most clipping tools scan your audio or video for loud moments. klypra reads your twitch chat instead, the same way a real viewer would, and reacts to what your community is actually reacting to.
audio-spike detection flags anything loud: a jumpscare, a mic bump, a sudden ad read. it has no idea whether your chat actually cared. twitch chat, on the other hand, is a live reaction feed written by the people who showed up to watch. when something genuinely funny, hype, or unexpected happens, chat reacts within seconds, and that reaction is measurable.
klypra's chat listener joins your channel's irc chat the same way a bot account would, and reads every message in real time.
a 10-message-per-second spike means something very different for a channel that usually sits at 1 message per second than for one that usually sits at 50. so klypra doesn't use a fixed number, it builds a rolling baseline for your specific chat, based on the median of your last several 5-minute snapshots (up to a 30-minute window). that baseline updates continuously while you're live, so it adapts as your chat naturally grows or slows down through a stream.
when the current message rate crosses a multiple of that baseline, klypra treats it as a candidate spike and starts scoring it.
a spike alone isn't enough to trigger a clip. klypra scores every candidate on a few signals at once:
those signals combine into a single moment score from 0 to 100. only spikes above your configured clip threshold actually turn into a clip request, everything else is noise that gets filtered out before it ever reaches your dashboard.
every account and every channel has its own adjustable clip threshold. lower it and you'll catch more moments, including smaller ones. raise it and klypra only fires on genuinely big reactions. there's no one right setting, it depends on how active your chat usually is and how selective you want klypra to be.