ῥαψῳδός · the song-stitcher

Take a piece of one thing.
Lay it on another.

A self-hosted video editor. Cut seconds from a link in the browser, or edit the whole film in a native editor that plays 4K in real time.

Get the source Run it locally

Rhapsode Pro: a graded, scene-cut sequence with a picture-in-picture, English and French subtitles, scopes and loudness meters

quick mode · the browser

Two pieces, one decision

A base, a clip on top, and how they meet: a dub, a picture-in-picture or a stack, with captions. Works on a phone.

Any link yt-dlp reads, or an upload. Long videos are fetched as a window around the moment you point at.
Frame-exact trimming against a proxy cut from the same file the render reads.
Stems. One video's voice over another's song, split on your GPU.
Share links that play inline in chat apps, and remix from any render.

The cut view: trimmers, a stage with a picture-in-picture, a timeline and the mode panel

pro · the native editor

The whole film, on your machine

C++ on FFmpeg and Vulkan. Ripple, roll, slip and slide; multicam synced by sound; Premiere's keys or a tighter set. Nothing needs an account or the network, the speech model included.

Colour wheels and a vectorscope beside the program monitor

Colour. Linear-light compositing, wheels, curves, LUTs, OpenColorIO with ACES 2.0, PQ and HLG, waveform, parade and vectorscope.
Sound. A keyframed mixer with track EQ, compression and limiting, a BS.1770 loudness meter, normalise, duck, noise reduction and voice isolation.
Delivery. YouTube, social, AV1, HDR10, ProRes and DNxHR presets on NVENC, a background queue, and smart render that re-encodes only what changed.
Interchange. OTIO, FCPXML, EDL and Premiere XML in and out. A quick-mode render opens as a Pro project; a Pro export publishes back as a share link.

The transcript panel in paragraphs with timecodes, beside the program monitor

edit by text

Delete the words, cut the film

Local transcription with word times. Select sentences and delete them, strip filler words and pauses, then turn the words into captions.

measured

Numbers from one machine

An RTX 4070 SUPER, the GPU idle before each run. Method and raw results in BENCHMARKS.md.

60 fps	three tracks of 4K60 HEVC 10-bit, graded and transformed, with no dropped frame in 60 s
86 fps	the same timeline unthrottled; Kdenlive's engine renders it at 15.6
26 ms	a seek on the editor's proxies (382 ms on long-GOP originals)
141×	real-time speed of transcription on the GPU
27 s	a 19-minute talk transcribed, cut to a quarter, captioned and exported
10 steps	to cut an interview with captions; Premiere's documented workflow takes 20

run it

One repository

The web app is one Node package; the editor is a CMake project beside it. Viewing renders is public, creating needs the key in .env.

Web: Node 24, ffmpeg and yt-dlp. Pro: Linux with a Vulkan GPU, SDL 3, FFmpeg 8. NVENC and NVDEC are used when present. MIT.

git clone https://github.com/ZahakJ/rhapsode
cd rhapsode && npm ci && cp .env.example .env
npm run build && npm start            # :8013

cmake -S pro -B pro/build
cmake --build pro/build -j
pro/build/bin/rhapsode-pro            # the editor

Take a piece of one thing.Lay it on another.