Settings

In the editor, press the gear at the foot of the rail on the right, or ⌘ + , on a Mac (Ctrl + , on Windows and Linux). The settings open in a dialog over the capture; Esc, a click beside it or the X in its corner closes it. The same settings open on a page of their own from the gear in the panel that opens when you click the toolbar icon, from Options when you right-click the toolbar icon, and from chrome://extensions, then Details, then Extension options. On that page the X closes the tab, and Esc does nothing, so a stray key press does not close it.

The dialog lists its sections down the left: Appearance, Capture, Editor, Notifications, Sensitive Text, AI Model, Privacy and Local Agents. Each setting is one row. The question mark beside a setting's name says what it does when you point at it. A choice is saved the moment you make it; a box you type in is saved when you click away from it.

Page width is not here. It is the one choice that changes with the page in front of you, so it lives in the panel that opens when you click the toolbar icon; Getting started covers it.

Appearance

Theme

The editor, the popup and the settings have a light and a dark theme: the sun or the moon. The sun-and-moon button at the foot of the editor's rail switches the same setting straight to the other one, and a fresh install starts light. A captured page keeps its own colors; Page color scheme, under Capture, decides those.

Canvas dots and canvas color

The canvas is the surface behind the capture in the editor, and behind these settings on their own page. Canvas dots, on by default, draws a grid of dots on it that moves with the page as you pan and zoom. Canvas color picks one of four grays: White, Soft gray (the default), Gray or Deep gray in the light theme, and Black, Soft gray (the default), Gray or Light gray in the dark one. Each theme keeps its own choice, so switch the theme to set the other's.

Capture

These are set before you scan. The capture lays the page out once and takes the preview and the PDF from that one layout, so none of them can be changed for a capture you already have.

Page color scheme

A page with a dark version and a light one is captured in whichever this says, and the choice reaches the PDF as well as the preview.

• Match the browser (the monitor), the default, captures the page the way it looks right now.
• Light (the sun) takes the light version whatever the browser is set to.
• Dark (the moon) takes the dark one.

Leaving it on the default is also what keeps the PDF matching what you saw on screen. Chrome prints every page in the light scheme unless it is told otherwise, so a dark page used to come back as a dark preview and a white PDF. The capture tells it, whichever way the page writes its two versions: with a prefers-color-scheme media query, or with light-dark() colors under color-scheme: light dark, the way MDN does, which a print would otherwise render light even when told.

Light and Dark also reach a page that keeps the theme you picked on the site, the way many sites built with Tailwind or next-themes do (ui.shadcn.com is one): it marks the theme with a dark class or a data-theme on the page, which the capture turns to the one you asked for and back after. Match the browser leaves that mark as it is, so the page comes out as you see it.

A page that has only one version looks the same whichever you pick, and so does a page that chooses its theme in JavaScript some other way: its own script has already run by the time the capture starts.

Videos

A video on the page can only ever come out as a still image, and this decides what that still is.

• Paused frame, the default, pauses the video and keeps the frame it stopped on, at its on-screen size.
• Blank leaves the video's area empty.

Open folded sections before capture

A capture takes the page as it stands, so a section that is closed on screen is closed in the PDF too. This switch decides whether the scan opens the folded parts first.

• Off, the default, captures the page exactly as you see it.
• On opens every closed section, every accordion and disclosure that opens in place, and every hidden tab of a tab set (shown one below the other), then waits a few seconds for images that are still loading. After the capture it folds everything back up, so the tab is left the way you had it. A scan takes a moment longer.

It never presses a plain "Show more" or "Load more" button, a link, or anything in a menu or the site's navigation. Those can load the next page of results, send a form or count a click, so a section behind one is still left closed. Nor does it open a disclosure that does not mark itself as one to the browser's accessibility tree; open that by hand before you scan.

Check accessibility

Off by default. On, each page you scan is also checked against WCAG 2.0, 2.1 and 2.2, levels A and AA, with axe-core, on this computer: nothing about the page is sent anywhere. The editor then has an Accessibility panel on its rail that lists what failed, rule by rule, and boxes each failing element on the page while it is open. The panel can save the capture with the boxes drawn on and the rules listed after it, as a PDF to pass on.

An automated check finds some of what fails WCAG, not all of it: whether a picture's text describes it, or the page reads in a sensible order, still needs a person. A scan takes a moment longer with it on. It checks the scans that follow; a capture already open is not checked again.

Clean up before capture

A capture keeps whatever the page drew: an ad between two paragraphs, a cookie banner pinned to the bottom, a chat bubble, a newsletter pop-up and the dark layer behind it. Tick what the scan should hide first:

• Ads
• Cookie and consent banners
• Chat widgets
• Pop-ups and overlays: a pop-up window, the dark layer behind it, and a bar asking you to open the page in an app

All four are on by default. Untick them all for a capture you keep as a record of how the page looked, which should keep everything. What is hidden is taken out for the length of the scan and the page is laid out again without it, so no empty box is left behind, in the PDF or the preview. After the capture it all comes back, and the tab is left the way you had it.

Ads are found by the services that serve them and by the markup that ad tags write, and a frame from another site at one of the standard ad sizes counts as one; never by a word in a class name. Banners and chat widgets are found by the elements the common consent and chat services put on the page, and by a fixed bar about cookies with a button in it. A pop-up is found by being marked as one, or by covering at least half of the window above the page.

When the scan hid something, the editor says what as it opens, in a notification with a Capture settings button that brings you straight back to these four. An integrity record counts them as well. Turn all four off when you want a record of how the page looked, since a hidden element is not in the file.

The rules will miss some: an ad served from the site's own address, a banner no service built, a pop-up that is neither marked nor large. An ad blocker in your browser already keeps ads off the page, and the capture takes the page as it stands.

Maximum page height

How far down a capture goes before it stops. A page with no bottom, an endless feed or a thread that keeps loading, has to stop somewhere, and a capture has always stopped at 60,000 pixels.

• 60,000 px, the default, is where a capture has always stopped.
• 120,000 px and 240,000 px go further, for a long thread, a long document or a feed.

Past 60,000 pixels the PDF is printed in pieces and joined back into one, so it keeps its text and its sharpness. The preview in the editor, and a PNG or JPEG made from it, is one picture of the whole page, and the taller the page the coarser it gets. A single long page export of more than 60,000 pixels comes out as several long pages. The width stays limited to 60,000 pixels.

Some pages draw only what is on screen and leave the rest as empty space until you scroll to it; a long ChatGPT conversation is one. When a scan finds a page like that cut short, it goes past this setting for that page, up to 10 million pixels, so the messages below the limit are not left blank. The setting itself stays as you chose it. Past a million pixels or so the editor's preview of such a page is too small to read, while the PDF keeps every message sharp and searchable.

Editor

Crop when the editor opens

Where the crop is when the editor opens.

• Whole page, the default, opens every capture uncropped.
• Main content opens on the page's article or column of text when the scan found one, and on the whole page otherwise.

The two carry the same icons as the two crop buttons in the editor's toolbar, which switch between them on any capture, and the crop can be dragged anywhere.

Remembered choices

Remember the last choices is off at first, and every capture then opens on the defaults; a look you want again is a preset, saved from the editor's Presets panel. Turn it on and the editor remembers the choices you made on the last capture: the pen's color and width, the text's size and face, the blur's strength and the redaction's color, the format, the paper, the page breaks, the frame, the stamp, the brand and the watermark, and a JPEG's quality. The next capture opens on the same ones, so the second scan of the day starts where the first one ended. Turned off again, what it remembered is kept for when it is turned back on. Reset to defaults puts them back the way they were installed, in the next capture and in an editor that is already open, the one you opened the settings from included. The brand's logo is kept on its own, and the reset leaves it; remove it from the editor's Stamp & Brand panel.

Tour of the editor

The first editor that opens after you install Page Scanner walks through it in a few steps: the crop, the tools, Output Format, Extract, the Assistant and these settings. Show the tour shows it again. From the editor it starts at once; from the extension's options page, the next capture opens with it. Next and Back (or the arrow keys) step through it, and Skip tour or Esc ends it.

Notifications

The notifications that pop up at the bottom of the editor. Errors and warnings always show, because each one says something went wrong or was left out of the file.

• What was hidden before capture: the notification that says what Clean up before capture hid. On by default.
• When something is done: that a file was downloaded, a picture copied, or hiding marks placed from Sensitive Text. On by default. Off, an export finishes without a word.
• How long a notification stays: 4 seconds (the default), 8, 15, or until you close it. A notification can always be closed sooner with its ×, and one for work still going, such as an export being built, stays until the work is done.

Sensitive Text

Always look for

Words and patterns for the editor's Sensitive Text panel to look for on every capture, with the emails, numbers and keys its Pattern Scan looks for. One per line: a word, found whatever its case, or /a regular expression/, such as a customer's name or the shape of your account numbers. They stay in this browser. One saved while the editor's Sensitive Text panel is showing behind the dialog is looked for there at once, once Pattern Scan has run.

AI Model

Which model the editor's AI features ask: AI Scan in the Sensitive Text panel, and a Markdown or Word export Tidied by a model. Each way the text stays on your computer. The Assistant and the side panel are the exception: they use Page Scanner's server, below, and not the model chosen here.

Model Server

One list holds every choice: Browser Built-in, WebGPU, and each server you can run yourself, marked (Local).

Browser Built-in, the default, is Chrome's own on-device model. It runs inside Chrome where your computer meets Chrome's requirements for it, and Chrome downloads it the first time a scan asks. The row says where it stands.

A (Local) server is a model you run yourself on this computer, with a program such as Ollama, LM Studio or llama.cpp: Ollama, LM Studio, llama.cpp, vLLM, Jan, LocalAI, KoboldCpp, MLX LM, or OpenAI-compatible for any server that answers the OpenAI chat API at /v1. "Local" because some of them also run as a service on the internet, and Page Scanner reaches only the one on your computer. Useful where Chrome has no model for your computer, in Edge, or when you want a model of your choice. The first one you choose asks Chrome, once, to let Page Scanner reach localhost; nothing else is reached. Then:

• Port: filled in with the server's usual one (Ollama 11434, LM Studio 1234, llama.cpp 8080, vLLM 8000, Jan 1337, KoboldCpp 5001). Change it if you started the server on another.
• Server model: the models the server has that can answer. Embedding models are left out, since they cannot. The refresh button beside it asks again after you add one to the server.

Page Scanner starts no server and downloads no model: start the server and pull a model in it first. On Ollama, for example:

ollama pull qwen3.5:2b

In our test on a MacBook Pro (M4 Pro), Qwen3.5 2B found about as much as Chrome's model in less than half the time. The smallest models we tried, around 300 million parameters, took prices and company names along with the rest.

WebGPU runs a small model in the browser itself, on your computer's graphics card, with no server and no program to install. Useful where Chrome has no model for your computer and you would rather not run a server. Then:

• The first line says what the browser reports of the graphics card: whether it has WebGPU, whether it computes in 16-bit numbers, and how much memory the computer has.
• Browser model lists the models Page Scanner has tested, with the one it recommends for that card marked. A model the card cannot run is listed but cannot be chosen. The line under it gives the download's size, how the model scored in our test, and how long it took.
• Download fetches the chosen model from Hugging Face, once, with its progress on the line. It goes on while you open another section or close the settings, and the line shows where it stands when you come back. It is kept in the browser's cache for the next time. The bin button that takes its place removes it again, keeping any file another downloaded model still uses.

Page Scanner recommends Qwen3.5 0.8B, 16-bit (2.0 GB) for a graphics card that computes in 16-bit numbers on a computer with 8 GB of memory, and Qwen3.5 0.8B, 4-bit (670 MB), which runs on any card with WebGPU, for the rest. On a MacBook Pro (M4 Pro) they took 13 and 11 seconds for every 1,800 characters, where Chrome's model took about 2. LFM2.5 1.2B (fast) answers in under a second and finds much less.

Nothing of yours goes with the download; your text never leaves the browser. These models are slower than Chrome's and find less, so check what AI Scan found before you hide it.

Try it

The plug button asks the model about a short made-up order with five things in it to find. A green check beside it means the model found all five and nothing else. A red cross means it did not, or could not be asked, and the line under Try it says what it found or why. Nothing from your captures is used. You can open another section or close the settings while it runs: coming back, you see it still turning, or how it did.

Page Scanner's server

For the Assistant and the side panel: whether their messages may go to Page Scanner's server, which reads every one of them. A model there reads each message and answers it in up to 8 steps, through the editor's controls or the scan's choices. With each step it is sent the conversation so far, the choices and the counts the extension gives back, never the page's text, pictures, address or title. It keeps a count of messages per install: 10 a month, free, each counted once however many steps it takes, and a message that fails before the model reads it is not counted. On by default; the line under it says how many of this month's are left. Off, neither reads anything. The privacy policy says what the server keeps. While Page Scanner has the server closed to the Assistant, the switch is off and says so, and while the Assistant itself is closed the row is not shown.

AI Scan's prompts

One row for each kind AI Scan looks for (Names, Postal addresses, ID numbers, Payment details), folded until you open it with its arrow, holds what the model is told that kind is and is not. The line "Your own prompt." under a name says you have changed it. Each model understands some prompts better than others, so they are kept per model: what you write for one model on Ollama does not change what Chrome's model is told. Empty a box to go back to Page Scanner's prompt, or press the undo button (Reset to default) that an edited row shows.

Privacy

Report usage statistics

The extension counts which of its features are used and sends those counts to Google Analytics, so the features people use are the ones that get the work. The switch is on by default.

• On, it sends one event each time something is used: the popup opening (and whether over a page it can scan, never which page), a scan ending (the width, whether it came out as a PDF with real text or as an image, and for an image or a failure the step where the PDF capture stopped (never the error's words), whether it stopped at the size limit, how the main content was found, whether you picked one element, whether collapsed content was opened first, whether it went past 60,000 pixels, whether clean-up hid anything, how long it took), a mark placed (its kind only), an export and its options (among them whether the crop was the whole page, the main content or your own), the Sensitive Text panel used (how many finds and how they were hidden, never a find), a file opened or refused (a picture or a PDF, never its name or content), a preset saved, applied or deleted (never its name), the Assistant and the side panel used (whether a plan was proposed, run, undone or rejected and how many lines it had, never what you said), a setting changed, and the extension being installed or updated. Each carries a random id made on this browser, the extension's version, which browser it runs in (Chrome or Edge) and which store it came from.
• Off, it stops every request before it is made, and the page your browser opens if you remove the extension no longer names the version, the browser or the store in its address.

What is never sent: the address, title or content of a page, anything you type on a capture, a file name, anything the local agent connection carries, or a model's name, a server's port or a prompt you wrote. The capture itself, its marks and the files you export stay on your machine either way. The request goes from the extension's own code to Google's Measurement Protocol, with no Google script, no cookie and no added permission.

The command line and the MCP server send nothing. A scan they ask for is counted once, by the extension. The privacy policy lists every event and how long Google keeps them.

Local Agents

This is where you connect the extension to the command line or an MCP server on your own machine: Connect, Connect through (Page Scanner Helper, or a port and a token), Group an agent's tabs (a green tab group for the tabs an agent opens), and a name for this browser, plus the port and the token when you pair that way. When the connection fails, a toast says why, such as a helper that is not set up yet. Local agents explains what that means and how to set it up.

Where settings are kept

In your browser's local storage, on this device. They are deliberately not synced through your Google account, so your settings, your captures and your pairing stay in this browser.