nvim-ask

An inline AI code assistant for Neovim, powered by the Claude CLI. Select code, ask a question, get a streaming response in a floating overlay — then apply it directly to your buffer.

Requirements

• Neovim >= 0.9
• Claude CLI installed and authenticated (claude in your PATH)

Installation

lazy.nvim

{
  "your-username/nvim-ask",
  config = function()
    require("nvim-ask").setup()
  end,
}

Local development

{
  dir = "~/path/to/nvim-ask",
  config = function()
    require("nvim-ask").setup()
  end,
}

Configuration

require("nvim-ask").setup({
  keybind = "<leader>ai",       -- keybind for visual + normal mode
  window = {
    width = 0.8,                -- fraction of editor width
    height = 0.8,               -- fraction of editor height
    border = "rounded",         -- border style
    min_width = 40,             -- minimum terminal width to open
    min_height = 15,            -- minimum terminal height to open
  },
  claude = {
    model = nil,                -- override model (e.g. "sonnet")
    max_tokens = nil,           -- max output tokens
    timeout = 120,              -- request timeout in seconds
  },
})

Usage

With a code selection

1. Visually select code (V, v, or <C-v>)
2. Press <leader>ai (or your configured keybind)
3. Type your instruction in the prompt (e.g. "refactor this to use async/await")
4. Press <CR> to send
5. Press <CR> in the response to replace the original selection

Without a selection

1. Press <leader>ai in normal mode
2. Type a general question
3. Press <CR> to send

You can also use the :NvimAsk command directly.

Keybinds (inside the overlay)

Key	Context	Action
<CR>	Prompt buffer	Send request
<CR>	Response buffer	Apply code to original buffer
y	Response buffer	Yank response to clipboard
r	Response buffer	Retry (re-send same prompt)
<Tab>	Any	Cycle focus to next section
<S-Tab>	Any	Cycle focus to previous section
q	Any	Close the overlay

How it works

• Invokes claude as a child process via vim.fn.jobstart
• Streams the response token-by-token using --output-format stream-json
• Parses markdown code fences from the response for syntax highlighting and apply/yank
• Disables all Claude tools (--tools "") so you only get text responses

License

MIT