Masky Chat by masky.ai
Masky Chat is a desktop HUD for streamers, Discord moderators, and viewers. A local AI scores every message for relevance, toxicity, logical fallacies, and manipulation — surfaces only what deserves your attention, translates chat into your language, and lets Masky avatars read the best messages aloud on stream.
Try it right now → you're already in the global room. Type a message and watch the judge score it before you send — the same system-1 verdict the app gives every chat message. If it would be blocked, you'll see exactly why.
The same judge working the demo above runs locally inside the app — no API key, no per-message cost, nothing leaves your machine.
Twitch, YouTube, Discord, and Facebook chat in one feed, each message scored for how much it deserves your attention right now. Two sliders — relevancy and safety — own the feed. Your mic feeds the judge context, so "relevant" means relevant to what you're saying.
On a cadence, the most relevant message is read aloud by a Masky avatar in an OBS-ready transparent overlay. Chatters with their own Masky avatar get read in their own face and voice. Leaderboards track who gets read the most.
Every message gets a safety score built from the persuasion-technique literature: loaded language, reputation attacks, false logic, coercion. Flags name the technique — a lens for the human decision, never an automated verdict.
Incoming chat translates locally into your language — on your machine, no cloud round-trip. Follow a stream in any language, and have readings spoken in the voice you choose.
The composer judges your draft as you type — relevancy, safety, kind, and any technique flag — so you know how the room's filters will treat your message before it exists.
Judging runs on a local model via Ollama. No API key, no subscription, no chat data shipped anywhere. Pick any model you like in settings.
Every judgment feeds the leaderboards — reads on stream, message kinds, and a board per manipulation technique. Click through the tags; this is the app's own trophy room (sample names shown).
One forward pass per message answers a structured battery: a 0–3 relevance rubric, a 0–3 manipulation rubric (double-weighted into the safety score), an outcome-steering probe, a persuasion-technique family, and a message kind. Scoring follows the SemEval-2023 persuasion-technique taxonomy, MAFALDA, and MentalManip. Sliders threshold the scores — sliding never re-runs inference.
Greetings and emote spam score low; direct questions, stream issues, raids, and safety reports score high.
100 is straight talk. The score drops as rhetoric turns to pressure: loaded language, smuggled presuppositions, guilt leverage, pile-on bait.
Confident flags name the dominant family — reputation attack, false logic, distraction, coercion — as a hypothesis for the human, never a verdict.
The room's judge blocked it before it ever left your browser — the same pre-send verdict the app gives.