openvim

A Neovim distribution with AI built in.

Inline autocomplete, inline edit, and an agent that edits your code from a command. Bring your own model: Claude, a local Ollama model, or OpenAI.

_{:chat editing a file in place. Illustrative demo.}

---

openvim is not a fork of Neovim's C core. It is a config and plugin layer that boots plain Neovim against a bundled setup, so it tracks upstream Neovim for free and stays easy to hack on in pure Lua.

Features

• Inline autocomplete Copilot-style ghost text as you type. <Tab> to accept.
• Inline edit Select code, describe a change, get it rewritten in place.
• :chat agent Tell it what to do and it edits the current file directly.
• Pluggable backends Claude, Ollama (local, free), or OpenAI. Your choice persists.

Quick start

git clone https://github.com/natedemoss/openvim
cd openvim

# pick a backend (see "Models" below), then:
nvim -u config/init.lua

Requirements: Neovim 0.10+ and curl on PATH.

Run it as openvim

Put bin/ on your PATH, then launch with one word:

openvim

Platform	Launcher
Windows (PS)	bin\openvim.ps1
Windows (cmd)	bin\openvim.cmd
macOS / Linux	bin/openvim

Backends

Pick a provider once and openvim remembers it (:OpenvimProvider, saved to ~/.openvim/settings.json).

# Claude (best quality)
export ANTHROPIC_API_KEY=sk-ant-...

# OpenAI
export OPENAI_API_KEY=sk-...

# Ollama (local, free, private) — see models below

Models (Ollama)

openvim uses two kinds of model: a base model for autocomplete (fill-in-the-middle) and an instruct model for :chat and inline edit. Qwen2.5-Coder is the best local code family right now. Pick a size to match your hardware.

Autocomplete (base / FIM)

Model	Size	Best for
qwen2.5-coder:0.5b-base	0.5 GB	CPU-only, snappiest (~1s)
qwen2.5-coder:1.5b-base	1.0 GB	CPU with a little patience
qwen2.5-coder:3b-base	1.9 GB	light GPU
qwen2.5-coder:7b-base	4.7 GB	GPU, best quality

Chat + inline edit (instruct)

Model	Size	Best for
qwen2.5-coder:1.5b	1.0 GB	CPU, quick answers
qwen2.5-coder:7b	4.7 GB	balanced quality (GPU recommended)
qwen2.5-coder:14b	9.0 GB	high quality, needs a real GPU
deepseek-coder-v2:16b	8.9 GB	strong alternative

# CPU-friendly default pairing:
ollama pull qwen2.5-coder:0.5b-base   # autocomplete
ollama pull qwen2.5-coder:1.5b        # chat / edit

On a CPU-only machine, use the small base model for autocomplete. A 7B instruct model on CPU can take minutes per response. A GPU changes everything.

Usage

Autocomplete

Type in insert mode, pause, and ghost text appears. <Tab> accepts, <C-]> dismisses.

Inline edit

Visually select lines, press <leader>ae (leader is space), and type an instruction:

add input validation

:chat agent

Open a file and tell openvim what to change. It rewrites the file in place.

:chat add type hints and a docstring
:chat refactor this into smaller functions

Commands

Command	Description
:chat / :Chat	Agent: edit the current file from an instruction
:OpenvimEdit	AI-edit the selected lines (use a visual range)
:OpenvimToggle	Enable / disable inline autocomplete
:OpenvimProvider	Switch backend (claude / ollama / openai)
:OpenvimStatus	Show current provider, model, and settings

Configuration

Defaults live in config/lua/openvim/init.lua. Override with a table:

require('openvim').setup({
  ai = {
    provider = 'ollama',
    debounce_ms = 250,
    models      = { ollama = 'qwen2.5-coder:1.5b-base' }, -- autocomplete
    chat_models = { ollama = 'qwen2.5-coder:7b' },        -- chat / edit
  },
})

How it works

openvim/
  bin/                  launchers (openvim, .ps1, .cmd)
  config/init.lua       entry point (nvim -u)
  config/lua/openvim/
    init.lua            defaults, branding, setup
    dashboard.lua       start screen
    settings.lua        persisted user settings
    ai/
      complete.lua      ghost-text autocomplete engine
      edit.lua          inline edit on a selection
      agent.lua         :chat whole-file agent
      backends/         claude.lua, ollama.lua, openai.lua

Backends expose two calls: complete() for fill-in-the-middle autocomplete and ask() for instruction-following chat and edits.

License

MIT