Web search and reading pages

When memory has no answer or fresh information is needed, the assistant can search the internet and read the pages it finds. It decides by itself when this is needed. You can also ask directly: "search the web for…", "open the page…".

Turn it on or off

Every connection on the Models screen has a Web access mode:

• Off: this model has no internet access.
• Automatic (recommended): the model searches when it needs to.

Two kinds of search

The model's built-in search. The provider searches and reads the results itself. Claude, GPT, Gemini, Grok and Kimi have it, and through OpenRouter so do the models that support it. The provider charges for it separately.

Fallback search. If the model has no built-in search, MindDuck searches by itself through the search service you chose and passes the results to the model:

Connect a service on the Models → Web search tab: paste the key and click Test and save.

Reading pages

To read a page it found, MindDuck downloads it. The method is chosen in Assistant → Web page reader → Reading mode:

• Local only (default): pages are read on the device, directly or in the background browser. Page addresses and content never leave the device.
• Automatic: this device first, then Jina Reader: on the device first, and pages that can't be read here through the Jina Reader service. It helps with sites that refuse requests from your device.
• Always use Jina Reader: every address is sent to Jina Reader.

Background browser

Many sites build their pages with scripts: the downloaded code holds almost no text. MindDuck opens such pages in a background browser, a hidden window on this device. The address isn't sent anywhere, and the window never appears on screen or gets in your way. While a page opens, the assistant's progress line shows "Opening the page in the background browser…" and "Waiting for the page to finish loading…".

The background browser opens the page without your sign-ins and cookies, like a private window, and doesn't let it open new windows, download files, ask for the camera, microphone or location, or reach your computer and home network. Turn it off with Use a background browser for pages built by scripts in the same place. Without it, such pages are read through Jina Reader in Automatic mode and may not be readable in Local only mode.

Long pages

The assistant gets a long page in parts, with an outline: it sees which part holds which section and can open the part it needs or search the page for words. Tables stay tables, and fields, buttons and switches show as short lines such as [Button: Save]: on pages that are applications they carry most of the meaning. The page loads once per answer; moving between parts doesn't load it again.

Saving images from a page

Images on a page reach the assistant with their addresses, so it can give you a direct link to a photo or save it into its answer. A saved image appears under the answer; click it to open it full size and press Save to keep it as a file. The assistant downloads it on this device, from public addresses only and without your sign-ins: JPEG, PNG, WebP and GIF up to 15 MB, at most 10 images per answer. Saving works with any model; a model that understands images also sees what it saved.

Looking at a page

Models that understand images can look at a page as well as read it: the background browser takes a picture of one screen, and the model sees the page as a visitor does. This helps when the layout, images, charts or the state of an application matter, and when you ask it to "take a screenshot and show me".

Before the first picture in a chat, a card The assistant wants to look at this page appears with the address and buttons:

• Allow once: one picture; next time the assistant asks again.
• Allow for this chat: no more questions in this chat until the app restarts. This permission is never saved.
• Deny: the assistant continues without pictures and doesn't ask again in this answer.

The picture the model looked at appears as a small preview under its answer; click it to open it in full. It is exactly the picture the model received. It is kept with the chat and is not sent to models again with later messages. A picture costs more tokens than text, so the assistant looks only when it helps the answer.

Pictures work on Windows, macOS, Android and iOS; on Linux the assistant only reads pages for now.

The assistant opens pages of local servers, for example a program on http://localhost:3000, only from the Local servers the assistant may open list in the same place. Write one address per line, with its port, and click Save addresses. Only these exact addresses open. Such pages are read on the device and never go to Jina Reader, and the list itself is stored on this device and isn't synced. Add only servers you trust: a page the assistant reads could ask it to open an address from the list.

What is sent where

• With built-in search, the query goes to the model's provider.
• With fallback search, the search service receives the query the model wrote. It doesn't get your conversation.
• Jina Reader receives the page address in Always use Jina Reader mode, and in Automatic mode only when the page couldn't be read on the device. Addresses of local servers from the list are never sent to it.
• The background browser works on the device: the site gets an ordinary request from your device, as with local reading, but without your sign-ins and cookies.
• A picture of a page goes to the model's provider only after you allow it.