Jupyter deploy

Jupyter Notebook Deploy

Create, execute, and deploy Jupyter notebooks as interactive HTML pages on Vercel.

Step 1: Understand the Request

The user will describe what they want analyzed, visualized, or computed. Clarify if needed, but generally proceed directly with creating the notebook.

Step 2: Create the Notebook

Use the NotebookEdit tool to create a .ipynb file in the working directory.

File naming: Use a descriptive slug based on the request, e.g., sales-analysis.ipynb, prime-numbers.ipynb. If the user specifies a name, use that.

Notebook structure guidelines:

• First cell: Markdown title and description of the analysis
• Import cells: Group all imports in one cell near the top
• Logic cells: Break the work into logical steps, each in its own cell
• Visualization cells: Use matplotlib/pandas plotting; always call plt.show() or use inline display
• Matplotlib setup: Always include this in your setup cell:

  import matplotlib
  matplotlib.rcParams['font.family'] = 'Noto Sans CJK JP'  # Japanese font support
  

  Use the default matplotlib style (white background, black text). Do NOT use dark themes — they are hard to read in notebook HTML output.
• Summary cell: Final markdown cell summarizing findings
• Only use libraries available in the container: numpy, pandas, matplotlib (standard library modules are also available)
• If the user needs a library that is not installed, tell them it is unavailable and suggest an alternative using the installed libraries

Important: Set the kernel to python3 in the notebook metadata.

Step 3: Execute the Notebook

Run the notebook to generate outputs (charts, tables, printed values):

jupyter nbconvert --to notebook --execute --inplace <notebook-file>.ipynb

This executes the notebook and writes outputs back in-place. Do NOT use jupyter execute as it may not persist outputs to the file.

If execution fails:

1. Read the error from the output
2. Fix the notebook using NotebookEdit
3. Re-execute
4. Retry up to 2 times. If it still fails, report the error to the user and stop.

Step 4: Convert to HTML

jupyter nbconvert --to html --template classic <notebook-file>.ipynb

IMPORTANT: Always use --template classic. The default lab template uses JupyterLab CSS variables (--jp-*) that are undefined outside JupyterLab, causing the page to render with no styling. The classic template produces self-contained, standalone HTML that looks correct in any browser.

By default, code cells are shown (notebook style with In [n]: prompts). If the user explicitly asks to hide the code, add --no-input to produce a clean report-style output.

The output file will be <notebook-file>.html in the same directory.

Step 5: Read the HTML

Read the generated HTML file using the Read tool.

If the HTML is too large (over 500KB), the Vercel deploy may fail. In that case:

1. Try --no-input if not already used
2. Reduce the number of high-resolution plots
3. Report to the user if it cannot be reduced

Step 6: Deploy to Vercel

Determine the project name:

• If the user specified a name, use it (lowercase, hyphens only)
• Otherwise, generate from the notebook filename: e.g., sales-analysis.ipynb becomes jupyter-sales-analysis
• Always prefix with jupyter- to avoid collisions with other projects

Deploy using the MCP tool:

mcp__vercel__vercel_deploy({
  name: "<project-name>",
  files: [
    { file: "index.html", data: "<the full HTML content>" }
  ]
})

Step 7: Return Result

Respond to the user with:

1. The deployed URL
2. A brief summary of what the notebook contains
3. Key outputs or findings (if applicable)

Example:

Notebook deployed!

URL: https://jupyter-sales-analysis-xxx.vercel.app

Created a sales analysis notebook with:
- Monthly revenue trends (line chart)
- Top 10 products by sales (bar chart)
- Summary statistics table

The notebook used sample data with 1,000 transactions.

Error Handling

• Notebook creation fails: Report the NotebookEdit error to the user
• Execution fails after retries: Share the error message and the notebook file (user can debug)
• nbconvert fails: Try without --no-input; if still failing, report error
• Vercel deploy fails: Check if HTML is too large; report error with details
• Missing library: Do NOT try to pip install (will fail as non-root). Tell the user which libraries are available and suggest alternatives