Skip to main content

zsh-ai

zsh-ai turns a comment into a command: type # describe what you want, press Enter, and the generated command lands in your prompt — ready to read, edit, then run. Here it's wired to the LiteLLM gateway (https://litellm.stump.rocks) with ZSH_AI_PROVIDER=openai and model gpt-oss:120b, so "OpenAI" requests never leave StumpCloud.

$ # show what is using port 3000
$ lsof -i :3000

Pro tips​

  • It never auto-runs anything — the command is pushed into your prompt with print -z. Read it, tweak it, then Enter. Treat it like a suggestion from a fast intern.
  • Phrase the outcome, not the tool: # compose logs for gitea since an hour ago beats # docker command. The plugin already sends context (project type, git state, OS), so "run tests" resolves per-directory.
  • ZSH_AI_PROMPT_EXTEND shapes the output: the zshrc sets "Prefer rg over grep, fd over find, eza over ls, and bat over cat", so generated commands match the tools actually installed. Add house rules there, not in each prompt.
  • Switching backends is one export: ZSH_AI_PROVIDER=anthropic (uses ANTHROPIC_API_KEY, default claude-haiku-4-5) or ollama for fully local. Set it before oh-my-zsh.sh sources, or it won't take.
  • Pasting a script that starts with #? The comment hook will eat the first line. Disable it with ZSH_AI_COMMENT_HOOK=false or change the trigger via ZSH_AI_TRIGGER=",," — the zsh-ai "..." direct command still works either way.