An automatic clip maker for Twitch that starts at the VOD
There is no Twitch integration here: no channel to connect, nothing watching while you are live. You export the VOD and upload the file, and every passage worth clipping comes back ranked, with the words that made it a candidate.
What the Twitch route actually is
How it choosesGetting a stream in means the file, unless it is also on YouTube
Nothing is connected to Twitch. There is no integration, no permissions screen and no channel linked to your workspace, and none is planned. A page that implied one would win the visit and lose the reader in the first minute, which is the wrong trade to make on purpose.
The usual route is export, then upload. Export the broadcast, wait for the download to finish, and drop the file on the page. Four containers are read — mp4, mov, mkv and webm — up to 8 GB. The app reads the file, not where it came from, so a stream that already lives on your disk skips the export.
Unless the broadcast is already on YouTube. The paste box takes a youtube.com or youtu.be address — a watch page, a Short, a live URL or an embed — so a broadcast that is also on your YouTube channel can be handed over as a link and fetched once on our side, with no export to wait for. You confirm the video is yours to use before the import starts, and what is stored is the sentence you were shown, the version of the terms, the time and the address. Imports are capped by the day — three on the trial, twenty on Starter, fifty on Creator, a hundred on Studio — and uploads are not.
And it has to fit. There is no cap on how long a single video may be, on any paid plan — a seven-hour VOD is read the same as a thirty-minute one, and simply spends credits by the minute. The free trial reads thirty minutes of video, once, with no card.
Live is not a thing it does. Nothing watches a broadcast and cuts as it goes. The stream ends, the VOD exists, and then this starts.
What each plan takesWhat a Twitch AI clip maker can read, and what it cannot
Selection happens entirely on the transcript. The audio is transcribed word by word with timings, and the model that ranks passages sees those words and nothing else — no frame is looked at until you have approved a moment and it is time to cut.
The talking parts
A story between queues, a take you did not know was a take, twenty minutes with a guest on the second camera. Anything with words under it can be ranked.
Not the gameplay
Nothing in the selection sees a frame. A clutch nobody spoke over does not exist as far as the ranking is concerned.
Not your chat
Chat is not read, and neither is a hype train. What the room reacted to is only visible here if someone said it out loud.
Fewer than you asked for
The selector is harsh on thin material and returns a short list rather than padding one. On a quiet stream that is the honest answer.
An AI clip maker for Twitch that offered to spot your best plays would be offering something this one does not do, and AI gaming clip maker is the wrong description of it altogether: it is a reader, and what it reads is speech. On a stream carried by what you say, that finds most of the value in eight hours of footage. On a silent grind it finds close to nothing, and the list says so by being short.
A long VOD read in windows, not in one pass
One pass favours the beginning. Read straight through, a long transcript anchors on what it sees first and stops picking from the back half. That was measured on whole episodes, and it is why windows exist at all.
Windows, in parallel. Around half an hour of content each, six at most, with three minutes of overlap on every inner edge so a passage sitting on a seam is still read whole by one of them. The finalists from all of them go into a single pool, and a final judge ranks them against each other rather than against their neighbours.
The opening is set aside. A model reads the start of the transcript and decides where the content actually begins, so the ranking is not spent on the part where you are greeting people and waiting for the raid to land.
Boundaries are words, not seconds. The model never returns a timecode. It quotes the words it wants to open and close on, verbatim, and the times are read off those words — so a cut can land anywhere except inside a word.
Then a stranger reads it. Every candidate is read again with the clip’s own lines and nothing else: no title, no transcript around it, no timecodes. If it only makes sense to someone who watched the previous two hours, the cut is made again with the report attached. A clip that still cannot be followed cold stays in the list but cannot score above 60, and it carries a line saying what was missing.
The same reading, on a podcastThe frame follows whoever is talking
Speaker tracking is TalkNet-ASD. On a co-stream, or a VOD with a guest on the second camera, the crop stays on whoever is actually talking rather than on whoever happens to be in shot. This is the part of a stream recording a fixed vertical crop handles worst.
Two ship by default; two more need a camera. Boxed 1:1, with the words composited behind the speaker — the subject is matte-cut and sits over the text — or full-frame 9:16 with karaoke captions timed off the transcript. Pick one, save it as a preset, and every moment you approve comes out the same way. A broadcast with a webcam over the game gets two more automatically: the screen filling the frame with the camera riding over it as a circle, the way it looked live, or the camera across the top with the screen filling the rest. Both are offered only where the shape watcher finds a camera in the picture, which on a gameplay VOD is most of it.
The banner guard is a lower-third detector. It measures the footage for a graphic frozen along the bottom of the frame and punches in only if it finds one. That is the only shape it looks for: an overlay somewhere else — a webcam border, an alert box in a corner — is part of the picture as far as it is concerned.
You decide, then it cuts. Nothing renders until you approve a moment, and nothing is posted anywhere for you. The finished MP4 waits in your library for you to download and upload wherever it is going.
Where the vertical file goes nextQuestions
The ones worth answering before you spend an evening exporting a VOD.
Does it connect to my Twitch account?
No. There is no Twitch integration, no permissions step and no channel to link. A broadcast arrives as a file you upload — mp4, mov, mkv or webm, up to 8 GB. The one address the paste box takes is a YouTube one, so a VOD that is also on your YouTube channel can come in as a link instead, once you have confirmed the video is yours to use.
Can it clip my stream while I am live?
No. Nothing watches a live broadcast. The stream has to end, the VOD has to be exported, and the file has to be uploaded before a word of it is transcribed.
How long a VOD can I upload?
There is no cap on any paid plan — Starter, Creator and Studio all read a VOD of any length; it simply spends credits by the minute. The free trial reads thirty minutes of video, once, with no card.
How does the Twitch clip maker AI decide what to keep?
It reads the transcript and scores each candidate on four pillars: hook, flow, value and travel. The travel score comes from a rubric distilled from real shorts that beat their own channel median views, so it is a prior about what travelled — not a prediction that this one will.
Will it find a clutch play with no commentary?
No. Selection runs on words alone, so a moment nobody spoke over is invisible to it. The streams it reads well are the ones where the talking carries the broadcast.
Does it upload the clip to TikTok or YouTube Shorts?
No. It hands you an MP4 in your library. Downloading it and posting it stays yours, on every platform.
Is the uploaded VOD kept, so I can cut more from it later?
Yes, for as long as your plan keeps clips: seven days on the trial, ninety days on Starter, a year on Creator, and on Studio for as long as the subscription stays active. While the file is there you can keep cutting new moments from a stream you have already paid to read.
The other ways in
Same pipeline, described from wherever your footage is now.

YouTube clip maker
Paste the watch URL and it is fetched once on our side, or upload the file. Only the passages you approve are ever cut.

Clipping software
What the whole pipeline does, end to end: transcript, ranked candidates, your decision, then the cut.

Podcast clip maker
Two hours read in thirty-minute windows with three minutes of overlap, so the back half is ranked too.

TikTok clip maker
Vertical 9:16 with word-by-word captions. It hands you the MP4 — the upload to TikTok stays yours.
