What it does

Butterfish wraps your existing shell, bash or zsh, on macOS or Linux. You use the shell as normal, and a command that starts with a capital letter is sent as a prompt to an OpenAI model. The model can see your recent shell history, so you can ask things like why a command failed without copying and pasting. It also offers command autosuggestions that you accept with Tab, and a Goal Mode. Starting a command with an exclamation mark lets the model propose commands toward a goal, which you run or edit; two exclamation marks run them without confirmation. The author says Goal Mode is hit or miss and works best on simple goals. The prompt wrappers are visible and editable in a local file, and a verbose mode shows the raw requests sent to the API. The default model is gpt-4-turbo, you can choose another, or you can point Butterfish at any server with an OpenAI-compatible API, such as a local model. Further commands cover one-off prompting, command generation, file summaries and embeddings of local files, and a Neovim plugin is available separately. Butterfish is open source under the MIT licence and you supply your own OpenAI API key. The page shows version 0.1.7, built in August 2023, and calls the tool experimental.

Capabilities

Wraps bash or zsh and keeps your shell history as context
Capital-letter prompts to ask an AI model from the command line
Command autosuggest accepted with Tab
Goal Mode agent that proposes commands toward a goal
Visible, editable prompt wrappers
Works with OpenAI-compatible servers, including local models

When to use it

You work mostly in a terminal and want AI help without leaving it
You want to ask why a command failed using the real shell output
You want to inspect and control the prompts sent to the model