Notebook Live Edit MCP
An MCP server that runs inside VS Code and lets an AI coding agent work in the notebook you have open: read cells, edit them, run them in the notebook's kernel, and look at the outputs, including plots.
Edits go through the live notebook editor, not the .ipynb file on disk. Your kernel state, unsaved changes and outputs stay intact, and you see every change happen in the editor.
Features
- Live editing: list, insert, edit and delete cells in the open notebook.
- Running cells: run a cell in the notebook's current kernel and get its output back: text, errors with tracebacks, and images such as matplotlib plots.
- "This cell": the agent can see which cell you have selected, so you can say "change this cell to ...".
- Diffs: every edit returns a line diff, and a VS Code diff tab beside the notebook shows the latest edit. The tab updates in place, wherever you move it.
- Long-running cells: a call never blocks for more than 50 seconds. A cell that is still running is reported as running, together with its output so far.
- Kernel control: interrupt or restart the kernel. Restarting asks for your confirmation in VS Code.
- Several windows: each VS Code window runs its own server, and an agent started in a window's terminal talks to that window.
Setup
- Install the extension.
- Add the MCP server to your agent's configuration, once. On the first start the extension offers this for each agent CLI it finds installed (Claude Code, Codex, or both). You can also run Notebook Live Edit MCP: Add MCP Server to Claude Code or ... to Codex from the command palette. Each runs that agent's
mcp add command, shown below, in a terminal. Nothing gets installed.
- Open a notebook, open a new terminal in that VS Code window, and start the agent there.
The extension sets NOTEBOOK_LIVE_EDIT_PORT and NOTEBOOK_LIVE_EDIT_TOKEN in the integrated terminals of each VS Code window. An agent started anywhere else won't have the token, and the server will refuse it.
Claude Code
claude mcp add --transport http -s user notebook \
'http://127.0.0.1:${NOTEBOOK_LIVE_EDIT_PORT:-7337}/mcp' \
--header 'Authorization: Bearer ${NOTEBOOK_LIVE_EDIT_TOKEN}'
Keep the single quotes, so the variables are expanded by Claude Code when it connects, not by your shell now. Claude Code follows the port of the window it was started in, so it works with several windows open.
Codex
codex mcp add notebook --url http://127.0.0.1:7337/mcp --bearer-token-env-var NOTEBOOK_LIVE_EDIT_TOKEN
Codex can't read the port from an environment variable, so it always connects to port 7337, which belongs to the first VS Code window that opened. In other windows the token doesn't match and Codex gets an error instead of reaching the wrong notebook. To use Codex in another window, close the other windows or give that window's workspace a fixed notebookLiveEdit.port and add a second Codex server for it.
| Tool |
What it does |
list_cells |
All cells with index, kind, execution count, status and source; marks the selected cell |
get_selected_cell |
The cell you have selected in VS Code |
get_outputs |
A cell's outputs (text, errors, images); can wait for a running cell to finish |
edit_cell |
Replace a cell's source; returns a diff and updates the diff tab |
insert_cell |
Insert a code or markdown cell |
delete_cell |
Delete a cell |
run_cell |
Run a cell and return its outputs, waiting at most 50 seconds |
interrupt_kernel |
Interrupt the kernel; variables are kept |
restart_kernel |
Restart the kernel; VS Code asks you to confirm |
Security
The server listens on 127.0.0.1 only, so other machines cannot reach it. Every request must carry a random token. Each workspace gets its own, created once and kept across reloads, so an agent started in one VS Code window can't reach the notebooks of another. Requests from web pages, which always send an Origin header, are rejected.
Anyone who has the token can run code in your kernel. It is visible to programs started from VS Code's terminals, and it is stored with VS Code's other extension data.
Settings
| Setting |
Default |
Description |
notebookLiveEdit.port |
7337 |
Preferred port. If another window already uses it, the window picks a free port and tells its terminals through NOTEBOOK_LIVE_EDIT_PORT. Reload the window after changing it. |
Dev containers and remote windows
The extension runs where the notebook's workspace is, so in a dev container it has to be installed inside the container.
Requirements
- VS Code 1.90 or newer
- The Jupyter extension, for the kernel commands
- An MCP client that supports Streamable HTTP and custom headers, for example Claude Code
Limitations
- Only runs started through the server are tracked as running. Cells you start yourself show their final status once they finish.
- Interactive outputs that need a browser, such as Plotly figures and widgets, are reported as omitted.
- The thing is vibe coded (and vibe documented), obviously. Use at your own peril...
License
Apache-2.0
