• No API key. It runs the claude CLI you are already signed in to, so replies come out of your Claude Code subscription (or whatever the CLI is set up with).
• A popup at your code, not a chat sidebar. It opens in the corner with the selection highlighted, and gets out of the way when you want: hide it or step back into the code while Claude keeps working, then come back with the same key.
• Answers or edits, from one key. Ask a question and the answer streams in. Ask for a change and the selected lines are rewritten in place (one u undoes it), or shown as a diff you accept first.
• Always on the current code, and it remembers its own changes. Every follow-up sends the lines as they are now, with diffs of what Claude changed and what you changed since, so "undo your second change" works.
• Reads like your editor. Code in answers is highlighted with treesitter, names in it jump to their definitions over LSP, the popup follows your colorscheme (or one of 9 themes), and your own instructions go into every chat.
• Pure Lua, zero dependencies, light. setup() loads one module and takes 0.74 ms, nothing runs while Claude is not replying, and a streamed answer costs 0.71 ms of CPU per piece, even on a 10,000-line file.

Requirements

• Neovim 0.10, 0.11 or 0.12. CI runs the tests on 0.10.0 and the latest stable release on every change.
• Claude Code: the claude CLI on your PATH, signed in (run claude once in a terminal).

Install

With lazy.nvim:

{
  "kova1max/float-claude.nvim",
  opts = {},
}

Any other plugin manager works too: add the repository and call require("float-claude").setup(). Then run :checkhealth float-claude to see that the CLI is found and no other mapping took the plugin's keys. The full manual is :help float-claude.

Usage

1. Select some lines (V), or a block (<C-v>).
2. Press <leader>ai and type what you want.
3. Press Enter.

Claude decides whether your prompt asks for an answer or an edit:

  two_sum.lua · L5–13 · 9 lines
  ❯ what's the time complexity?
    O(n²): the two loops compare every pair in nums.
    Space is O(1).

  ❯ █

The header names the file and the lines Claude sees. Past questions are dimmed, replies are indented under them, and a fresh ❯ waits for a follow-up. In Normal mode, <leader>ai asks about the current line, or goes back to the open popup.

Ask

The answer streams in. Names from the code and code blocks are highlighted in the file's language. Keep asking: follow-ups continue the same conversation.

Edit in place

Ask for a change ("use a table of seen numbers, one pass", the GIF at the top) and the selected lines are rewritten in the buffer, streaming in line by line. The popup stays open with refactored 9 → 9 lines · u to undo, so you can keep going ("also rename x"). Each edit is one u. Select with <C-v> and the edit replaces just the block's columns on each line.

Review before it lands

With review_edits = true, an edit shows as an inline diff (removed lines red, new lines green) and the buffer stays untouched until you decide: <leader>yy accepts, <leader>nn discards.

Hide it, keep going

<C-q> hides the popup and <C-o> steps back into the code with the popup still on screen. Either way Claude keeps answering or editing. <leader>ai brings you back to the same conversation, and require("float-claude").status() puts a spinner (then ● for an unseen reply) in your statusline. Up to 5 chats live side by side, one per selection (<leader>an / <leader>ap / <leader>al), and <leader>aH reopens a saved chat, also after restarting Neovim.

Jump from the answer to the code

With an LSP server attached, the names in an answer that the server knows are underlined. In the conversation (<C-w>w from the prompt, or a click), Enter or a click jumps to the definition and K shows the hover.

And also

• One-key actions: <leader>ae explain, <leader>af fix the diagnostics, <leader>at tests, <leader>ay types, <leader>ad doc comments, <leader>as simplify; for Markdown, commit messages and comments <leader>ar reword, <leader>ah shorten, <leader>ag grammar. Add your own in actions.
• Several ranges at once: <leader>am collects separate selections, then one question (or one edit, one u) covers them all.
• @ mentions: @diagnostics, @<buffer> or @<path> sends more context, with completion.
• Model and effort per chat: :FloatClaudeModel sonnet, :FloatClaudeEffort high, shown in the header, switched without restarting the chat.
• Command palette: :FloatClaude lists everything through vim.ui.select, so a picker that replaces it (telescope, fzf-lua, snacks) shows it.
• Your instructions: prompt = "answer in Ukrainian" (or a function of the filetype), and an opt-in per-project file.
• Themes: theme = "auto" follows your colorscheme; or tokyonight, catppuccin, gruvbox, nord, onedark, rose-pine, dracula, night-owl, float-claude; :FloatClaudeTheme to try them.

Every feature has its section in :help float-claude (doc/float-claude.txt).

Keys

Key	Where	Action
<leader>ai	Visual	Ask about the selection (v, V: lines; <C-v>: the block's columns)
<leader>ai	Normal	Back to the open popup; with none open, ask about the current line
<CR>	Popup	Send the prompt
<S-CR> or <C-j>	Popup	New line in the prompt (<C-j> if your terminal does not send Shift-Enter)
<C-u> / <C-d>, wheel, <Up> / <Down>	Popup	Scroll the conversation
q	Popup, Normal mode	Close and end the chat; a half-written edit is undone
<C-q>	Popup	Hide it; the chat goes on
<C-o>	Popup	Back to the code; the popup stays on screen
<leader>yy / <leader>nn	Popup, with review_edits	Accept / discard the diff
<CR>, click / K	Conversation, on an underlined name	Jump to its definition / show its hover
<leader>ae, <leader>af, ...	Normal, Visual, popup	One-key actions
<leader>am / <leader>ax	Visual, Normal	Add the selection to the ranges to ask about together / clear them
<leader>an / <leader>ap, <leader>al	Normal	Next / previous open chat, or pick one
<leader>aH	Normal	Reopen a saved chat
@	Popup	Mention a file, a buffer or @diagnostics
:FloatClaude	Normal, Visual	Pick any of the above from a list

Every key is a setup option, and any of them can be false.

Configuration

Every option, with its default (opts for lazy.nvim, or require("float-claude").setup({...})). A table option (like cursor_offset) is replaced as a whole, so give all of its fields; actions is the exception, merged action by action.

{
  keymap = "<leader>ai", -- Visual + Normal mode: ask about the selection / back to the popup; false to map require("float-claude").ask yourself
  cmd = { "claude" }, -- the CLI, with extra arguments if you like
  model = false, -- each chat's model to start with: "opus", "sonnet", "haiku", "fable" or a model name; false: the CLI's default
  effort = false, -- each chat's effort to start with: "low", "medium", "high", "xhigh" or "max"; false: the CLI's default
  mentions = { max_lines = 500 }, -- @ mentions in the prompt, each sending at most max_lines lines; false: off
  prompt = false, -- your instructions for every chat: a string, or function(ctx) returning one (ctx: buf, file, filetype)
  project_prompt = false, -- e.g. ".float-claude.md": that file, from the file's directory or above, is added
  context_lines = 40, -- unselected lines sent above and below the selection
  diagnostics = true, -- send the diagnostics in the selected lines too
  actions = { -- one-key actions; each is merged on its own, false removes one
    explain = { key = "<leader>ae", prompt = "Explain what this code does." },
    fix = { key = "<leader>af", prompt = "Fix the problems the diagnostics report in this code; with none listed, fix the bugs you find." },
    tests = { key = "<leader>at", prompt = "Write unit tests for this code, in the project's test style, as an answer with a code block; do not edit the code." },
    types = { key = "<leader>ay", prompt = "Add type annotations to this code, in the language's usual way. Change nothing else." },
    docstring = { key = "<leader>ad", prompt = "Add a doc comment to each function in this code, in the language's usual style. Change nothing else." },
    simplify = { key = "<leader>as", prompt = "Simplify this code without changing what it does." },
    reword = { key = "<leader>ar", prompt = "Reword this text so it reads clearly; keep its meaning and tone.", prose = true },
    shorten = { key = "<leader>ah", prompt = "Shorten this text; keep its meaning.", prose = true },
    grammar = { key = "<leader>ag", prompt = "Fix the spelling, grammar and punctuation of this text. Change nothing else.", prose = true },
  },
  prose_filetypes = { "markdown", "gitcommit", "text" }, -- buffers offering the prose actions instead of the code ones
  max_chats = 5, -- chats open at once (one shown, the others hidden); one more ends the one used longest ago
  history = { max = 100, days = 30 }, -- save chats to reopen them; either can be false; false: off
  status = { spinner = { "◐", "◓", "◑", "◒" }, unread = "●" }, -- what status() shows
  width = 50, -- popup width in columns, capped to the screen; longer lines wrap
  max_height = false, -- popup height cap in rows; false: as tall as the screen allows
  placement = "top-right", -- or "cursor": next to the cursor
  top_right_offset = { row = 0, col = 0 }, -- "top-right": rows from the top, columns from the right edge
  cursor_offset = { row = 1, col = 3 }, -- "cursor": rows below, columns right of the cursor (above it when there is no room below)
  theme = "auto", -- or a preset: night-owl, dracula, tokyonight, catppuccin, gruvbox, nord, onedark, rose-pine, float-claude
  colors = {}, -- single colors over the theme, e.g. { bg = "#000000", prompt = "#ff9e64" }
  review_edits = false, -- true: show an edit as a diff and apply it only once accepted
  close_after_edit = false, -- true: close the popup (ending the chat) once an edit is applied
  close_on_leave = false, -- true: leaving the popup (other than with unfocus_key or a jump) closes it
  -- Normal mode keys for the open chats:
  next_chat_key = "<leader>an", -- show the next chat
  prev_chat_key = "<leader>ap", -- show the previous one
  chats_key = "<leader>al", -- pick an open chat from a list
  history_key = "<leader>aH", -- pick a saved chat to reopen
  add_range_key = "<leader>am", -- Visual and Normal mode: add the selection (or the current line) to the ranges to ask about together
  clear_ranges_key = "<leader>ax", -- Normal mode: clear them
  -- Keys in the popup:
  submit_key = "<CR>", -- send the prompt (Insert and Normal mode)
  newline_keys = { "<S-CR>", "<C-j>" }, -- Insert mode: a new line in the prompt
  close_key = "q", -- Normal mode: close the popup and end the chat
  hide_key = "<C-q>", -- hide it; the chat goes on, keymap brings it back
  unfocus_key = "<C-o>", -- back to the code, the popup stays on screen
  accept_key = "<leader>yy", -- with review_edits: accept the diff (Normal mode)
  discard_key = "<leader>nn", -- with review_edits: discard it
  scroll_up_key = "<C-u>", -- scroll the conversation half a screen
  scroll_down_key = "<C-d>",
  line_up_key = "<Up>", -- scroll it a line (in a multi-line prompt, first move between its lines)
  line_down_key = "<Down>",
  jump_key = "<CR>", -- in the conversation, on an underlined name: go to its definition
  hover_key = "K", -- in the conversation, on an underlined name: show its hover
  insert_keys = { "i", "a" }, -- in the conversation: back to the prompt
}

For a statusline, require("float-claude").status() returns a spinner while a hidden chat answers, ● for a reply you have not seen, and "" otherwise. With lualine:

sections = {
  lualine_x = { function() return require("float-claude").status() end, "filetype" },
}

What Claude sees

The file name and filetype, the selected lines, up to 40 lines around them, and the diagnostics in the selected lines, plus whatever you @-mention. Every follow-up resends the selected lines as they are at that moment, so Claude always works on the current code, and starts with diffs of what changed since its last reply: its own edits (applied, accepted or discarded) apart from yours.

The CLI runs with no tools, MCP servers, hooks or settings files: Claude reads only what the popup sends and changes only the selected lines, in your buffer, one u away. Your CLAUDE.md files and settings.json are not used; auth comes from your claude login or from environment variables such as ANTHROPIC_API_KEY in the shell Neovim runs in.

How it compares

Facts from each project's README (November 2026). Pick by the shape you want: float-claude is for quick questions and edits about the code in front of you; for an agent that runs commands and edits across files, use Claude Code itself (claudecode.nvim puts it in a split), avante.nvim or codecompanion.nvim.

	float-claude.nvim	claudecode.nvim	avante.nvim	codecompanion.nvim	CopilotChat.nvim
UI	popup at the selection	Claude Code's terminal UI in a split	sidebar	chat buffer, inline	chat window (split or float)
Talks to	claude CLI, headless	claude CLI, over its IDE protocol	API providers (Claude, OpenAI, Gemini, Copilot, ...), ACP agents	HTTP adapters (Anthropic, OpenAI, Copilot, Ollama, ...), ACP agents	GitHub Copilot, plus custom providers
Claude subscription without an API key	yes	yes	yes (auth_type = "max", or ACP)	through the Claude Code ACP agent	no (Claude models through Copilot)
Edits	in place, streamed, one u; optional diff review	diffs to accept or deny	diffs applied to the file	approve or reject agent changes	diffs to accept
Dependencies	none	none (snacks.nvim optional)	plenary.nvim, nui.nvim, a make build step	plenary.nvim, nvim-treesitter	plenary.nvim, curl, make tiktoken
Neovim	0.10+	0.8+	0.11+ (its badge: 0.12+)	not stated	0.10+

float-claude and claudecode.nvim work side by side; their suggested keys both live under <leader>a, so check :checkhealth float-claude for clashes.

FAQ

Do I need an API key? No. float-claude starts the claude CLI, which uses whatever it is signed in with: a Claude Pro or Max subscription, or an API key, Bedrock or Vertex set up in its environment.

Does it send my whole file? No: the selection, up to context_lines (40) lines around it, its diagnostics, and what you @-mention.

Can Claude run commands or touch other files? No. The CLI runs with no tools; a reply is either an answer or the replacement for the selected lines.

Which model does it use? The CLI's default, unless you set model. :FloatClaudeModel and :FloatClaudeEffort switch the open chat's.

Does it work in tmux or over SSH? Yes, it is all in the terminal. If Shift-Enter does not reach Neovim, <C-j> adds a new line.

Is it slow on big files? No. Measured on a 10,000-line file with an LSP server attached: 0.71 ms of CPU per streamed piece of an answer and 0.5 ms per piece of an edit; while an answer streams and its names are looked up, the editor is never busy for more than 61 ms at a time. The claude process ends when its chat closes or Neovim quits. tests/bench.lua measures it on your machine.

Troubleshooting

• Updated the plugin and it misbehaves? Restart Neovim: a running Neovim keeps the old code.
• Start with :checkhealth float-claude. It checks the Neovim version, that setup() ran, that the CLI runs, and which of the plugin's keys another mapping took.
• "claude exited" or no reply: run claude in a terminal and sign in. Behind a custom install, point cmd at it.
• A key does nothing: another plugin maps it (:checkhealth float-claude names it); change the key in setup().
• No @ completion menu on Neovim 0.10: add noselect to completeopt, or press <C-x><C-o>.
• Code in answers is not highlighted: install the treesitter parser for that language (Neovim ships Lua, Vim script, C, Markdown and a few more).
• Names in answers are not underlined: they need an LSP server attached to the file.

Development

Run the tests (headless; a fake CLI stands in for Claude and a fake LSP server for a language server, so no network or account is needed):

nvim --clean --headless -l tests/run.lua

ONLY=<text> runs just the tests whose name contains the text.

Measure what the plugin costs (setup time, what runs while idle, CPU per streamed piece of a reply on a 10,000-line file, the longest the editor is blocked, memory), with the same fake CLI and LSP server:

nvim --clean --headless -l tests/bench.lua

The GIFs are recorded with VHS from the tapes in assets/, with the fake CLI scripting Claude's replies: assets/record.sh (needs vhs, and lua-language-server for lsp.tape).