Disk space
Ask in plain words and get a command you can read before it runs. This one includes hidden folders and lists the biggest first.
Type what you need. Press your shortcut.
Review the command. Hit Enter when ready.
Works with Claude Code, Codex CLI or Gemini CLI
Describe what you need in plain language at your shell prompt.
Ctrl+G by default. Choose your key during installation.
The risk and a failure check show under the command. Read or edit it, then press Enter. Nothing runs automatically.
Six everyday requests, with what hey-shell actually returned.
Ask in plain words and get a command you can read before it runs. This one includes hidden folders and lists the biggest first.
Commands that change things are flagged. Killing processes is rated MEDIUM, so you see it before you press Enter.
Not quite right? Add a # note after the command and press your shortcut again. The command changes instead of starting over.
Run hsf right after a command fails. You get a corrected version on your next prompt, with what was wrong. Nothing is re-run.
gti should be git.Ask with hse to get the command with a short explanation of what it does and what to watch out for. The command is then left on your next prompt.
Got a command from somewhere? hsc looks for known reasons it would fail: syntax errors, missing programs, no git repo, no tmux server. It never runs the command. Here, from a shell with no tmux running.
Real output with the default agent (Claude Code, Haiku model). bash, zsh and fish work the same way, and so do Codex CLI and Gemini CLI; answers vary by agent and model. Nothing runs until you press Enter.
None of this is needed for the main flow. It is there when a command is not quite right, or fails.
Not quite right? Add # only .ts files after the command and press your shortcut again. It changes the command instead of starting over.
Run hsf right after a command fails. You get a corrected version on your next prompt, with what was wrong. Nothing is re-run.
hse <request> adds a short explanation, and hsw "<command>" explains a command you found somewhere, before you run it.
The agent is told which tools you have (rg, fd, jq…) and what the project is (pnpm, Makefile targets, compose). Names only, never file contents.
Deletes, force-pushes, overwrites, cloud deletes and download-and-run (curl … | /bin/bash) are flagged in amber or red. Anything else shows Risk: none recognized: no known pattern matched, which is not the same as safe.
Missing programs, syntax errors and wrong context (no git repo, no tmux server) show as Check: LIKELY FAIL. It never runs the command.
Nothing is logged unless you turn history on, and then only for you. Your folder is sent as ~/project, without your user name.
The same shortcut and commands in all three, with commands written for your shell: set -gx in fish, export in bash and zsh.
hey-shell asks the AI agent CLI you already use and are logged in to, and never runs what comes back. Agents can use their own tools while they answer, so here is exactly what hey-shell restricts for each:
Text only. No tools, no MCP servers, no saved session.
No writes. A read-only sandbox your Codex config cannot loosen. It may read files in the folder while answering.
Not restricted by hey-shell. Gemini CLI has no documented way to turn its tools off, so this relies on its own defaults.
Your terminal. Your shortcut.
~/hey-shell/install.sh --key alt+o
One line. It downloads the newest release and asks for your key and short aliases. Then open a new terminal.
Add bash -s -- --agent codex (or gemini) to use those instead of Claude Code.
The script is about 60 lines, or install by hand: git clone https://github.com/alparslanpolat/hey-shell ~/hey-shell && ~/hey-shell/install.sh
hse how to free up disk space for the command with an explanation, or hsc "<command>" to check one before you run it.
Generated commands are only placed on your prompt, never executed. Always read them before pressing Enter.