What it does
You email Rover and tell it what to watch, in your own words — not a keyword, a brief. "Nvidia, but only the open-weight models and developer tooling, not the share price." "What Jensen Huang is saying in public — the keynotes and the interviews, not the results." "What the UK food industry is doing about sugar reduction."
Every morning it reads what has appeared since yesterday, throws away almost all of it, and sends you what is left. On a quiet day it says so and sends nothing else. That is the feature, not the failure.
Why we built it
Because the alternatives fail in two opposite directions. A keyword alert sends you everything that mentions a name, which is mostly noise and a share price. An AI summariser sends you something fluent that you cannot check — and in this field, fluent and wrong is the expensive combination.
We were already running an AI newsroom here, and had learned the hard way that the model is not the thing to fix. A better model still invents a quotation now and then. What stops it reaching a reader is architecture: make the machine prove each claim against a page it fetched itself, and throw away anything it cannot. Rover is that same gate pointed at one person's interests instead of a publication's.
The experiment: an agent that lives in your inbox
The other half of the experiment is the interface, or rather the absence of one. There is no app to install, no account to create, no password, no dashboard to check and no settings page. Everything happens by email — the one tool everybody already has open, already knows how to use, and already has on their phone.
- Signing up is emailing Rover and saying what you want watched. That is the whole form.
- Setting it up is the sentence you already wrote. Rover reads your own words back to you — the subject, the angle, what you said you did not want — and if it has read you wrong, you reply and say so.
- Steering it is replying. Add a subject, drop one, tell it the last briefing was off-target. Two links under each finding mark it as more or less like what you want.
- Adding something can be as little as forwarding an article. Rover reads it, works out the one or two subjects it is actually about, and tells you what it took — your covering note outranks the piece itself.
- Asking is replying with a question. It researches the answer the same way it researches a briefing, and if it cannot stand anything up it says so rather than dressing it up. Five a day, because each one is a real research run.
- Leaving is one click, or the word STOP in a reply. A second button deletes everything held about you, immediately.
We wanted to know whether an agent could be genuinely useful with an interface that thin — whether the thing people find hard about AI tools is the models, or the onboarding, the tabs and the remembering to go and look. Our bet is that a briefing you did not have to go anywhere to read beats a better one behind a login. Rover is how we find out.
How it proves what it says
Every factual statement in a briefing carries a quotation, and that quotation must appear literally in the page we downloaded. Not similar to it — literally in it. A model that invents a fact invents the quote along with it, and the check kills both. Nothing that fails is flagged, softened or hedged; it is dropped before you see it.
- Quotes are matched verbatim against the fetched source. Punctuation and encoding differences are forgiven; a changed number is not.
- Figures in the prose must already appear in a verified quote. An invented percentage cannot survive the sentence it sits in.
- Dates are about the event, not the page. A press release published this week about something from last year is last year's news, and it is dropped.
- Links are attached by the machinery, never typed by the model, so a citation cannot point at the wrong thing.
- People are quoted from recordings where we can get them. A journalist's paraphrase is someone else's sentence; a transcript is theirs.
The result is a briefing that is shorter than it could be and slower to impress. We think that is the right trade. If Rover has nothing that clears the bar, it tells you it found nothing rather than filling the space.
Ninety seconds, whatever you ask for
Watch ten things and you do not get ten sections. The shape is fixed — a fifteen-second scan on top, then at most three stories, each with why it matters to you and a suggested next step — and the findings compete for those places on how well they are corroborated, how recent they are, whether they develop something you were already told, and how closely they fit the brief you wrote. Anything that was watched but did not make the cut is named in one line at the end, so a cut never reads as a gap.
It also remembers what it has already told you. The same story does not arrive twice in different words; if it genuinely moves on, it arrives as a development and says what changed.
What it is not
It is not advice — not financial, legal or professional — and it does not make decisions. It summarises public information and shows you where it came from. It has no opinion about what you should do next, and it will not pretend to.
Where it is up to
Rover is in a closed beta, deliberately small — small enough that the interface question gets a real answer rather than an average one. Every briefing carries a one-click way out and a button that deletes everything we hold about you; what we hold, who processes it and for how long is set out in the privacy policy, and the rules of use in the terms.
Most AI tools ask you to trust the output, and to come to them. This one is built so you need do neither.