All builds

Developer tool

whatcmd

Type plain English at your shell prompt and get the right command for your OS and tools. A local model, no API keys, and nothing runs until you press Enter.

median end to end, on an M4 Pro
217 ms
overall on my 500-case benchmark
92.7%
of non-shell requests refused
98.8%

Replay of whatcmd in a terminal: it is installed with npm and set up with whatcmd init, then plain-English requests after /w become shell commands on the prompt, such as finding files bigger than 100MB, refining them with /w+, and finding what is using port 8080. A joke is refused, and a risky rm -rf gets a warning.

Type /w followed by what you want, such as /w find files bigger than 100MB, and the line is replaced with the right command for your machine: find . -type f -size +100M. It knows your OS, shell, GNU or BSD flags and installed tools, and it never runs anything until you press Enter.

What it does

  1. Reads the man page or help text of the tools you name, so flags are real and not guessed.

  2. Refine the last command instead of retyping it, and "fix that" knows the exit status.

  3. Warns before damage, such as recursive deletes, dd, piping curl to a shell and force pushes.

  4. Looks things up on the web only when the local answer fails, through tiny-browse.

How it is built

  • An output gate rejects non-shell text, repeated requests and copied examples, with targeted retry hints for missing programs and GNU-only flags on macOS.
  • Grounding in local docs lifts made-up-tool cases from 0% to 85%.
  • 100% repeatable on 50 cases run three times.
  • CI, CodeQL, dependency review, secret scanning and a tag-driven release that ships the tested tarball.

Honest limits

  • Small models still pick wrong flags sometimes, so read before you press Enter.
  • Benchmarked on macOS only, on one machine, with cases I wrote by hand.
  • No fish or Windows shortcut yet.