Zsh Ai Cmd

zsh-ai-cmd

Natural language to shell commands with ghost text preview.


Install

Requires `curl`, `jq`, and an API key for your chosen provider (or a Claude subscription for the `claude-code` provider).

# Clone
git clone https://github.com/kylesnowschwartz/zsh-ai-cmd ~/.zsh-ai-cmd

# Add to .zshrc
source ~/.zsh-ai-cmd/zsh-ai-cmd.plugin.zsh

# Choose your provider (default: anthropic)
export ZSH_AI_CMD_PROVIDER='anthropic'  # or: openai, gemini, deepseek, ollama, copilot, claude-code

# Set API key for your chosen provider
export ANTHROPIC_API_KEY='sk-ant-...'
export OPENAI_API_KEY='sk-...'
export GEMINI_API_KEY='...'
export DEEPSEEK_API_KEY='sk-...'
# Ollama and Copilot need no key (local services)
# Claude Code uses your existing Claude subscription (requires: npm install -g @anthropic-ai/claude-code && claude login)

# Or use macOS Keychain
security add-generic-password -s 'anthropic-api-key' -a "$USER" -w 'sk-ant-...'

Usage

1. Type a natural language description
2. Press Ctrl+Z to request a suggestion
3. Ghost text appears showing the command: find large files → find . -size +100M
4. Press Tab or → to accept, or keep typing to dismiss

When the model sees more than one reasonable approach, the ghost text shows a cycle counter (`⟲ 1/3`) — press `Ctrl+Z` again to cycle through the alternatives, then `Tab` to accept the one showing.

Destructive suggestions (`rm`, `kill`, `sed -i`, history rewrites, …) are tinted red and marked with `⚠` so you notice before accepting. Providers without structured output (`copilot`, `claude-code`) return a single suggestion and can't flag destructive commands.

Configuration

ZSH_AI_CMD_PROVIDER='anthropic'              # Provider: anthropic, openai, gemini, deepseek, ollama, copilot, claude-code
ZSH_AI_CMD_KEY='^z'                          # Trigger key (default: Ctrl+Z)
ZSH_AI_CMD_HIGHLIGHT='fg=8'                  # Ghost text style (zsh region_highlight format)
ZS