pynote

PyPI distribution: jupyter-pynote · Python import: pynote · Lab extension id: pynote

Claude-powered chat side panel for JupyterLab 4 and Jupyter Notebook 7.

Open any notebook, crack open the side panel, and talk to Claude about the code. When Claude proposes changes to the notebook, type apply in chat and they're written into real cells — no dialog box, no reload, no token-pasting.

Features

• Right-hand side panel docked into the Lab shell; follows the currently active notebook.
• Chat with Claude about the current notebook's cells (their source is sent as context each turn).
• Structured cell edits via Claude tool-use — replace existing cells or insert new ones by stable cell id, not brittle index.
• apply / cancel are typed in chat; no modal or popup.
• Edits go through Lab's own shared model, so undo / autosave / RTC all Just Work.
• Server extension proxies the Claude call, so ANTHROPIC_API_KEY lives in the Jupyter process — never in the browser.

Prerequisites

• Python 3.9+
• JupyterLab 4 or Notebook 7 (pip install jupyterlab)
• Node 18+ and jlpm (ships with JupyterLab)
• An Anthropic API key

Install alongside your existing Jupyter

From jupyter-pynote/:

# 1. Install the Python side (server extension) in editable mode
pip install -e ".[dev]"

# 2. Install npm deps and build the labextension assets
jlpm install
jlpm build

# 3. Wire the labextension into JupyterLab
jupyter labextension develop . --overwrite

# 4. Enable the Jupyter server extension
jupyter server extension enable pynote

# 5. Provide your API key (or put it in a .env you source before launching)
export ANTHROPIC_API_KEY=sk-ant-...

# 6. Launch Jupyter
jupyter lab

Verify the extension is loaded:

jupyter labextension list      # should show: pynote v0.1.0 enabled ok
jupyter server extension list  # should show: pynote enabled

Open any .ipynb. You'll see a pynote tab in the right sidebar.

Model / tokens

Override via env vars before launch:

export PYNOTE_MODEL=claude-sonnet-4-6     # default
export PYNOTE_MAX_TOKENS=4096             # default

How the chat works

1. You type a question — "plot the last column" or "explain cell 3". pynote sends your message plus a snapshot of every cell's id, type, and source to the server extension.
2. The server forwards it to Claude with a system prompt that describes the propose_notebook_changes tool.
3. Claude replies with prose, and optionally calls the tool with a list of changes {op, target_cell_id, source}.
4. The panel shows the prose + a compact preview of each proposed change.
5. You type apply → pynote walks the change list and calls sharedModel.setSource(...) / sharedModel.insertCell(...) inside a single transaction. One undo reverts the whole batch.
6. Type cancel / discard / no to drop the proposal, or just keep chatting — a follow-up question clears the pending proposal automatically.

There's no apply button, no modal, no popup — everything is in the chat line.

Smoke test without Lab

Quickly verify Claude integration without running Jupyter:

export ANTHROPIC_API_KEY=sk-ant-...
python scripts/smoke_test.py

Expected: a proposal with at least one replace or insert_after change targeting the empty cell.

Dev workflow

# In one terminal — rebuild on TS changes
jlpm watch

# In another — Lab picks up the changes on refresh
jupyter lab

On Python-side changes, restart Lab.

Uninstall

jupyter server extension disable pynote
jupyter labextension uninstall pynote
pip uninstall jupyter-pynote

Troubleshooting

• Panel doesn't appear: run jupyter labextension list; if pynote isn't there, redo step 3.
• Panel says "No notebook open": that's literal — open a .ipynb in the same Lab window.
• ANTHROPIC_API_KEY is not set: the server process (the one running jupyter lab) didn't inherit the env var. Export it in that shell before launch.
• 403 / xsrf in browser console: shouldn't happen since requests go through @jupyterlab/services — but if you see one, disable any other extension that proxies Lab API calls.

License

MIT — see LICENSE.