Describe what should happen on a website, and an agent does it in a real browser you can watch and take over. The description is drawn back as a blueprint before anything runs, and the same drawing is what the API returns.
Accounts are open. Sign in with a Google, Apple or GitHub account, or with an email address and a verification link. One sign-in works for the app, the API and the live browser view.
Every account owns a personal workspace and can create or join others. Invite people by email, give them a role, and work in whichever workspace the task belongs to. The plan and the credits belong to the workspace; your browsers stay yours.
Each person’s browsers are separate Chromium profiles on their own displays, listed by id and reachable only by their owner. Open a session to watch the agent work, take the controls when a site asks for a login or a file, and hand them back. Files come from your own device through the live view.
A description with a time or an event in it now gets two actions in the composer: Run once, and Turn on. Turn on keeps it running on the schedule the words describe. Each automation has its own page with its blueprint and every run, and can be paused or given a monthly credit cap.
Reading a description is now done by the model, and returns the same shape the app draws: a trigger, the steps with their connections, and a note when something is ambiguous. POST /describe/read takes the text and a timezone and returns the plan.
Runs are charged in credits by what the model actually used, never less than one credit, and the charge is shown on every run. Free grants 200 credits a month and Pro 2,500; buy more from Billing when a month runs hot. Any automation can be capped at a monthly limit.
Browsers are now registered in the control plane with the node that holds them. The app is sent to the right node for each session, nodes report in on a heartbeat, and a browser on a node that has gone quiet is placed again on the next request.