Glyph Reach

Glyph Reach connects your own Glyph account's AI chats to this browser so they can find things on a page, click, type, fill forms, scroll, read, take screenshots and follow downloads — in your real, logged-in Chrome, on the sites you direct them to. It is built for Rogue AI staff and for clients running their own Glyph. It is not a consumer product and it does nothing on its own: every action comes from a command you issued in your own chat. How it works 1. Pair once. In Glyph, open Settings > Glyph Reach and choose "Pair this browser" (on glyph.rogueai.tech that is one click; elsewhere you paste the code into the extension's popup). Your browser exchanges a short-lived, single-use code for a token that belongs to that browser alone. No shared secret is compiled into this extension. 2. Ask. In any of your chats, ask for what you want done on a page. The chat sends one command at a time to the browser you paired. 3. Stay in control. Revoke any browser from the same settings page at any time; revoking stops work in flight, not just the next command. If you click or type in a tab yourself while a command is running there, including inside a frame on the page, Reach stops that command and pauses automation on that tab until you tell your chat to carry on (the chat is told only the kind of input — a click, a key, an edit, or the tab being brought forward — never which key or what text); between commands the tab is simply yours. Scrolling with the mouse wheel does not stop a command, and Reach may scroll the page to reach what it acts on. One narrow case may not stop it: an edit you make without a key or a click — text you drop in from another application, or dictation — that lands while the page is still handling Reach's own typing, or whose text is part of the text Reach is typing into that page at that moment. The page's own reactions to what Reach typed or clicked (a field that reformats itself a moment later, an editor that inserts a template) are Reach's doing and do not stop it when they come within about a second and a half of Reach's input; a reaction later than that stops the command, and the chat is told a person used the tab. What is new in 3.0.0 - It can see far more of a page. The page reader now walks inside shadow roots, iframes and custom components, and can address plain text and click targets that carry no accessibility role, so pages that previously came back almost empty are usable. - Actions are hit-tested before they fire. If something covers the element, the click is refused and named rather than landing somewhere else. Inside a frame, a click is pressed only once the pointer is seen to land on the intended point, and a click aimed by screen position reports where it actually landed; it is never reported as done when it landed elsewhere. - A result says what was observed, not what was hoped: whether the field now holds the typed text, whether the page received the click or key, whether a drop was accepted, whether the click made the page load a new page or open a new tab. What the page's markup only predicts is labelled as a prediction. When nothing could be observed, the result says that too. - It works in a background tab. Reach does not need to bring a tab to the front, focus your window or steal your cursor, and a Reach click is exactly a click: no key is held down with it, so the page does what it would do for yours. When the page opens a new tab or window — after a link, Enter, a form, or the page's own script — Reach first asks whether the browser shows that Reach caused it: it came from the tab Reach is working in while you were not on that tab and had not been on it in the six seconds before, or it appeared during one of Reach's own clicks or key presses there while the command ran (a key press is Enter, Tab, a space or a key the chat asked it to press, never a letter of the text it types). When the page brings forward a tab the chat already has (a link or script aimed at that tab's window name), Reach counts it as its own only when that tab began loading during one of Reach's own clicks or key presses: a tab coming forward is never, by itself, a sign that Reach did it, because your own choice of that tab looks the same. Only a tab that passes one of these tests is given to the chat, and only such a tab is ever moved. If you are watching or using the tab Reach is working in, nothing is moved: the new tab or window opens exactly as it would for your own click. If you are somewhere else — another tab, another window, another application — Chrome can show the new one in front of you for a moment (on our test machines usually a tenth of a second or less, up to about three-quarters of a second when the computer was heavily loaded), and keys you type or clicks you make in that moment go to the new page. Reach then puts back what you had in front (a new window it cannot put behind is minimised) and never gives focus to a window it opened. It looks again immediately before it moves anything: if you have already moved on or chosen a tab or window yourself — the new tab included — Reach leaves you where you are, and if your choice lands in the same instant as Reach's own move, Reach brings your choice back. Everything else stays where Chrome put it, even in front of you, and is not given to the chat: a tab you open yourself, including one that appears a few seconds after your click; one the page opens while you are on that tab or after you have been on it in those six seconds; a tab of the chat's that the page brings forward later on its own (a timer re-using its window name); a tab of yours that a link on the page Reach clicks is aimed at; a tab, yours or the chat's, that you have moved into a window of its own, even when a click of Reach's loads a page into it; and a pop-up the page re-focuses without loading anything, which Reach cannot tell from your own click on that window. For a new tab it cannot attribute, the chat is told only that one opened, and to ask you …