notebook-editor-mcp

# Local Notebook Agent Editor ## Demo https://github.com/user-attachments/assets/e52a4ee7-3bb2-4953-a515-4c32f8742c6f A local FastAPI + React editor for working on Jupyter notebooks with scoped AI agent editing. You open a local `.ipynb` file (or a project folder), edit and execute cells, and save in place. Each agent turn has a **scope**: in **Blocking** mode the agent may only rewrite the source of cells you explicitly mark editable; in **Trusted** mode the whole notebook is editable and the agent may add, delete, reorder, and retype cells. Either way the backend validates and applies (or rejects) every change — the agent never mutates your notebook directly — and every change is reviewable as a diff and undoable. Everything runs on your machine and binds to loopback only. Read [Security Limits](#security-limits) before using it with untrusted notebooks. ## Install You work in the browser tab. An MCP client — Claude Code, for instance — starts the editor and opens it for you. [Running it locally](#run-it-locally-instead) is for working on the editor itself. Install it into the environment your notebooks run in, so your cells keep the packages they already have: ```bash /abs/path/to/.venv/bin/pip install notebook-editor-mcp claude mcp add agent-notebook -- \ /abs/path/to/.venv/bin/notebook-editor-mcp \ --workspace-root /abs/path/to/project ``` That environment is the one your cells run in. Left alone, the editor executes them wherever it was installed — which is the right answer here and the wrong one for `uvx` or `pipx`, where a notebook opening with `import pandas` fails on its first cell. So to keep the editor out of your project instead, point `--kernel-python` at the environment your cells need: ```bash claude mcp add agent-notebook -- \ uvx --from notebook-editor-mcp notebook-editor-mcp \ --workspace-root /abs/path/to/project \ --kernel-python /abs/path/to/.venv/bin/python ``` `--workspace-root` confines the editor to one directory. Us…