Architecture

Privacy isn't a policy page. It's an architecture.

There is no RefineIQ server. By default, prompt refining runs entirely in your browser — so there is nothing to phone home to, breach, or quietly re-policy later.

01

You select text

In ChatGPT, Claude, Gemini, or any text box — the Refine pill appears where you work.

02

Inference stays local

SmolLM2 runs in your browser via WebGPU (WASM fallback). Your draft never hits a RefineIQ API.

03

You review & replace

Plain-text rewrite under 500 characters. Esc restores the original. Offline after the one-time download.

What “on-device” means here

  • Model download is consented and one-time. Default ~380 MB, or a lighter ~130 MB tier. Cached locally; then RefineIQ works offline.
  • WebGPU first, WASM fallback. Modern Chromium gets the fast path; older setups degrade gracefully instead of failing closed.
  • Voice Draft stays local too. Speech-to-text uses an on-device model — RefineIQ deliberately avoids Chrome's cloud speech API.
  • Optional BYOK is still middleman-free. Your OpenRouter or Hugging Face key talks HTTPS to that provider directly. There is no RefineIQ proxy.

What we ask Chrome for — and why

Permissions are justified host-by-host in the privacy policy. Short version: storage for the model and settings, optional microphone for Voice Draft, and site access so the Refine pill can appear where you type.

Verify it yourself

In on-device mode, Chrome DevTools → Network should show no RefineIQ backend traffic for refine requests — because there is no backend. That's the product, not a slogan.

Go deeper

Free. Private. On-device.

Add to Chrome — 0 servers