New advertisers: 50% off your first month with code FIRSTMONTHSee slots

Cloud GPU video editing that offloads voice cleanup and heavy rendering from a local machine.

version
1.4.1
GitHub stars
32
View on GitHub

What it does

  • Adds 5 skills (clean-voice-sidon, add-intro-outro, find-script, save-script, complementary-plugins) for remote video and audio jobs
  • Runs the voice-track cleanup on a Modal GPU with Sidon, then loudness-normalises and remuxes without re-encoding the video
  • Enforces a 30-second test run before full jobs and uses two-pass loudness normalisation to -14 LUFS by default
  • Designed for laptops without a GPU, with light ffmpeg work kept local and requiring a Modal account

Video editing and rendering on rented cloud GPUs (Modal): offload heavy jobs like AI voice cleanup from a local machine.

Add marketplace

/plugin marketplace add danielrosehill/Claude-Code-Plugins

Run it inside Claude Code once, then install the plugin below.

Install

/plugin install remote-video-editing@danielrosehill

Run it inside Claude Code after adding the marketplace.

Files

Discovered from source

Add the intro and outro

The stings are short MP4s with their own jingle, one pair per orientation, listed in preferences. This skill normalises everything to one format, joins intro, video and outro with short cross-fades (video and audio), and renders the result.

Preferences

remote-video-editing/preferences.json (lookup order as in clean-voice-sidon):

KeyMeaning
stings{"intro_16x9": url, "outro_16x9": url, "intro_9x16": url, "outro_9x16": url}; URLs or local paths
sting_fade_sCross-fade length, default 0.5
render_notesWhere to run long encodes (e.g. a server rather than a laptop). Follow them.

If stings is not set, ask for the files. The plugin does not ship any brand assets.

Orientation

Read the displayed orientation, not the coded size: phones store portrait video as 1920x1080 with a rotation tag. Check with ffprobe -v error -select_streams v:0 -show_entries stream=width,height:stream_side_data=rotation -of compact IN. A rotation of 90 or -90 swaps width and height. ffmpeg applies the rotation automatically when it decodes.

  • Portrait (9:16) video: use intro_9x16 / outro_9x16, canvas 1080x1920.
  • Landscape: use the 16:9 pair, canvas 1920x1080.
  • 16:9 cut of a portrait video (for YouTube's main feed): canvas 1920x1080, the video scaled to 1080 high in the centre, over a blurred, darkened, zoomed copy of itself; use the 16:9 stings. The [v1] chain below becomes:
    [1:v]fps=30,split[a][b];
    [a]scale=1920:1080:force_original_aspect_ratio=increase,crop=1920:1080,scale=480:270,boxblur=20:2,scale=1920:1080,eq=brightness=-0.06:saturation=0.9[bg];
    [b]scale=-2:1080[fg];[bg][fg]overlay=(W-w)/2:0,setsar=1,format=yuv420p[v1]
    
    Test it on an 8-second cut (-ss 60 -t 8) and show a frame before rendering the whole thing.

Render

W:H is the canvas, F the fade (0.5), DI the intro duration, DV the video duration (ffprobe format=duration). Offsets: first fade at DI-F; second at DI+DV-2*F.

ffmpeg -y -i intro.mp4 -i main.mp4 -i outro.mp4 -filter_complex "
 [0:v]fps=30,scale=W:H,setsar=1,format=yuv420p[v0];
 [1:v]fps=30,scale=W:H:force_original_aspect_ratio=decrease,pad=W:H:(ow-iw)/2:(oh-ih)/2,setsar=1,format=yuv420p[v1];
 [2:v]fps=30,scale=W:H,setsar=1,format=yuv420p[v2];
 [0:a]aresample=48000,aformat=channel_layouts=stereo[a0];
 [1:a]aresample=48000,aformat=channel_layouts=stereo[a1];
 [2:a]aresample=48000,aformat=channel_layouts=stereo[a2];
 [v0][v1]xfade=transition=fade:duration=F:offset=DI-F[v01];
 [v01][v2]xfade=transition=fade:duration=F:offset=DI+DV-2*F[v];
 [a0][a1]acrossfade=d=F[a01];[a01][a2]acrossfade=d=F[a]" \
 -map "[v]" -map "[a]" -c:v libx264 -preset medium -crf 20 -maxrate 10M -bufsize 20M -c:a aac -b:a 192k -movflags +faststart out.mp4
  • Several source clips: join them first (same filter pattern, or the concat demuxer if they share a codec), then treat the result as main.mp4.
  • This is a full re-encode: minutes of every core for a 10-minute 1080p video. Run it on Modal, not locally and not on a home server: put the arguments after the inputs (from -filter_complex to +faststart) one per line in render.args, then $MODAL run "${CLAUDE_PLUGIN_ROOT}/scripts/ffmpeg_modal.py" --inputs intro.mp4,main.mp4,outro.mp4 --args-file render.args --out out.mp4 (modal_command and run_notes from preferences say where Modal runs). Join several clips in the same graph (concat filter) rather than as a separate pass. Delete the job's folder in volume video-editing-work afterwards.
  • xfade needs both inputs on the same timebase: put fps=30,settb=1/30 on every video branch (including after a concat), or ffmpeg fails with "First input link main timebase (1/30) do not match".
  • Bitrate: keep the cap (-crf 20 -maxrate 10M -bufsize 20M). Plain CRF 18 on noisy phone footage came out at 24 Mbps, about 2 GB for 11 minutes (2026-11-13). YouTube re-encodes anyway and suggests 8-12 Mbps for 1080p uploads.
  • Verify: output duration is about DI + DV + DO - 2*F, and check frames at DI-F/2 (mid-fade), the middle and the end.

Handheld footage: stabilise in the same render

Phone videos shot handheld benefit from light stabilisation (the user, 2026-11-13, after the first Hebrish weekend update went out without it). Use ffmpeg's vid.stab (libvidstab; Debian's ffmpeg, which the Modal image installs, has it; check with ffmpeg -filters | grep vidstab). Two passes, both on Modal:

  1. Detect (fast, analysis only, writes a transforms file), per source clip: ffmpeg -i main.mp4 -vf vidstabdetect=shakiness=5:accuracy=15:result=main.trf -f null - Run it as its own ffmpeg_modal.py job (args -vf / vidstabdetect=...:result=/work/<job>/main.trf / -f / null, output -), or add a detect step to the script. The .trf must sit beside the input in the volume.
  2. Transform inside the main render, first thing on that clip's video branch, before grade and scale: vidstabtransform=input=main.trf:smoothing=20:zoom=2:optzoom=0,unsharp=5:5:0.4 smoothing around 15-30 (frames) keeps natural motion; a fixed zoom=2 (percent) hides the moving borders without the pumping optzoom=1 causes on a talking head. The light unsharp offsets the resampling softness.

Test on an 8-second cut first and compare a still and a short clip with the user before the full render. If libvidstab is missing, deshake is a one-pass fallback (weaker). AI stabilisers are not needed for this.

Making stings

The stings in use were rendered from the brand SVGs with a small Pillow script plus a generated jingle. Find it with find-script (tags logo, sting, intro), and don't regenerate the logo with a video model: video models distort letterforms.

Readme

Remote Video Editing

Agent skills for video editing and rendering on rented cloud GPUs (Modal). The local machine does the light ffmpeg work; the GPU-heavy steps run remotely, so a laptop without a GPU can still use model-based tools.

Skills

  • clean-voice-sidon: clean up a video's voice track with the Sidon v0.1 speech-restoration model on a Modal GPU, then two-pass loudness-normalise and remux it without re-encoding the video. Tests on 30 s first.
  • add-intro-outro: put your intro and outro stings on a video (16:9 or 9:16 to match), with short fades; also a 16:9 blurred-background cut of a vertical video.
  • find-script: before writing a new job script, search your private script library for one to reuse or adapt.
  • save-script: after a job, save its script to the library with notes on what it was used for.
  • complementary-plugins: list and install the plugins that go with this one.

Complementary plugins

  • video-editing: local editing and rendering (Kdenlive, melt, ffmpeg render profiles, silence cutting, subtitles).

Requirements

  • ffmpeg locally.
  • A Modal account with the CLI set up (pip install modal or uv tool install modal, then modal setup).

Configuration (optional)

remote-video-editing/preferences.json under the first data root that exists: $CLAUDE_USER_DATA, ~/.claude-user-data, or ${XDG_DATA_HOME:-~/.local/share}/claude-plugins:

{
  "modal_command": "modal",
  "run_notes": "where and how to run Modal, if not on this machine",
  "loudness_lufs": -14,
  "true_peak_db": -1.5,
  "scripts_repo": "owner/Cloud-Video-Editing-Scripts",
  "scripts_repo_path": "~/repos/Cloud-Video-Editing-Scripts"
}

scripts_repo is a private GitHub repo used as a library of past job scripts (index.json catalogue plus scripts/<slug>/ folders); save-script offers to create it if it is not set. Requires gh and jq.

Installation

claude plugins marketplace add danielrosehill/Claude-Code-Plugins
claude plugins install remote-video-editing@danielrosehill

Details

Marketplace
danielrosehill/Claude-Code-Plugins
Source
danielrosehill/Remote-Video-Editing
Author
Daniel Rosehill
Category
Media Editing
License
MIT

More from danielrosehill/Claude-Code-Plugins

All 264

Similar to remote-video-editing

More Media & Audio/Video