FlowClip

Local, private clip editor for making viral shorts — TikTok, YouTube Shorts, Reels.

Sibling project to FlowLocal; same idea, for video instead of voice: everything runs on your machine, nothing is uploaded, nothing is metered, no subscription.

Download

Grab the latest installer from the Releases page.

FFmpeg is bundled, so there is nothing else to install. Windows may warn that the publisher is unknown — the installer is not code signed. Choose More info → Run anyway if you are happy to.

Early build. Multi-clip editing, layout compositing, image overlays, captions, music, text and colour all work end to end. Stickers and transitions are not built yet.

What it does today

Edit a project, not just a clip. Several clips sit end to end on one timeline with a single playhead running across all of them, so you can watch the piece through and judge whether a cut lands. Trim by dragging a clip's edges, split it at the playhead with S, and reorder or remove clips in place. Every clip keeps its own trim, layout and captions, because two recordings rarely want the same treatment. Ctrl+Z undoes anything, and projects save to a small file that stores the edit rather than the footage.

Layout — the interesting part. Rather than cropping a 16:9 gameplay clip down to a vertical slice and throwing most of the frame away, you compose the output from regions: take the facecam from one corner, the gameplay from the middle, the minimap from another, and stack them into a 9:16 canvas. Each layer has:

• Fill or Fit — crop the layer to its slot, or keep the whole frame intact
• Blurred backdrop — fitted layers sit on a blurred copy of themselves instead of black bars
• Rounded corners and borders
• Reorderable z-order, snapping to edges and centre lines, and arrow-key nudging

A layer can also draw from an image instead of the footage — a logo, watermark or badge. Image layers are always contained and never padded or backed, so a transparent PNG stays transparent rather than arriving in a black box.

Layout presets solve themselves against the canvas you pick: a stack on 9:16, a larger gameplay band on 1:1, and full-screen gameplay with picture-in-picture insets on 16:9.

The output preview is a canvas replaying the same maths FFmpeg will apply, so what you arrange is what renders.

Captions — transcribed on your machine with whisper.cpp. Nothing is uploaded, and there is no per-minute charge. You get word-level timings, so the caption pops word by word rather than sitting there as a block of subtitle. Four styles, a colour picker for the spoken word, and the caption block is dragged into place on the output like any other layer.

The transcript is editable line by line — rewrite a line and it keeps its timing, with the words re-spaced across it. You can also write captions from scratch: Add line drops one at the playhead, so a clip with no speech at all can still be captioned.

Voice activity detection is on by default. Without it whisper spreads word timings across silence — given three seconds of quiet before speech it puts the first word at 0:00, so the opening caption sits on screen before anyone talks.

Music and effects — drop in a bed that plays under the whole video, and effects that land wherever the playhead is. Music is mixed after the clips are joined, so it carries across cuts rather than stopping at the end of whichever clip owned it, and the video stream is copied through that pass so nothing is re-encoded. Everything plays in the preview too, since a mix you cannot hear is a mix you cannot judge.

Text — hooks, punchlines and labels, each with its own time range and dragged into place on the output. Outlined or on a filled panel, in whichever font and colour. Text is timed against the finished piece rather than a clip, so a hook can run across a cut.

Adjust — every audio level in one place, so a clip can be balanced against its music instead of guessing from two separate panels. Plus per-clip colour: five looks, or brightness, contrast and saturation by hand. Deliberately limited to what maps onto both FFmpeg's eq filter and a canvas filter, so the preview cannot drift from the export.

Roadmap

Milestone	What lands	Status
M1	Load · preview · trim · reframe · export	✅
M1.5	Multi-region layout compositor, packaged installer	✅
M2	Local Whisper captions, styled and editable	✅
M3	Multi-clip timeline, split, undo, project files	✅
M4	Images and logos as layers	✅
M5	Music and sound effects	✅
M6	Text overlays, per-clip audio and colour	✅
M7	Stickers, transitions, effects	⬜

Shortcuts

Key	Action
Space	Play / pause (loops inside the trimmed selection)
I / O	Set clip in / out point at the playhead
← →	Step one frame — or nudge the selected box in Layout
Shift + arrows	Jump 5 seconds — or nudge a box further
Shift + drag	Resize a layout box keeping its shape
S	Split the clip at the playhead
Ctrl + Z / Ctrl + Shift + Z	Undo / redo
Ctrl + S / Ctrl + O	Save / open a project
Enter	Commit a caption line you are editing

How it works

• Electron + React + TypeScript for the shell and UI.
• FFmpeg does all cutting, scaling, compositing and encoding. It is bundled and spawned from the main process; the renderer never touches it.
• Layout export builds a split / crop / scale / overlay filter graph, one branch per region, composited onto a generated canvas.
• Rounded corners come from a generated alpha mask that is rendered once and cached — the per-pixel expression that draws one is far too slow to run per frame.
• whisper.cpp transcribes locally; captions are burned in as an ASS track, one event per word so the timing is exactly what whisper reported. A word stays on screen until the next one begins, otherwise the line blinks out during every pause between words.
• Multi-clip export renders each clip through the same composite path and joins the results. Rendering separately is what lets every clip keep its own layout and captions, and since the segments share a canvas, frame rate and audio format, the join is a stream copy rather than a second encode. Silent sources get generated silence, or they cannot be joined at all.
• Project time and clip-local time are different things — the timeline runs 0..total while each video element still seeks in its own source — so the conversion lives in one place (src/shared/timeline.ts) rather than being re-derived per component.
• The renderer has no Node access; it talks to the main process over a narrow preload bridge (src/preload/index.ts).

Development

npm install
npm run setup:ffmpeg
npm run setup:whisper
npm run dev

setup:ffmpeg fetches a static FFmpeg build into resources/bin so nothing needs installing system-wide. If FFmpeg is already on your PATH, FlowClip uses that instead.

setup:whisper fetches whisper.cpp into resources/whisper and prunes it — the release archive is 68 MB of server, streaming and test binaries plus a 49 MB OpenBLAS that made no measurable difference here (1.60s vs 1.63s on an 11s clip), so only what whisper-cli loads is kept, taking it under 10 MB. Speech models are downloaded on first use rather than bundled.

Building the installer

npm run dist

Output lands in release/. Two things that will bite you on Windows:

• Close any running FlowClip first. A running copy holds d3dcompiler_47.dll open, and the build fails with Access is denied while clearing release/win-unpacked.
• win.signAndEditExecutable is off deliberately. electron-builder otherwise downloads a toolchain whose archive contains macOS symlinks, which Windows refuses to create without Developer Mode or elevation — and it re-downloads to a fresh directory each attempt, so a partial cache never helps. scripts/after-pack.cjs stamps the icon and version metadata with rcedit instead, which needs no download.

The multi-size .ico is built by scripts/make-ico.mjs rather than an image library, because generated ICOs are routinely rejected by Explorer when the directory entries are subtly wrong.

Licence

FlowClip's own source is MIT.

The installer redistributes FFmpeg under the GPL v3. FlowClip invokes it as a separate executable and contains no FFmpeg code, so the two remain separate works. The GPL build is used because H.264 export depends on libx264, which LGPL builds omit. Source and build scripts: FFmpeg, BtbN/FFmpeg-Builds. Full details in THIRD-PARTY-NOTICES.md.