Show HN: Mu – an agent with actual command line experience

Mu, a new shell plugin for zsh and fish, lets users press Tab to enter agent mode, where commands are sent to an LLM instead of being interpreted as shell script, with executed commands appearing in zsh_history. The plugin, which provides a single bash tool and special commands for file editing and multi-modal input, aims to offer a seamless switch between normal shell work and agentic work for command line users.

I'm an old school user that finds the AI agent TUIs too magical, so I experimented with a new agent UX. I implement mu as a shell plugin zsh and fish for now . Pressing Tab enters agent mode, where 1. the command line prompt is changed to displays model name and context usage, and 2. a preexec hook sends commands to an agent instead of interpreting them as shell script. Thats it. Each LLM prompt starts a regular process that reads stdin and writes stdout, with no terminal magics. The actually executed commands appear literally in zsh history, just like other shell commands. This allows a seamless switch between normal shell work and agentic work, with mixed history that uses regular scrollback buffer. I also experimented with a minimum agent design, where only a single bash tool is provided. Special commands are provided for reliable file editing and multi-modal input. So far it worked well for me, as a day to day command line helper. Comments URL: https://news.ycombinator.com/item?id=49550418 https://news.ycombinator.com/item?id=49550418 Points: 1 Comments: 0