tau.nvim

A minimal Neovim plugin for applying LLM edits to a visual selection or range.

Tau sends your instruction, selected code, nearby context, and optional context files to an OpenAI-compatible /chat/completions API, then streams the proposed edit back into Neovim for review.

Install

Requirements: Neovim 0.10+, Bun, and an OpenAI-compatible API endpoint.

Using lazy.nvim:

{
  "timothyckl/tau.nvim",
  lazy = false,
  build = "cd cli && bun run build",
  config = function()
    require("tau").setup({
      api_url = "https://api.openai.com/v1", -- Tau appends /chat/completions
      api_key = vim.env.OPENAI_API_KEY,
      model = "gpt-4o",
    })
  end,
  keys = {
    { "<leader>t", ":Tau<CR>", mode = "v", desc = "Tau: edit selection" },
    { "<C-t>", ":TauContext<CR>", mode = "n", desc = "Tau: context files" },
  },
}

Using vim.pack:

tau.nvim requires Bun to build the CLI binary at cli/tau. The plugin is not usable until that build has completed, because :Tau, :TauContext, and :TauCancel all rely on the compiled CLI being present.

Minimal install:

vim.pack.add({
  {
    src = "https://github.com/timothyckl/tau.nvim"
  },
})

Complete example with automatic rebuilds on install and update:

-- Define this before vim.pack.add() so fresh installs also build cli/tau.
vim.api.nvim_create_autocmd("PackChanged", {
  callback = function(ev)
    local data = ev.data
    if data.spec.name ~= "tau.nvim" then
      return
    end
    if data.kind ~= "install" and data.kind ~= "update" then
      return
    end

    local result = vim.system({ "bun", "run", "build" }, {
      cwd = data.path .. "/cli",
      text = true,
    }):wait()

    if result.code ~= 0 then
      local output = vim.trim(
        (result.stderr and result.stderr ~= "") and result.stderr or (result.stdout or "")
      )
      if output == "" then
        output = ("bun run build exited with code %d"):format(result.code)
      end
      vim.notify(
        ("tau.nvim: failed to build cli/tau\n%s"):format(output),
        vim.log.levels.ERROR
      )
    end
  end,
})

vim.pack.add({
  { src = "https://github.com/timothyckl/tau.nvim" },
})

require("tau").setup({
  api_url = "https://api.openai.com/v1", -- Tau appends /chat/completions
  api_key = vim.env.OPENAI_API_KEY,
  model = "gpt-4o",
})

vim.keymap.set("v", "<leader>t", ":Tau<CR>", { desc = "Tau: edit selection" })
vim.keymap.set("n", "<C-t>", ":TauContext<CR>", { desc = "Tau: context files" })
vim.keymap.set("n", "<leader>T", ":TauCancel<CR>", { desc = "Tau: cancel request" })

The PackChanged hook rebuilds cli/tau after every install and update, and vim.system(...):wait() keeps the build synchronous so the CLI is present before first use.

If you hit ENOENT: no such file or directory ... /cli/tau, the CLI was not built yet. Rebuild it manually with:

cd <plugin>/cli && bun run build

For Ollama-compatible local servers:

require("tau").setup({
  api_url = "http://localhost:11434/v1",
  api_key = "ollama",
  model = "qwen2.5-coder:7b",
})

Usage

Select code in visual mode, then run:

:Tau rewrite this to be simpler

Or run :Tau without an instruction to open the instruction picker. While reviewing a streamed proposal:

Key	Action
<CR>	Accept edit
<Esc>	Reject edit
r	Regenerate edit

Tau automatically includes nearby lines from the active file. To add more files as context, press <C-t> in the instruction picker or run :TauContext.

In the context picker:

Key / Symbol	Meaning
<Space>	Toggle file under cursor
<Esc> / q	Close picker
◎	Active file, always included
●	Selected context file

Selected context files persist for the current Neovim session.

Configuration

require("tau").setup({
  api_url = "https://api.openai.com/v1",
  api_key = vim.env.OPENAI_API_KEY,
  model = "gpt-4o",

  -- optional
  context_lines = 30,
  timeout_ms = 60000,
  context_window = 128000,
  temperature = 0.2,
  max_tokens = 4096,
  top_p = 1,
  debug = false,
  keys = { context = "<C-t>" },
})

Option	Default	Description
api_url	required	OpenAI-compatible API base URL
api_key	required	API key; prefer environment variables
model	gpt-4o	Model name
context_lines	30	Lines above and below the selection to include
timeout_ms	60000	Request timeout in milliseconds
context_window	unset	Model context window, used for token estimates
temperature	unset	Sampling temperature, 0 to 2
max_tokens	unset	Maximum response tokens
top_p	unset	Nucleus sampling value, 0 to 1
debug	false	Write diagnostics to ~/.local/state/tau/diag.log
keys.context	<C-t>	Open context picker from the instruction picker

Commands: :Tau [instruction], :TauCancel, :TauContext.

License

MIT