nlsh — Natural Language Shell

Type plain English at your shell prompt. Get a shell command back, confirm it, and run it — all without leaving the terminal.

$ list all files modified today
⟳ thinking...
  find . -maxdepth 1 -newer $(date -d today +%Y-%m-%d) -type f
[y] run  [e] edit  [n] cancel

No cloud. No API keys. No mode switching. Runs entirely on-device — either via Apple Foundation Models or a local Ollama model.

Requirements

• macOS 26 or later
• Apple Intelligence (System Settings → Apple Intelligence), or
• Ollama — nlsh will offer to install it automatically during setup

Install

One-liner (build from source)

Unfortunately Requires Rust and Xcode Command Line Tools due to apple's HORRIBLE notarization systems.

git clone https://github.com/TheActualJacob/nlsh.git && cd nlsh/nlsh && cargo build --release && sudo ./target/release/nlsh --install

Then set as your default shell:

chsh -s /usr/local/bin/nlsh

One-liner (pre-built, when a release is available)

curl -fsSL https://github.com/TheActualJacob/nlsh/releases/latest/download/install.sh | sh

First launch

The first time you run nlsh, a setup screen appears:

  ╭────────────────────────────────────────────╮
  │         nlsh · model setup                 │
  ╰────────────────────────────────────────────╯

  Choose an AI backend:

  ▶ 1  Apple Intelligence   on-device, no download
    2  Ollama local model   ~986 MB download

  [↑↓ / 1-2] navigate   [Enter] confirm   [q] quit

• Apple Intelligence — uses the on-device model, no download needed
• Ollama — downloads qwen2.5-coder:1.5b (~986 MB); nlsh installs Ollama automatically via Homebrew if it isn't already present

At the end of setup you'll be asked whether to set nlsh as your default shell.

Run nlsh --setup at any time to switch backends or re-configure.

Usage

From here, use it exactly like zsh. Shell commands run normally. When you type something that isn't a shell command, nlsh routes it to the model:

$ show disk usage for each folder here, sorted largest first
⟳ thinking...
  du -sh */ | sort -rh
[y] run  [e] edit  [n] cancel > y

# command runs...

Confirmation prompt

Key	Action
y or Enter	Run the command
e	Open the command in $EDITOR to modify before running
n or Esc	Cancel

Flags

Flag	Description
--setup	Re-run model selection / backend configuration
--dry-run	Print the generated command but don't run it
--no-hist	Prefix generated commands with a space (hides them from history if HIST_IGNORE_SPACE is set in .zshrc)

How it works

nlsh wraps zsh in a pseudoterminal. Every line you type is classified before being sent to the shell:

• Recognized command (exists in $PATH, is a builtin, alias, or function) → forwarded directly to zsh.
• Unrecognized input → sent to the configured model for translation into a shell command.

When a TUI application starts (vim, less, htop, ssh), nlsh detects the alternate screen and switches to full passthrough mode — every keystroke goes straight to the pty with zero interception.

Troubleshooting

"Apple Intelligence unavailable — NL routing disabled"

• macOS 26+ required
• Apple Intelligence must be enabled: System Settings → Apple Intelligence & Siri
• Your device must be eligible (M-series Mac or A17 Pro iPhone/iPad and later)
• Switch to Ollama with nlsh --setup if Apple Intelligence isn't available

"Ollama unavailable — NL routing disabled"

Ollama isn't running. Start it with ollama serve, then open a new nlsh session. Or switch backends with nlsh --setup.

Generated commands aren't going into shell history

Pass --no-hist and add setopt HIST_IGNORE_SPACE to your ~/.zshrc.

Uninstall

sudo rm /usr/local/bin/nlsh /usr/local/bin/nlsh-model
# Remove /usr/local/bin/nlsh from /etc/shells
chsh -s /bin/zsh   # if you set nlsh as your default shell

License

MIT