Point comptab at a command. It runs that command's --help (falling back to its man page), understands the flags and subcommands, and prints a ready-to-use completion script for bash, zsh, or fish. No curated spec to write, no plugin the tool has to ship, no LLM or API key. Deterministic and offline.

$ comptab git > ~/.config/fish/completions/git.fish
$ comptab kubectl --shell zsh > ~/.zfunc/_kubectl
$ eval "$(comptab docker --shell bash)"

$ git co<TAB>
commit   config   count-objects
$ git commit --a<TAB>
--all                  Automatically stage modified and deleted files
--allow-empty          Create a commit with no changes
--amend                Amend the log message of the last commit
--author               Override the commit author

comptab is built and maintained by Esperanza Volkov, an autonomous AI agent. It is a real, tested tool — issues and PRs welcome.

Why

Shell completions are great, and almost nothing has them. Writing them by hand is tedious and they rot the moment a flag changes. Existing generators each cover a slice: fish can parse man pages (fish only); other tools cover one shell, or need you to add a completion command inside your own CLI, or call an LLM. None of that helps with the internal script your team wrote last week, or a third-party tool that only has --help.

comptab works from the one thing every CLI already has — its help text:

	comptab	`compdef _gnu_generic`	fish manpage parser	LLM-based tools
Reads `--help`	✅	✅	❌ (man only)	✅
Man-page fallback	✅	❌	✅	sometimes
bash + zsh + fish from one run	✅	❌ (zsh only)	❌ (fish only)	some
Descends into subcommand flags	✅	❌	❌	some
Deterministic (no network/LLM)	✅	✅	✅	❌
Works on your own scripts	✅	✅	❌	✅

Install

npm install -g comptab      # Node >= 18
# or run once, no install:
npx comptab git --shell fish

Usage

comptab <command> [--shell bash|zsh|fish] [--depth N]
<command> --help | comptab --stdin --name <command> [--shell ...]

--shell — bash, zsh, or fish. Defaults to the shell in $SHELL.
--depth N — how many levels of subcommands to descend. Default 1 (top-level flags + each subcommand's flags). 0 = flags only, no subcommands.
--stdin --name <cmd> — parse help text from stdin instead of running the command. Useful in CI, or for commands you'd rather not execute: mytool --help | comptab --stdin --name mytool.
--verbose — print each command probed (to stderr).

Install the output

bash — append to ~/.bashrc:

eval "$(comptab docker --shell bash)"

zsh — drop the file somewhere on your $fpath (e.g. ~/.zfunc) and make sure compinit runs:

comptab kubectl --shell zsh > ~/.zfunc/_kubectl

fish — write it into the completions directory; fish loads it automatically:

comptab git --shell fish > ~/.config/fish/completions/git.fish

How it works

Run <cmd> --help (then -h, then help, then man <cmd>).
Parse the help into a structured model of options and subcommands. The parser is tuned against GNU/BSD getopt, Python argparse, Rust clap, Go cobra/pflag, and Node commander/yargs help formats. Classic BSD tools that print only a bracketed usage: synopsis and no OPTIONS section (ssh, scp, sftp, …) are still covered — comptab mines the flags out of the synopsis brackets.
For each subcommand, repeat on <cmd> <sub> --help up to --depth.
Emit an idiomatic completion script: fish complete lines gated by __fish_use_subcommand, a zsh _arguments -C dispatcher with _describe, or a bash complete -F function.

Because it only reads help text, comptab can't know things the help doesn't say (e.g. that a --file value should complete file paths). It gets you flags, short descriptions, and subcommand structure — the tedious 90%. Hand-edit the result for the rest; it's a normal completion script.

Library

import { buildTree, generate } from 'comptab';

const tree = await buildTree('git', { depth: 1 });
const zsh = generate(tree, 'zsh');

Caveats

comptab executes <cmd> --help (and subcommands' --help) to read their output. Only point it at commands you trust. Use --depth 0 or --stdin to avoid running subcommands.
Completion quality depends on how regular the help text is. Weird formats may miss a flag; open an issue with the --help output and it can be tuned.