salt

How much do you swear at your coding agents?

salt reads the session logs that Claude Code, Codex, and Cursor already keep on your machine, counts the profanity in the prompts you typed, and publishes the counts to a link you can open and share. A sodium panel for your terminal.

npx salt-ai
bunx salt-ai

Nothing lands on your PATH or in your project. The runner caches the download, and that is the only trace it leaves.

The scan is local; the report is not. Publishing is the default and the only way to see the dashboard. The tool prints what it is about to upload before it does, and --json keeps everything on your machine.

Options

| Option | Effect | | ---------------------- | ------------------------------------------------ | | --no-open | Do not launch a browser | | --harness <list> | Limit to claude, codex, cursor | | --since <date\|span> | Only count prompts since 2026-01-01 or 30d | | --json | Print the report to stdout instead of publishing | | --lexicon <path> | Use an alternate lexicon file |

What gets published

The whole report goes up, repository names included. Anyone with the link can read it, and the link expires 30 days after it is created.

The report holds counts and matched swear words, never the text of your prompts. Project names are cut down to the repository's directory name, so no absolute paths leave your machine, and your home directory is never counted as a project.

For a fully offline run, --json scans, prints the report to stdout, and exits without making a single network request. Same numbers the dashboard would show, minus the dashboard.

Accuracy

Counting profanity in developer prompts is mostly a filtering problem. Two things ruin the count if you let them.

Only what you typed counts. Every harness routes machine-generated text through the same channel as real prompts, so salt drops tool results, system reminders, slash-command envelopes, sub-agent delegations, approval-assessment transcripts, and automation heartbeats. Harnesses also rewrite a session's entire history into a new file on every fork and resume. One prompt appeared 405 times in the corpus this was built against. Replayed prompts collapse to their first occurrence.

Code never counts. salt strips fenced blocks, inline backticks, and file paths before matching, so assert, class, pass, and analysis cannot register as swears. Matching is word-bounded against an allowlist on top of that, and f*ck folds onto its canonical spelling, so censoring yourself changes nothing.

The report's methodology section lists what was actually read, including what could not be. No grain of salt required.

Custom lexicon

Season to taste. Create ~/.config/salt/lexicon.toml:

remove = ["hell"]          # stop counting these
allow  = ["cassowary"]     # never match inside these words

[add]
blast = "mild"             # word = mild | medium | strong | acronym

Where it reads from

| Harness | Location | | ----------- | ----------------------------------------------------------------------- | | Claude Code | ~/.claude/projects/**/*.jsonl | | Codex | ~/.codex/sessions/**/*.jsonl, ~/.codex/archived_sessions/**/*.jsonl | | Cursor | ~/.cursor/chats/**/store.db (opened read-only, immutable) |

Development

pnpm workspace, three packages:

| package | what it is | | --------------- | ------------------------------------------------------------------ | | packages/core | the report schema shared by the CLI, the Worker, and the dashboard | | packages/cli | the scanner/matcher/publisher, tested and compiled with Bun | | packages/site | the landing page, the report dashboard, and the Cloudflare Worker |

pnpm install
pnpm -r typecheck
cd packages/cli  && bun test          # scanner/matcher/report/publish suites
cd packages/site && pnpm test         # worker contract tests in real workerd
cd packages/site && pnpm dev          # landing page + dashboard + worker locally
cd packages/cli  && bun src/main.ts --json   # run the CLI from source

Point the CLI at a local worker with SALT_HOST=http://localhost:5173.