Divinci Local Inference — Privacy Policy

Last updated: June 2026

This policy applies specifically to the Divinci Local Inference Chrome extension. For Divinci AI’s website, apps, and services generally, see our main Privacy Policy.

Divinci Local Inference runs an open-weight AI model (Google’s Gemma 4) locally in your browser on your GPU, and — when you choose to sign in — connects that local assistant to your Divinci account for optional cloud-assisted features. This policy explains exactly what stays on your device and what, in specific situations, is sent to Divinci.

The short version: By default the extension is local-only — your chats with the on-device model never leave your computer. Some optional, clearly-controlled features (signing in, page-aware answers, and account-mode chat) do send data to Divinci. Those are described below. We do not sell your data, show ads, or use it to track you across the web.

1. What stays on your device (default)

• Your chats with the local Gemma model. Prompts and responses are computed on your GPU and are not logged, stored, or transmitted by the extension. (Exceptions: the two optional features in §3 and §4.)
• The model files, cached in your browser after first download.
• Your settings (selected model, inference defaults, privacy toggles), stored locally in your browser.

When you are not signed in, the extension sends no browsing information to Divinci.

2. Signing in to Divinci (optional)

If you click Sign in / Sign up, the extension completes a standard OAuth sign-in with Divinci’s identity provider (Auth0). On success we receive and store on your device an access token and your basic profile (email, name, and avatar URL) so the extension can show who you’re signed in as and make authenticated requests on your behalf. The access token never leaves the extension’s background service worker. You can sign out at any time from the toolbar popup, which deletes the stored tokens.

3. Web browsing activity (only while signed in and the panel is open)

To tell you whether the page you’re viewing is covered by Divinci’s shared public-web knowledge index and to ground answers in it, the extension — only when you are signed in and have the Divinci side panel open on a page — sends the following to Divinci’s API:

• The page’s address, reduced to its origin and path only. The query string and fragment (the parts after ? and #, which can contain search terms, tokens, or personal identifiers) are stripped before sending.
• A one-way fingerprint (hash) of the page’s visible text, used to detect whether our index is up to date. The page’s actual content is not sent — only this hash and the trimmed address.

Important limits:

• It happens only while the side panel is open on a page. With the panel closed, the extension sends nothing about the pages you visit.
• Sensitive sites are skipped entirely — the extension does not send anything for sign-in/account pages, banking and financial sites, webmail, healthcare portals, local/private addresses, or non-standard ports.
• It is used to look up and freshen the public-web index, not to build a profile of you or to target advertising.

The shared index itself is built by Divinci crawling publicly accessible web pages on its own servers; this extension does not upload page content to build it.

4. Page-aware answers & account-mode chat (optional)

• Page-aware answers (grounding). When a page is in the index and you send a message in the side panel, the extension sends your message and the trimmed page address to Divinci to retrieve relevant context, which is then given to the local model. So in this case your chat message does leave your device. You can turn this off — see §5.
• Account-mode chat. If you enable “Use my Divinci account” for chat, your conversation is sent to Divinci’s servers (to run server-hosted models and tools) and stored as a transcript on your account, the same as chatting on chat.divinci.app. Leaving this off keeps chat fully local.

5. Your privacy controls

In the popup under Advanced settings → Privacy:

• Retrieve Divinci page context — when off, the extension never sends your message for page-aware answers (your chat query stays on your device). Default: on.
• Allow Divinci to use my account chats — when off, the extension asks Divinci not to use your account-mode chats to improve its services. Default: on. (This sends an opt-out signal with your requests; the actual handling is enforced by Divinci’s servers.)

You can also stay signed out (fully local) or sign out at any time to stop all of §2–§4.

6. Where data goes

• huggingface.co (and CDN cas-bridge.xethub.hf.co) — to download the model files, subject to the Hugging Face Privacy Policy.
• Divinci’s identity provider (Auth0) — only during sign-in.
• Divinci’s API (api.divinci.app) — for the signed-in features in §3 and §4.

7. What we do not do

• We do not sell or rent your data.
• We do not show ads or use your data for advertising or cross-site tracking.
• We do not send the content of the pages you visit (only the trimmed address and a one-way hash, per §3).
• We do not transmit anything about your browsing when you are signed out or when the side panel is closed.

8. Permissions

• offscreen — run the WebGPU model.
• storage — store settings and the cached-model preference locally.
• identity — complete OAuth sign-in to your Divinci account (§2).
• host permissions (api.divinci.app and the Auth0 sign-in origin) — make the authenticated requests in §2–§4.
• content script on all sites — draw the side panel and, only while it’s open and you’re signed in, run the page-index check in §3. The script reads the page title, address, and visible text locally to compute the hash; it does not transmit page content.
• externally_connectable (Divinci AI domains only) — let chat.divinci.app use the local model via a chrome.runtime port.

9. Open source

The extension is Apache-2.0 licensed; source is available at github.com/Divinci-AI/gemma-gem.

10. Changes to this policy

If we change how the extension handles data, we will update this policy and bump the extension version (shown on the chrome://extensions card).

11. Contact

Questions? Email mike@divinci.ai.