gutcheck

grep for meaning. Ask a plain-English question about every line, paragraph, JSON record or diff hunk, then explore the answers in a live terminal view. One binary, runs on your CPU. No API key, no bill, nothing leaves your machine.

curl -fsSL https://raw.githubusercontent.com/sfmqrb/gutcheck/main/install.sh | sh   # Linux x86_64, macOS arm64
brew install sfmqrb/tap/gutcheck                                                    # Homebrew
cargo install --git https://github.com/sfmqrb/gutcheck                              # from source

The first run downloads a 1.3 GB model to $XDG_CACHE_HOME/gutcheck.

Use

gutcheck -n "is this a bug report?" issues.txt             # keep the matching lines, grep-style
gutcheck -i "does this log line describe an error?" app.log   # explore live: threshold, sort, re-ask
gutcheck @secret -r src/                                   # named questions: hunt credentials in a repo
gutcheck --top 5 -f "..." app.log                          # the 5 most likely, ranked, with heat bars
gutcheck -c bug,billing,question --tally "what is this about?" support.txt   # a labelled histogram
git diff | gutcheck --diff "removes a permission or security check"          # judge each hunk
gutcheck --field message "is the user frustrated?" events.jsonl              # JSON field (or --csv)
tail -f app.log | gutcheck -v "is this routine noise?"                      # streams; -v inverts

grep habits carry over: -n -H -l -q -m -v -r -C, several questions with -e (--all for "every one"), exit codes 0 (match), 1 (none), 2 (error). Pipes get plain text; a terminal gets color. Every flag: command reference.

Made to be explored

• -i, a live view. Scores appear as they arrive, best first. Slide the threshold with the arrow keys, press / to ask a new question, enter to print the matches.
• --auto. Finds the cut-off from the scores themselves, so you don't guess a number.
• Named questions. @secret, @pii, @security, @error, @bug, @spam... or your own in ~/.config/gutcheck/questions.
• Heat bars, --rank, --tally. Answers you can read at a glance.
• A receipt on every run. 72 records · 8 model calls (64 reused) · 4.4 s: what the run cost and what it saved.
• Three models. -M multilingual (100+ languages, fastest), english, and typed, the most accurate on English (named questions use it). Models and accuracy.

Built for logs

Identical lines are answered once. With -f, lines that differ only in numbers, timestamps and ids share one answer, so a log with a few dozen line shapes costs a few dozen model calls. On real logs (Loghub samples):

gutcheck --estimate "..." big.log shows the number of model calls before you spend them. Method: benchmarks. Getting the best answers: tips.

More

Command reference · Models and accuracy · Benchmarks · Tips · How it works and credits

gutcheck is an independent project, not affiliated with the Laya or Jev authors. MIT licensed; the Apache-2.0 models are downloaded at runtime.