jove.nvim

Edit Jupyter .ipynb notebooks in Neovim as if they were native Python buffers — with rendered cell outputs, real LSP/copilot/treesitter, and proper round-tripping to disk. No otter, no quarto, no temp files, no lost outputs.

How it works

.ipynb on disk (JSON)
    │  BufReadCmd  ── jupytext CLI via stdio (no temp file) ──▶
buffer  buftype=acwrite, filetype=python
    │  pyright/ruff/copilot/treesitter attach natively
    │  molten attaches → kernel + inline outputs
    │  BufWriteCmd ── jupytext --update via stdio ──▶
.ipynb on disk (JSON, outputs preserved)

Requirements

• Neovim ≥ 0.10
• jupytext on $PATH
• molten-nvim — optional but required for kernel execution and rich outputs
• nvim-various-textobjs — optional, for ic/ac cell text-objects

Install (lazy.nvim)

{
  "nghiant03/jove.nvim",
  lazy = false,                        -- BufReadCmd must be registered before .ipynb is opened
  dependencies = { "benlubas/molten-nvim" },
  opts = {
    auto_kernel = true,                -- MoltenInit from kernelspec on open
    auto_import_outputs = true,        -- restore outputs from .ipynb JSON
    auto_export_outputs = true,        -- merge molten outputs back on :w
    keymap = {
      run_cell  = "<leader>x",
      next_cell = "]h",
      prev_cell = "[h",
    },
  },
}

opts = {} (or config = true) is required — without it lazy.nvim never calls require("jove").setup(...), and the keymaps under keymap = {} are not registered. The BufReadCmd/BufWriteCmd handlers are installed from plugin/jove.lua at startup either way, but you'll lose the cell motions and run_cell binding.

Setup

1. Install jupytext

pip install jupytext      # or: pipx install jupytext / conda install jupytext
jupytext --version        # must be on $PATH for the nvim process

If you launch Neovim from a conda env that doesn't have jupytext, prepend the env's bin/ to vim.env.PATH early in your init.lua, or set opts.jupytext = "/abs/path/to/jupytext".

2. Install molten-nvim (recommended)

Molten provides the kernel and inline outputs. Without it jove still edits notebooks, but :JoveRunCell and outputs do nothing. Minimal lazy spec:

{
  "benlubas/molten-nvim",
  build = ":UpdateRemotePlugins",
  init = function()
    vim.g.molten_image_provider       = "image.nvim" -- or "snacks.nvim"
    vim.g.molten_auto_open_output     = false
    vim.g.molten_virt_text_output     = true
    vim.g.molten_virt_lines_off_by_1  = true         -- correct for `# %%` cells
    vim.g.molten_wrap_output          = true
  end,
}

You also need a Jupyter kernel that matches your notebook's metadata.kernelspec.name — typically python3 from ipykernel:

pip install ipykernel
python -m ipykernel install --user --name python3

Verify inside nvim with :lua =vim.fn.MoltenAvailableKernels().

3. Avoid conflicts

• Do not load jupytext.nvim alongside jove — both register BufReadCmd on *.ipynb. Jove will hard-refuse and warn on startup; remove one.
• Do not duplicate molten's auto-init autocmds. If you copied the BufAdd *.ipynb → MoltenInit + MoltenImportOutput snippet from molten's README, delete it: jove already does this when auto_kernel = true and auto_import_outputs = true. Otherwise molten initializes twice and logs errors.

4. Verify

:checkhealth jove

Expect green checks for: Neovim ≥ 0.10, jupytext binary found, molten detected, no jupytext.nvim conflict. Then open any .ipynb:

nvim notebook.ipynb

You should land in a Python buffer split into # %% cells, with kernel status visible via molten and outputs rendered as virtual text.

Commands

Command	Action
:JoveRunCell	Run cell under cursor via molten
:JoveRunAbove	Run all cells from top to cursor
:JoveRunAll	Run every cell
:JoveNextCell / :JovePrevCell	Jump between cells
:JoveInitKernel	Start molten kernel for this notebook
:JoveImportOutputs	Re-display outputs stored in the .ipynb JSON
:JoveExportOutputs	Merge current molten outputs into .ipynb on disk
:checkhealth jove	Verify deps, versions, and conflicts

Comparison

	jove.nvim	jupytext.nvim	quarto-nvim + otter	jupynium.nvim
Native python LSP	yes	yes	proxy/chunked	yes
Inline outputs	via molten	no	via molten	via browser
No temp files	yes	no (writes sidecar)	yes	yes
Output persistence	yes (JSON merge)	no	manual	yes
Browser required	no	no	no	yes
Scope	small	small	wide (.qmd)	wide

Non-goals

• Markdown prose rendering between cells (markdown cells stay as # %% [markdown] comment blocks).
• Browser sync à la jupynium.
• A custom Jupyter kernel client (delegated to molten).
• Reimplementing jupytext format conversions in Lua.

License

MIT