Your agents get a tab in the Chrome you are already signed in to
No fresh profile and no sign-in walls: each session drives a background tab in your own Chrome, and the evidence lands in the thread.
Most agent browsing starts in a fresh, empty browser: no cookies, no SSO, no staging login. Codecast takes the opposite route. A Chrome extension, paired once per machine, lets every agent on that machine work in the Chrome profile you already use. If your Chrome can open a dashboard behind SSO, the agent can read it. Nothing about your profile is copied or synced anywhere.
There is no browser button to press: you ask an agent in plain words ("check the checkout on staging") and watch it happen in Chrome and in the thread. Setup is once per machine. Install the Codecast extension from the Chrome Web Store, then in Agent features turn on the Browser card and click Pair; the card reads "Chrome connected". If an agent reaches for the browser before that, a Connect your Chrome card appears under its step with Install extension and Pair, and offers to tell the agent to carry on once it is done.
What you see in Chrome
Every agent tab opens in the background, inside a red tab group named Cast. Each session gets one tab and reuses it. Your window, tabs and focus stay put; agents act only on tabs they opened.
A driven page wears a thin red border (hidden from screenshots), the toolbar icon carries a CAST badge, and Chrome shows its debugging banner. When the work is done, the agent closes its tab.
Your screen changes only when the agent needs you (you asked to see the page, or it hit a sign-in): it brings its tab forward once, says what to do, and waits.
What you see in the thread
Each browser step is a row in the conversation with two pills:
- open tab raises the agent's tab in your Chrome, or offers to bring the page back if it was closed.
- watch live streams the tab into a pane beside the thread (or docked above it), the address flashing whenever the agent navigates.
The live view has one control that matters: Take the wheel sends your clicks and typing to the page, for a login the agent cannot do. Hand back (or Esc) and it carries on, signed in.
An agent can also offer you a page (a staging build, a failing dashboard): a chip appears beside the session title and a globe on its inbox card. Nothing opens until you click; an agent never moves what you are looking at.
How the agent reads a page
Agents do not guess at pixels. The page is read as a short list of things you can act on, each with a ref like #e7, and every action names a ref (or just "find Place order, click"). After a refresh, a stale ref is found again at the same position rather than pointing at whatever moved there.
Agents batch the steps they can see ahead into one run, which stops at the first failure and reports what ran and what never did. Waits are for text or a URL, never a fixed sleep.
Evidence lands without asking
A screenshot renders under the step that took it. An annotated shot draws the refs on the image, so you and the agent mean the same button by #e7, and a phone-sized viewport shows what a phone shows. When any step fails, the row brings its own context: console errors and failed requests, newest first, plus a screenshot. Network logs, vitals and recordings are there when debugging needs them.
Safety model
Your signed-in browser is a real grant, so the connection never leaves the machine: the extension talks only to a bridge on 127.0.0.1, and after pairing each side proves it holds the secret without sending it. An audit trail records the sites agents visit, by origin only, since full URLs carry tokens. A project can list the sites agents may browse; landing off the list is flagged.





