Quickstart

Two ways in: use the hosted demo (nothing to install), or add the extension to your own JupyterLab or JupyterLite.

Try the hosted demo

1. Open the live demo in ChatGPT's in-app browser, or in Google Chrome with WebMCP enabled.
2. Wait for the status bar (bottom right) to read WebMCP ready. If it says WebMCP unavailable, this browser doesn't expose document.modelContext: the notebook still works, but there are no agent tools.
3. Double-click customer-analysis.ipynb in the file browser and wait for the kernel to go idle.
4. Give your agent the prompt below and watch the notebook.

Open customer-analysis.ipynb. The conversion rate looks wrong to me:
read the conversion-rate cell and check its denominator against how
eligible_sessions is defined further up. If it's wrong, fix only that
expression, rerun the cell, and leave a review comment on the line you
changed explaining why.

You should see the targeted cell ring and report Applying…, converted / visitors become converted / eligible_sessions, a ±2 changed button appear (click it for the diff), the execution count increment, and a review thread show up in the Agent panel.

Then try the parts that aren't about prompting

• Highlight eligible_sessions and ask the agent what you selected. It reads the exact substring.
• Right-click a cell → Agent Access → Hidden, then ask the agent to read it. It will behave as if the cell doesn't exist, while you can still edit it.
• Edit a cell without saving, then ask the agent to rewrite it. The write is refused and your text is untouched.
• Switch the Agent panel to Propose mode and ask for another edit. It waits for your Accept or Deny. See Propose mode.

Other seeded notebooks: needs-review.ipynb (deliberate problems, good for a review task) and reviewed-analysis.ipynb (a finished human-and-agent session with review threads). scratch.ipynb is nearly empty, for experiments.

Add it to your own notebooks

pip install jupyterlite-webmcp
jupyter labextension list   # should show jupyterlite-webmcp enabled OK
jupyter lab

For a JupyterLite site, add jupyterlite-webmcp to your requirements.txt next to jupyterlite-core and the Pyodide kernel, then run jupyter lite build. The install guide covers each platform and how to verify it.

Check the tools by hand

In a WebMCP-capable browser, open the devtools console:

const tools = await document.modelContext.getTools();
tools.map(t => t.name);                                        // 22 names
const ctx = tools.find(t => t.name === 'jupyter_get_context');
JSON.parse(await document.modelContext.executeTool(ctx, '{}'));

Chrome passes arguments and results as JSON strings; each result parses to a { content, structuredContent, isError } envelope. The tool reference lists inputs, outputs, bounds and error codes.