Email and password, then a six-digit code from your authenticator. Tick trust this device and it stops asking on this browser for thirty days — on this browser only, and you can revoke it later from Settings without touching the others. Forgot the password? Forgot your password? under the form sends a single-use link to your e-mail.
If somebody invited you, you did not start here: you opened the link they e-mailed you, which named the team and let you choose your own name and password.
A set is not a TV. It is a TV plus the camera pointed at it, and (optionally) the smart plug on its socket — because everything Screenwhere does needs both halves: the camera to see the screen and the box to press the buttons. Add one under Devices → Add: catalogue fields, the LAN addresses, and whether this set may be recorded at all. Then Generate pairing command and run that one command on the box in the TV's network: it works once, for 30 minutes, and the box configures itself from it.
Nothing reaches into your network from outside. The box on the TV's LAN dials out to the relay and keeps that connection; the cloud never opens anything inwards.
Open the set. You get the live camera view and a remote deck beside it. Only one person holds a set at a time — presence shows who, and the hand-over is explicit, so two people cannot quietly fight over the same screen. Everything meaningful you press is written to the audit log.
Record while you work. The recording keeps the video and the key track — every button you pressed, timed against the picture — so review is not guesswork: scrub the video and the presses move with it. Trim, replay onto a live TV, export. Mark one ★ and the storage sweep will leave it alone.
A scenario is that key track turned into something repeatable, and it is not tied to the TV
it was recorded on. Write or record one in the editor, keep it in the library under folders and tags, and run
it against any set in the project. Steps are short to write — press with times
for repeats, screenshot to keep a picture with the run, and an app picker for launching apps
(with an optional payload, a deep link that opens the app on a chosen screen) —
and checks compare what the camera sees with the text or picture you expect.
Watches run scenarios on a schedule and tell you when a screen stops looking like itself. Runs is the history: every execution and its verdict, kept apart from the watch settings on purpose — a run history is not a configuration. A failed step says whether it is the first failure or a repeat, and when a screen was meant to change, Adopt as the new reference makes the new look the expected one. Failures reach you in the app, by webhook or by e-mail.
Scenarios are versioned, and more than one line of work can exist at a time. Releases is where those versions live; promotion is a deliberate act, because production should be somewhere you get to, not wherever somebody last saved. A promotion's run can be followed live with Watch live.
Four roles. Read-only may look and nothing else; Standard user works; Team admin manages a team; Superadmin runs the installation. A role belongs to a membership, not to a person — the same account can be a Team admin in one team and Read-only in another — and teams decide who sees which TVs, so a person's reach is the team's reach, not a global switch.
Invite by e-mail rather than creating accounts for people. The invitation is bound to that address, names the team, and can be revoked under Teams until it is used; a bad or spent link gets one neutral refusal, so it cannot be used to discover what teams exist.
Projects keep one client's work genuinely apart — scenarios, recordings, builds and watches belong to a project; televisions and people do not, because a screen in the corner of the office is not a client's property. Day is the screen you can put in front of that client: one day of runs, what was found, and nothing about how it is wired.
Cameras and availability answers “is it even online” first: last-seen per set, 48-hour availability bars for plugs and boxes, and each camera's drift from where it was calibrated. Audit answers “who did that”, with security events called out from ordinary ones. Installation — for whoever runs the installation — holds the quotas, the disk reserve and the operations alerts. Data and retention, from Settings, says how long everything is kept and exports a team's data.
Everything above has a machine-facing side: an MCP server that lets an AI agent list sets, launch apps, send keys and text, write and run a scenario or a whole suite, read the results back, and install and debug your own app build. Same permissions as a person — an agent is not a way around the roles. See MCP.
Stuck on how any of it is wired? That is architecture. Getting our app onto a Samsung TV in the first place is cert-gen.