The browser is a rail view like the others, and it is a real one. The page on screen is the page itself rather than a picture of one, which is what lets the agent read it and work on it.
Its own tabs, logins, and history
The browser keeps a profile of its own, separate from the browsers you use outside the app. Sign in to a site here once and the session stays, so the next visit, and the next conversation that browses alongside you, arrives already signed in. Nothing is copied from your everyday browser, so there are no cookies or saved passwords to carry over.
Every page opens in its own tab. The strip beside the page is where you move between them, and closing a tab takes that page's history with it. A tab you leave open keeps its place, and its back and forward history, when you switch to another view and come back.
One field for addresses and searches
The address field takes both kinds of input. Type an address, or type the words you would have typed into a search engine: text with a dot, an IP address, or localhost opens as an address, and anything else goes to the engine. A search never needs a special gesture.
DuckDuckGo is the engine the app starts on and Google is the other one it ships. You can choose between them, and add your own with a query template, in settings.
Pages served over https or http open. An address that uses another scheme is refused, which includes local files and script addresses. A development server on localhost or on your own network opens like any other site.
The page is the context
Open the agent beside a page and that conversation keeps the page as its context. The panel says Ask the agent about the page on this tab until you start one, and starting it is what creates the conversation: a tab you only visited never gets a transcript you did not ask for.
From there you can ask about what is on screen instead of describing it. The agent reads the page you are looking at, including what its console says and what a screenshot shows, so a question about a broken layout or a request that never came back has the evidence in front of it.
Ask for the code behind the page to change and the answer arrives in that same panel, beside the page it is about. A finished turn still raises its own notification entry, so you can leave, answer something else, and come back to the page with the reply waiting next to it.
Where manual browsing ends and the agent starts
Browsing by hand is the default. You hold the address, the tabs, and the pages, and the agent sits beside a page and reads it when you ask.
The agent drives a page it opened for itself. Navigating, clicking, typing, and reloading belong to those pages, not to the tab you are steering; reading is the part that reaches the tab in front of you. Two conversations never drive the same page, so a worker in the background cannot move the page you are reading.
The agent has no way into a browser you opened outside the app, and no way to click around other applications. Its reach ends at the browsers inside CodeInOven, and it acts on a page only while you are working in one.
What the browser asks you first
A page that wants your camera, your microphone, your location, or your screen asks, and the app starts from deny. A download goes through the save dialog rather than landing in a project by default, and it keeps its progress visible until you act on it.
A page that opens a window, such as a sign-in or a checkout, is hosted inside the browser instead of becoming a separate window that can hide behind the app. When the page closes that window, the browser follows it.
Settings for the browser sit with the rest of the app's settings, so the engine you search with, the sites you allowed something, and the downloads you kept are all in places you can find again.