Warden reads the comments on a page and rewrites the hostile ones, so the argument survives and the abuse does not. It runs as you scroll, in place, with no separate dashboard to check.
Free tier: 10 rewrites a day, no card required.
What it actually does
The same comment, at each of the three strictness settings. Warden keeps whatever point was being made and changes only how it lands.
"this update is absolute dogshit and the devs are clueless who clearly never played their own game"
left as written — the target is a thing, not a person
"This update is really bad and the developers seem out of touch with their own game."
"This update has some rough edges, but I hope the developers can improve it in the future."
The complaint survives every setting. What changes is the register — from blunt, to plain, to hopeful. Nobody is told the update was good.
"the housing shortage is a policy failure, though these parasites flooding in have made it worse"
"the housing shortage is a policy failure"
"I really hope we can sort out the housing shortage"
The argument about housing is kept in full. The claim about a group of people is removed rather than reworded — not softened into "the influx of people", which would hand back the same claim in politer words. That distinction is the whole design, and the thing most of the models we tested got wrong.
"mate you are so bad at this game it is actually impressive"
"you're really struggling with this game"
"you're still learning, keep at it"
Aimed at a person, but it is about how they played — not who they are. So the observation stays, and only the sting is removed.
"what a pathetic little loser you are, nobody cares what you think"
"I'll leave this one."
"Everyone has rough days."
Here there is no argument to save — it is only an insult. Warden steps away rather than inventing a point nobody made, and at the strictest setting answers it with something kinder instead.
How it works
When a page loads, Warden collects the comment text you can see on it. Nothing from other tabs, nothing from your history, and no images.
Ordinary text is discarded before any judgement is made — on a normal page that is roughly two thirds of it. Only what survives that filter counts against your daily allowance.
What remains is sent to a large language model with a prompt written and tested for this one task. Anything it flags is rewritten in place, next to the comment that triggered it.
Three settings: attacks on people only, all hostility, or anything negative at all. You can change it at any time, and the page updates without re-reading anything.
What it is, plainly
Comment text is sent to DeepSeek, whose servers are in China, to be judged. It is not processed on your machine. We do not store the text of the comments we judge, and we never sell data. Read the privacy policy.
Warden runs on a general-purpose language model. What is ours is the prompt, the guard rails around it, and the test suite that keeps it honest — not the model weights.
It sometimes rewrites something that did not need it, and it can miss things. It is a filter you control, not a moderator, and the original text is always one click away.
Measured, not claimed
Those figures come from a 43-case internal test set, run repeatedly. It is a small set that we wrote ourselves, so treat it as evidence that the thing works rather than an industry benchmark. We publish the numbers we have rather than adjectives we cannot support.
Ten rewrites a day, no card. If you want more, Guard is £24.99 a month for 100 a day and Pro is £39.99 for 300.