🪄Ensorcell -- AI for Jupyter Notebooks

Your notebook, enchanted. A full AI agent embedded right in your cells -- no context switching, no chat sidebar, no copy-paste. Claude sees what you see, works where you work, and has full access to your tools.

en·sor·cell /ɪnˈsɔːrsɛl/ verb -- to enchant; to bewitch. From French ensorceler. See also: what happens to your notebook cells when AI is embedded right where you work.

Works with any Jupyter notebook in VS Code. Install the extension, make sure you have Claude Code CLI, and you're ready to go.

Why Ensorcell?

If you already work in Jupyter notebooks, you already know the best way to think: prose and code together, step by step, building understanding as you go. Ensorcell adds one thing -- Ctrl+Alt+Enter -- and suddenly your notebook can think along with you!

No new tool to learn. No cloud platform to upload to. No chat sidebar to copy-paste from. No terminal window to multitask in. Just a full AI agent, right here in your cells, sharing your context, running on your machine.

A Full Agent, Not Just Q&A

Ensorcell doesn't limit Claude to answering questions. Claude has full access to your filesystem, terminal, web search, and every other tool in Claude Code. Ask it to research, write code, analyze data, read files, search the web -- whatever the task requires. Debugging a data pipeline? Ask Claude to read the error, check the source file, and suggest a fix -- all without leaving your notebook. Your notebook becomes a command center, not just a conversation.

Shared Context by Default

Every AI tool asks you to explain what you're working on. Ensorcell doesn't -- Claude sees every cell above your prompt automatically. Markdown, code, outputs, prior conversations -- it's all there. No dragging cells into a chat panel, no special syntax to reference previous results, no manually curating what the AI can see. You and Claude are literally looking at the same thing.

Code cell outputs are included too. Your model training cell threw a cryptic CUDA error? Claude already sees it -- just ask what went wrong.

Context You Control

You decide what Claude sees, two ways:

By position -- ask from cell 10 and Claude sees cells 1 through 10. Ask from the top and Claude starts fresh.

By visibility -- click the eye icon on any cell to hide it from Claude's context. On code cells, you get two independent toggles: one for the source, one for the output. Hide a massive data dump but keep the code visible. Hide the boilerplate but keep the results. Toggle state persists when you save.

Sometimes you want to drill deep into a tangent -- a five-exchange rabbit hole exploring some detail that caught your attention. That's great! That's how learning works. But when you move on, those exchanges are still eating, or otherwise misdirecting / polluting, context even though they're no longer relevant. Toggle them hidden: they stay in your notebook as a record of your thinking, but Claude stops seeing them. Context stays clean without losing your work.

This is dialogue engineering -- the methodology pioneered by Jeremy Howard at Answer.AI -- implemented natively in the tool you already use.

Local-First

Ensorcell runs on your machine. Claude Code CLI has access to everything you do: your filesystem, your terminal, your development environment. Fine-tuning a model on your local GPU? Claude can see your training logs, read your config files, and help you iterate -- all running on your hardware. Things a cloud tool can't touch.

And because Ensorcell is a VS Code extension, it works where the coolest Jupyter notebook users already are.

How It Works

1. Write your prompt in a markdown cell
2. Press Ctrl+Alt+Enter (or click the ✨ sparkle icon in the cell toolbar)
3. Your cell gets a [Me to Claude] header with a timestamp
4. The ensorcell animation plays while Claude works
5. Claude's response appears in a new [Claude] cell below, with elapsed time

That's it. Everything else is already how notebooks work.

System Prompts

Shape Claude's behavior with a layered system prompt:

Built-in default -- Ensorcell ships with a system prompt that orients Claude to the notebook environment, explains the conversation markers, and encourages proactive tool use. You never have to configure this.

Per-workspace -- set ensorcell.systemPromptFile to a markdown file in your workspace for project-specific instructions.

Per-notebook -- add [system] cells directly in the notebook for instructions that travel with the document:

[system]

You are a close reading companion helping me analyze
this article. Be critical, cite specific passages,
and challenge claims when appropriate.

All three layers stack. Type [system] in any casing -- it renders as bold [System] with an amber border. System cells are excluded from conversation context -- Claude receives them as instructions, not as part of the dialogue.

Place them anywhere. Use as many as you want. Change Claude's role mid-notebook by adding a new one. Doing close reading in the first half and data analysis in the second? Add a new system cell at the transition point.

Visual Cell Styling

AI cells are visually distinct so you can scan your notebook and know what's what:

• Gold -- your prompts ([Me to Claude])
• Violet -- Claude's responses ([Claude])
• Amber -- system prompts ([System])

Each cell type gets a colored left border, a gradient background, and an icon label in the cell status bar.

Use Cases

Software development -- Your API integration is returning 403s and you can't figure out why. Write the request in a code cell, run it, and ask Claude. It sees your code, the error response, and your earlier cells where you set up the auth headers. It can also read the API docs from your project files and search the web for known issues -- a full coding agent, not just a code explainer.

Research and close reading -- Read an article section by section. After each passage, ask Claude to explain unfamiliar concepts, challenge the author's claims, or connect ideas to things you've discussed earlier. Your conversation builds cumulatively -- by the end, you have a notebook that's both the text and your annotated dialogue with it. Inspired by AI-assisted close reading as practiced by Jeremy Howard and fast.ai.

Data analysis -- Your scatter plot looks wrong but you're not sure why. Ask Claude -- it can see both your plotting code and the rendered output. It suggests checking for outliers, you add a filter cell, rerun, and ask again. The whole investigation lives in one notebook.

Learning -- Working through a textbook or course? Build a notebook of notes, questions, and AI-assisted explanations. Don't understand a proof? Ask. Want a simpler example? Ask. The notebook becomes both your study material and your tutor.

Why This Exists

I built Ensorcell because I love notebooks. They're my preferred environment for everything -- programming, research, studying, data exploration. Literate programming is just how my brain works best. I also really dig VS Code, and the way Jupyter notebooks work in it.

When I discovered dialogue engineering and the excellent SolveIt, I wanted that same workflow -- AI as a thinking partner, shared context, iterative small steps -- but I wanted it here, in my notebooks, in VS Code, on my machine, with access to my local tools and data. So I built it.

Requirements

• Claude Code CLI installed and authenticated
• VS Code with the Jupyter extension

Configuration

Keybinding

Ctrl+Alt+Enter -- fires only when focused on a markdown cell in a Jupyter notebook. Does not conflict with standard notebook keybindings (Shift+Enter for run cell, Ctrl+Enter for run in place).

Debugging

Every invocation logs a summary to the Output panel (Ctrl+Shift+U > select "Ensorcell"):

[6:07 PM] Prompt: "What does the code in the above cell mean?..." (cell 2)
[6:07 PM] Context: 1,176 chars, 3 cells (1 hidden)
[6:07 PM] Sending to Claude...
[6:07 PM] Response: 884 chars, 9s, 1 turn, $0.06

For deeper investigation, enable ensorcell.debug in settings. This writes a full forensic log for every invocation -- the complete notebook snapshot, the exact context and system prompt sent to Claude, the full response, and Claude's raw JSON with token counts, cost, tool usage, and permission denials. Open the log directory from the command palette with Ensorcell: Open Debug Logs.

Build from Source

npm install
npx tsc -p ./
npm run package
code --install-extension ensorcell-0.0.1.vsix --force

Reload VS Code after installing (Ctrl+Shift+P → "Developer: Reload Window").

First magic moment: Open any .ipynb, add a markdown cell, type a question, press Ctrl+Alt+Enter.