; and every chart as a PNG.

What this cannot do

• It does not run code and cannot re-run cells. What you see is what was saved; a notebook saved with “Clear all outputs” shows code and Markdown only.
• It does not show live widgets or JavaScript charts. ipywidgets, Plotly, Bokeh and Altair outputs need JavaScript and a kernel. Each one is named in a sentence where it would be, and a static picture saved beside it, if any, is shown.
• It does not edit. Nothing is saved back into the .ipynb.
• HTML written inside Markdown cells appears as text, and pictures linked from the web are not loaded but named, so opening a notebook never contacts anyone.

Useful to know about .ipynb files

What is inside an .ipynb

Plain JSON. A cells list holds every cell’s source as text; each code cell keeps its execution_count and an outputs list, where a chart is a base64 PNG (about a third larger than the picture) and a pandas table is both text/html and text/plain. metadata.kernelspec names the kernel, and nbformat 4 has been the layout since IPython 3 (2015). The older nbformat 3 put cells inside worksheets and had separate heading cells; this page reads both.

A worked example

The BSD-licensed test notebook test4.ipynb from the nbformat project reads as: nbformat 4.0, no kernel recorded, 9 cells (4 code, 5 Markdown), 4 with outputs. Its code cells ran as In [1], In [3], In [7], In [6], so it is flagged as run out of order, with numbers 2, 4 and 5 missing. Its largest output is a 12.2 KB PNG (a 520 × 67 logo); a %%javascript cell’s output is named and not run, and the <script> inside its HTML output is removed.

Why a notebook is huge

Almost always the outputs, not the code. Every inline chart is stored as base64 text, a printed DataFrame keeps both its HTML and its text form, and some libraries embed their whole JavaScript bundle (several MB) in one output. The “Largest outputs” table shows where the bytes are; clearing those outputs in Jupyter before sharing or committing to git shrinks the file.

The .py download

Each code cell follows a # %% In [n] line, the “percent” cell format that VS Code, Spyder and Jupytext understand, so the file still runs cell by cell. Lines starting with % (IPython magics such as %matplotlib inline) or ! (shell commands) are commented out, because plain Python cannot run them; a whole %%bash or %%time cell is commented out too.

Without uploading it

nbviewer shows only notebooks at a public web address, and GitHub only notebooks pushed to a repository. Here a private notebook, with its data in the outputs, stays on your computer. The HTML outputs are cleaned and drawn in a sandboxed frame that is not allowed to run scripts or load anything; the page checks after drawing that no notebook code ran.