What a Browser Extension Can Read and Do: Read-Only vs. Automation
How the browser scopes what an extension can touch
An extension declares which permissions it needs when it is installed, and the browser both prompts you from that declaration and enforces it afterwards. For reading webpages, the part that matters most is site access: to do anything on a site, an extension first has to hold access to that site. The declared scope can be as narrow as one domain or as broad as every site, and Chrome states the range in the install prompt with a line such as “Read and change all your data on all websites.”
Once an extension holds site access, it can inject a script into the page, and that script sits in the same position as the page’s own scripts, able to read the page structure and its text. The same position also means it is in a position to click buttons, type into inputs, and submit forms. The browser hands over one set of capabilities, and how far an extension goes with them is its own decision, which is the starting point for everything below.
There is also a more restrained way to declare access, where the permission takes effect only at the moment you click the extension’s icon, applies to that one tab, and gives it nothing until you click. Chrome offers a control of its own as well: on an extension’s details page you can set site access to “on click,” “on specific sites,” or “on all sites,” and that setting stays yours to change at any time after installing.
Where read-only and automation differ
A read-only tool takes the text off the page, hands it to a model, and shows you the answer. It changes nothing about the page itself, and after a refresh everything looks the way it did before. The risk with this kind of tool concentrates in one question: where does that text go, and how is it handled once it arrives.
An automation tool runs through part of a workflow for you, so it has to genuinely operate the page by clicking buttons, filling forms, and triggering submissions. Those actions change the state of your accounts: a sent message cannot be recalled, a placed order needs a refund to undo, a changed setting has to be changed back by hand. The cost of a mistake is a different kind of thing here, one being content ending up somewhere it should not, the other being something happening in your account that you never intended.
Signed-in content makes the difference worth attention. An extension runs inside your own browser session, so a page script can read whatever you can see on screen, and that includes internal company tools, mail, and admin dashboards. The extension cannot bypass a login, and a page you cannot open stays closed to it too, so the two questions to ask are about everything outside that boundary: does this content leave your machine, and is this extension able to act on those pages for you.
What to check in the two minutes before installing. The store listing has a privacy practices section where the developer declares what data is collected, whether it is sold to third parties, and whether it is used for purposes unrelated to the core feature; a declaration that does not match behavior is grounds for removal, which is what gives that section weight. If the feature description promises form filling, acting on your behalf, or running through flows unattended, judge it as an automation tool. After installing, your browser’s extension details page lists the permissions it actually declared and lets you narrow site access there. Last, check whether it publishes a privacy policy that states how data is handled.
Using it in the browser
By the standard above, an AI extension built for reading should keep its own capabilities limited to reading text.
PageGrok is a Chrome side panel built for reading webpages. When you open the panel it reads the text already rendered in your active tab, and you can also clip one or several regions of the page and ask about those instead. It clicks nothing on the page, types into nothing, submits nothing, and fetches no other URL. The rule for what it can read is simple: text you can see in the tab is text it can read, including content that only appears once you are signed in and pages rendered dynamically by the front end.
At install time Chrome tells you it can read data on the sites you visit, which is what reading the current page requires. With that access, the only thing it does is take the text. You can set site access to take effect only when you click the icon, and the side panel works the same way afterwards.
Where the content goes is your choice. With built-in AI, what you submit is sent to PageGrok’s service to generate an answer; the service does not store page text, questions, or answers, and does not use them for training. With Ollama, LM Studio, or oMLX on Apple Silicon running on your own machine, page content moves only between your browser and that model.
Further reading: Summarize any webpage, Local AI in the browser.
FAQ
Does an AI reading extension need setup before I can use it?
Install PageGrok, open the side panel, and you can ask about the page you are on. Built-in AI runs on a lightweight cloud model, with no account and no API key required. It only reads text already rendered in your active tab; it does not click, fill forms, or fetch other URLs.
Can an extension see content that only appears after I sign in?
Yes. An extension runs inside your own browser session, so a script in the page can read the text you can see on screen, including content that only appears once you are signed in. What an extension cannot do is bypass a login: a page you cannot open, it cannot open either. The thing worth checking is where that content goes next.
How do I tell a read-only extension from one that can act for me?
Read what it advertises. Anything that fills forms, places orders, or runs through a flow on your behalf is an automation tool and needs to click and type in the page. Reading, summarizing, translating, and question answering usually need nothing beyond text access. After installing, your browser's extension details page lists the permissions it actually declared.
What can PageGrok read?
Anything rendered as text in your active tab: ordinary articles and documentation, dashboards and internal tools that appear only after you sign in, and pages the front end assembles after load. On a long, dense page you can clip one or several topic regions so only those go to the model, which keeps the answer on point.
Try it on the page you are reading right now
Built-in AI works with no setup and no account. You can also connect a local model.