Why local-first matters
Many paraphrase services ask writers to paste drafts into a remote form and pay for stronger limits. RhetoriLex takes a different route. The public app is static, installable, and free. It can build a safe prompt without any endpoint, or call only the OpenAI-compatible endpoint that the user enters.
This keeps the default tool useful for researchers, students, supervisors, editors, and Indonesian writers who need careful English without turning every sentence into a paid request. The app does not store manuscripts on a RhetoriLex server because the current web version has no server-side manuscript database.
Safe setup choices
| Choice | How it works | Privacy check |
|---|---|---|
| Prompt only | RhetoriLex builds the instruction and invariant checklist. | No AI call leaves the browser. |
| Local endpoint | Use an OpenAI-compatible local server such as a local model gateway. | Keep endpoint on your own machine or network. |
| Provider endpoint | Enter your own endpoint, model, and token when your provider permits browser calls. | Your text goes to that provider, not to RhetoriLex. |
Why the hosted web limit is local
The public page applies a three-call daily guard in browser storage. This is a courtesy limit, not account security. It prevents accidental overuse while keeping the static site free and deployable on GitHub Pages.
A true shared-key hosted service needs a backend. The planned design is a small Worker API, a D1 table for anonymous daily quota, a KV cache for rate-limit decisions, Turnstile for abuse resistance, provider tokens stored only as server secrets, and manuscript logging disabled by default. That backend is intentionally separate from the current static app.