Ducky provides the most ergonomic LLM chat interface for developers. With a minimal UX inspired by text messaging apps, ducky lets you access most powerful LLMs from within the terminal.
git clone https://github.com/gregriff/ducky.gitcd duckygo install .
Note: this installation process is required because I use a custom fork of
termenv.
ducky run [model name]
All latest OpenAI and Anthropic models are supported, given an API Key
Run
ducky --helpto see all flags and options
Edit the $XDG_CONFIG_HOME/ducky/ducky.toml that was created for you.
- Markdown rendering of responses (can customize colors and more)
- Syntax highlighting of code blocks (configurable, and per-language highlighting coming soon)
- Responsive resizing of all elements on screen during terminal window resizing, even during response streaming
- Intelligent resizing of prompt input to maximize main content area
- Graceful handling of API errors
- Why the terminal?
I like to juggle several chats at once, and I'd rather use a terminal multiplexer instead of having several LLM browser tabs open. Also, all IDEs have a terminal, so any developer can easily incorporate this tool into their existing workflow. In addition, LLM's are currently accessed most easily via browsers or native apps. On older linux hardware like Raspberry Pi's, these GUIs either don't exist or run slowly. A lightweight terminal-based LLM client allows even these machines to access LLMs.
- Does this support agentic workflows?
No. I'll let the IDEs and model providers handle that. This is a simple prompt/response loop.
- Why not use the IDE's LLM interface?
I believe LLMs work best in a seperate window when writing software. This makes the programmer think more about their prompts and discourages vibe coding.
- How to select text and copy to clipboard?
tldr; For now, you are only able to select and copy text that is visible on screen, using either the terminal emulator or multiplexer's dedicated text-selection keybind.
Unfourtunately, this is made tricky by the nature of the terminal. Since ducky is a uses the terminal's fullscreen mode for a polished feel, it captures all mouse/scroll input from it. While terminals usually have a keybind for overriding this (on alacritty its
shift+left clickto begin highlighting text, on others itsfn+left click), these solutions only allow copying text visible on the screen. Tmux provides the same behavior with itscopy-modeby pressingleader+[. So, if while selecting text with either of these methods, you drag the mouse to the top of the screen in order to scroll the text up and continue copying it, the screen will not scroll up, because the mouse and scroll inputs are being temporarily handled by the terminal emulator or multiplexer. A solution to this is for me to implement a text selection/copying feature myself, which is planned.
Glamour is the package responsible for rendering Markdown. It can be configured with different styles, which are JSON files that define mappings of strings to Markdown tokens, as well as colors, spacing options, and "code themes", which determine how code-block Markdown is rendered. These code themes can be changed with the code_block.theme property and can include any theme from this list
To create your own style, copy a file from glamour's built-in styles and modify it to change colors or other properties. Of note is the top-level document.block_prefix property, which if not set to "", will result in spacing inconsistencies when a response stream completes. Then, place the JSON file in $XDG_CONFIG_HOME/ducky/styles, and refer to it in the CLI args or the .toml config file by absolute path.
Development is done on macOS, using alacritty and tmux, or Zed's terminal. Testing will be done on fedora and raspberry pi OS soon.
This project is a rewrite of the now-unmaintained GPT-CLI Python repository. I'm doing this rewrite to learn Go and get experience with its concurrency patterns, standard library, package ecosystem, and idioms.