Clip anything by prompt
Type what you want — “every time they mention pricing”, “the funniest 30 seconds” — and get clips that match, plus an optional time range.
vs. Opus: prompt + explicit time-range filter, on every plan.
Open source · self-host freeMIT
Turn long videos into ranked, ready-to-post clips in one click.
Drop in a podcast, webinar or stream. Latch finds the moments, scores them with a virality score you can actually read, reframes to follow the speaker, adds animated captions in 30+ languages, pulls B-roll, and schedules the posts.
No credit card. Runs on your machine or our cloud.
Import from · publish to
- YouTube
- TikTok
- X
- Twitch
- Zoom
- Google Drive
source · 1:12:43 · 1080p16:9
877972816674
6 candidates · ranked
0:349:16
Live preview is a CSS illustration — real output uses your video.
Built for podcasters, educators, agencies and streamers who post every day and want to know why a clip was picked.
- ~3 min
- average turnaround per hour of source video*
- 30+
- caption languages via Whisper
- 6
- explainable score dimensions, with reasons
- 0
- vendor lock-in — export FCPXML, Premiere XML, EDL
*On a single consumer GPU with faster-whisper large-v3; CPU-only is slower.
How it works
From a two-hour recording to a week of posts
Three steps. Every one of them is inspectable — no black boxes between your video and the clips.
- 01
Drop a link or a file
Paste a YouTube, podcast or Zoom link — or upload an MP4 up to several hours long. We extract audio, normalize it, and transcribe with word timestamps and speaker labels.
ingest → transcribe
- 02
AI finds and ranks the moments
Chapters are detected from topic shifts and scene cuts. Candidate windows snap to sentence boundaries, then each is scored on hook, flow, value, trend, emotion and clarity — with written reasons.
chapters → candidates → score
- 03
Edit, brand, export or schedule
Tweak captions, swap the aspect, toggle jump-cuts, apply your brand kit. Download MP4 + SRT, hand the timeline to Premiere or Final Cut, or schedule straight to your channels.
plan → render → publish
Features
Everything the paid tools do — with the lid open
Twelve things Latch does out of the box. Each card notes how it differs from the closed alternative.
Explainable virality score
Six sub-scores (hook, flow, value, trend, emotion, clarity) and plain-English reasons for each pick. Heuristic first, LLM-refined when you add a key.
vs. Opus: a single 0–99 number with no breakdown.
Speaker-tracking reframe
Face and speech tracking keep the active speaker centred when going 16:9 → 9:16. Two speakers? Auto split-screen.
vs. Opus: tracking without automatic split-screen.
Animated captions
Six presets — Pop, Karaoke, Box, Bounce, Minimal, Outline — keyword highlight, emoji, per-platform safe zones. Word-level timing.
vs. Opus: 30+ Whisper languages instead of 20+, safe zones built in.
Filler & silence removal
“Um”, “like”, dead air — detected from the transcript and waveform, turned into jump-cuts you can toggle one by one.
vs. Opus: every cut is listed and reversible, not baked in.
AI B-roll
Keyword-driven stock footage from Pexels, or your own local library, dropped in fullscreen, picture-in-picture or top-third.
vs. Opus: B-roll on all plans, including self-host, and your own library.
Hook titles, descriptions & hashtags
Each clip ships with a title, a hook line, a description and hashtags tuned to the transcript — editable before you post.
vs. Opus: generated with the same model you bring, not a locked one.
Brand kits
Logo placement and scale, primary and accent colours, intro/outro clips, watermark text and a progress bar. Save as many as you need.
vs. Opus: unlimited kits on self-host; no plan gating.
Multi-aspect export
9:16, 1:1, 16:9 and 4:5 from the same plan, up to 4K, with loudness normalisation and per-platform encode presets.
vs. Opus: 4K and every aspect on every plan.
Timeline export
Hand the whole cut to Premiere Pro (XML), Final Cut (FCPXML) or any NLE via EDL — including reframe keyframes and cuts.
vs. Opus: XML export is Pro-only there; here it's standard.
Social scheduler
Queue clips to YouTube, TikTok, Instagram, LinkedIn and X. Each platform's length, caption and aspect rules are validated before you hit save.
vs. Opus: validation errors are surfaced in the UI, not at publish time.
Self-host & API
One `pip install`, a CLI, a REST API with SSE progress. Bring your own Whisper and LLM keys. No per-minute credits — ever.
vs. Opus: API is Business-tier; ours is MIT-licensed and included.
Captions
Six caption styles, word-accurate, safe on every platform
Captions come from word-level timestamps, not sentence guesses. Pick a preset, set your highlight colour, and Latch keeps text out of the TikTok, Reels and Shorts UI zones automatically.
- Pop, Karaoke, Box, Bounce, Minimal, Outline — or tune colours, size and position.
- Keyword highlight and optional emoji, generated from the transcript.
- 30+ languages via Whisper; burn in or export SRT/ASS instead.
Bold, outlined, the current word scales up and turns your highlight colour.
- Keyword highlight
- on
- Emoji
- auto
- Safe zone
- TikTok
- Words / line
- 4
- Position
- center
- Font
- Inter
Stopeditingbyhand🎬startpostingdaily
Reframe
Landscape in, speaker-centred portrait out
Face detection plus who-is-talking from the transcript drive a smooth crop path. When two people trade lines, Latch switches to a stacked split-screen instead of whip-panning.
- Modes: track, center, split, static — or auto.
- Keyframes are exported with the timeline, so your editor can adjust them.
- Works without a GPU; OpenCV/MediaPipe are optional accelerators.
face 0.98
16:9 source · 2 speakers detected
split
9:16 · auto split
Virality score
A score with receipts
Every candidate gets six sub-scores and a list of the concrete reasons behind them. Heuristics run on every clip; when you add an LLM key, the reasons get sharper and the score is marked “hybrid”.
- Hook, flow, value, trend, emotion, clarity — 0–100 each.
- Reasons cite timestamps and quotes so you can verify them.
- Re-rank with a topic prompt without re-transcribing.
Candidate 03 · 0:41
“Why does nobody talk about this?”
- Hook91
- Flow82
- Value88
- Trend67
- Emotion90
- Clarity85
Why this scored 84 · hybrid
- Opens with a direct question in the first 2.1s (“Why does nobody talk about this?”).
- Self-contained: no unresolved references to earlier parts of the episode.
- Two laughs and a tonal shift detected between 0:14 and 0:22.
- Topic “creator burnout” trending up 38% on our 30-day keyword index.
- Minor: one filler cluster at 0:31 — jump-cut suggested.
Scheduler
Queue a week of posts in a minute
Pick a clip, a platform and a time. Latch checks length, aspect ratio, caption limits and hashtag count against each platform's current rules before it saves — and tells you exactly what to fix.
- YouTube Shorts, TikTok, Reels, LinkedIn and X adapters.
- Dry-run mode stores the plan without publishing, so you can review.
- Per-platform captions, or one caption for all.
MonTikTok09:00YouTube17:30
TueLinkedIn12:00
WedInstagram08:30
ThuX18:00
FriTikTok09:00
SatYouTube11:00
Sun
Instagram Reels: caption is 2,340 chars (max 2,200). Trim before publishing.
Compare
Latch vs Opus Clip
A straight feature comparison. Where they are ahead, we say so.
† Opus Clip facts per Opus public pricing and feature pages, September 2026. Plans and limits change; check their site for current details. Opus Clip is a trademark of its owner; Latch is not affiliated.
Pricing
Free forever on your hardware. Flat plans in the cloud.
No per-minute credits. Cloud plans are not live yet — join the waitlist and we'll email you when they open.
Self-host
Available nowFreeMIT-licensed. Your GPU, your keys, no limits.
- Unlimited source minutes
- Every feature, every export format
- CLI + REST API + SSE progress
- Bring your own Whisper / LLM / stock keys
- Community support on GitHub
Cloud Creator
Coming soon$12/ monthFor one channel that posts every day.
- 600 source minutes / month
- 1080p exports, no watermark
- 3 brand kits
- All caption presets & languages
- Email support
Most complete Cloud Pro
Coming soon$24/ monthFor creators and small teams who want it all.
- 1,500 source minutes / month
- 4K exports
- AI B-roll + scheduler
- Team: 3 seats
- API access
Team
Coming soon$59/ monthFor agencies running many channels.
- 5,000 source minutes / month
- 10 seats
- SSO
- Priority render queue
- Everything in Pro
Cloud prices shown are planned launch prices in USD and may change before launch. Source minutes reset monthly and do not roll over.
FAQ
Questions people ask before they clone the repo
How accurate are the clips?
Candidates always snap to sentence boundaries, so you never get a clip that starts mid-word. Ranking is heuristic by default and improves noticeably when you add an LLM key. In our internal test set of podcast episodes, the top-3 picks matched a human editor's picks roughly two times out of three — good enough to save hours, not good enough to skip a 30-second review.
Which languages are supported?
Transcription and captions run on Whisper, which handles 30+ languages well (English, Spanish, Portuguese, German, French, Italian, Dutch, Japanese, Korean, Mandarin, Hindi and more). Titles, hooks and descriptions are generated in the transcript's language when an LLM key is present.
What hardware do I need to self-host?
ffmpeg and Python 3.11+. Any modern CPU works; a GPU with 8 GB+ VRAM makes Whisper large-v3 roughly 10× faster. Face tracking uses OpenCV/MediaPipe and is optional — without it, reframe falls back to center or static modes.
What happens to my videos and data?
Self-hosted: nothing leaves your machine except calls you configure (your LLM provider, your stock-footage provider). Cloud (when it launches): media is stored encrypted, processed in an isolated worker, and deleted 30 days after the project is deleted. We don't train on your content.
How does the virality score work?
Six sub-scores from 0 to 100 — hook (how the first seconds grab), flow (self-contained, no dangling references), value (concrete takeaways), trend (topic momentum), emotion (tonal shifts, laughter, emphasis) and clarity (speech rate, filler density). Each score comes with written reasons that cite timestamps, so you can check them. The total is a weighted blend you can tune.
Can I finish the edit in Premiere or Final Cut?
Yes. Export a Premiere Pro XML, a Final Cut Pro FCPXML or an EDL for the whole project. Cuts, reframe keyframes and caption timing are included where the format supports them.
Is there a watermark?
Never on self-host. Cloud Creator and up are also watermark-free. There is no free cloud tier planned, precisely so we never have to watermark.
How is this different from Opus Clip?
The engine is open source (MIT) and runs on your own hardware with your own API keys, so there are no per-minute credits. The score is explained rather than a single number, every cut is listed and reversible, timeline export works on every plan, and the REST API is included. Opus has a longer track record and a hosted product today; our cloud plans are still on a waitlist.
Your next fifty clips are already in your last recording.
Paste a link and see ranked candidates in minutes — or clone the repo and run the whole thing on your own machine tonight.