Remember a scene but not where it is? Describe it the way you'd tell a friend, "when he admits the robot hand freaks him out", and Jellymeme finds it in your Jellyfin library, trims it to the line you want, burns in the subtitles or your own caption, and hands you a GIF, a clip or a still.

_{Made by Jellymeme from one sentence. Tears of Steel © Blender Foundation, CC BY 3.0.}

Two ways in:

• One scene. Describe a single moment — "when he sets his foot on the grill" — pick it from the results, trim it exactly, caption it, and export a GIF, a clip, or a still image. When one query answers more than once — every "never seen it" in the show — tick as many results as you want and they become the clips of one cut.
• A whole montage. Paste a wall of text — the replies from a "best scenes in this show" thread, a list you wrote yourself, whatever shape it arrives in — and every line becomes a clip, matched, ordered and editable.

Either way you land in the same editor: nudge the in and out points, swap a match it got wrong, burn in captions, export.

It talks to Jellyfin over its HTTP API and never needs access to your media files.

If Jellymeme is useful to you, starring the repo helps other people find it, and armanckeser.com/subscribe has ways to hear about new releases.

---

How it works

How it works

Jellyfin  ──subtitles──▶  parse → window → embed  ──▶  SQLite + sqlite-vec
   │                                                          │
   │                          paste ──▶ split ──▶ embed ──▶ KNN search
   │                                                          │
   └──media over HTTP range──▶  ffmpeg: clip → caption → concat → MP4/GIF

Indexing. For each episode, Jellyfin is asked for its subtitle track as SRT. It transcodes ASS/SSA and embedded MKV tracks server-side, so Jellymeme never runs ffmpeg to extract subtitles. Cues are cleaned (speaker labels, sound effects, markup) and grouped into overlapping dialogue windows — a single cue like "No." carries no meaning, and a described scene spans several lines. Each window is embedded locally with all-MiniLM-L6-v2 via transformers.js and stored in sqlite-vec alongside the text.

Matching. Pasted text is split into descriptions, embedded, and searched with KNN scoped to the selected title. Because windows overlap, a naive top-k returns five views of the same ten seconds — so results are de-duplicated by time to yield genuinely distinct moments. That's what makes "try another match" show you something new.

Rendering. ffmpeg reads each clip directly from Jellyfin's static stream endpoint using HTTP range requests, seeking straight to the timestamp rather than downloading the file. Each segment is normalised to identical parameters so they can be concatenated with a stream copy, and captions are burned with libass.

Why not a Jellyfin plugin?

A plugin lives inside your Jellyfin install and can never be shared with anyone who isn't on your server. Running as a sidecar keeps that door open, and keeps the ML and video work out of a .NET plugin sandbox that was never built for it.

Jellyfin does the parts it's good at

Job	Who does it	Why
Library, metadata, episode lists	Jellyfin	It already scanned and organised everything
Subtitle extraction and format conversion	Jellyfin	Stream.srt converts any text track server-side
Preview playback in the browser	Jellyfin transcoder	Chrome won't decode most library MKV/HEVC
Frame-accurate clip export	ffmpeg	Jellyfin's stream endpoint has startTimeTicks but no end
Still frames and GIF palettes	ffmpeg	Not something a media server exposes
Semantic search	Jellymeme	Jellyfin only does substring title search

Requirements

• Jellyfin 10.9+ (developed and verified against 10.11.7) with an API key
• Node.js 20+
• A text-based subtitle track on the content you want to index — SRT, ASS/SSA, or embedded MKV subs. Bitmap subtitles (PGS, VOBSUB) are image-based and are skipped; the UI reports exactly which episodes were skipped and why.
• ffmpeg is not required system-wide — a static binary ships via ffmpeg-static.
• A font must be available for burned captions. Most systems have one; see JELLYMEME_CAPTION_FONT below.

Getting started

npm install
npm run dev

Open http://localhost:3000, go to Settings, and enter your Jellyfin URL and an API key (Jellyfin → Dashboard → Advanced → API Keys). The key is stored in the local SQLite database and is never sent to the browser — artwork, previews and media all proxy through the server.

Then:

1. Search for a show or film and click Index. Indexing is resumable; each episode commits as it finishes.
2. Open the title. Pick One scene to describe a single moment and choose from the results — tick several results instead to cut them all together — or Whole montage to paste a list where every line becomes a clip.
3. Edit: adjust each clip's start and end in either direction, cycle alternate matches, search manually for the ones that landed wrong, caption the cut and then any clip that wants its own, size and place the text, rewrite individual subtitle lines, mute clips.
4. Export as MP4, GIF, WebM, or a single frame as PNG or JPG.

Trimming, captions and still images

Trimming

Each edge of a clip moves both ways. Positive extends the clip outward, negative pulls it inward — a matched dialogue window is often longer than the moment you actually want, so tightening matters as much as padding. Drag the slider, nudge with the ± buttons, or use the arrow keys for 250 ms steps.

Captions

Three modes: none, the real subtitles for that moment (re-fetched from Jellyfin at render time with their original timings), or your own custom text in meme style — uppercase, white with a heavy outline, top-centre. Captions are burned with libass, so they wrap properly and appear identically in a still and in the video of the same moment.

Set the mode once for the whole cut; any clip can then be given its own instead.

Size and placement. A slider from 50% to 300% of the default type size, and a choice of top, bottom, or wherever the mode normally puts it. The size is a multiplier rather than a point size, so captions stay the same relative size whether you export at 320px or 1080px. Set for the cut, and overridable per clip — a clip can be made larger while still following the cut's words.

Editing the real subtitles. With subtitles on, each clip lists the lines that fall inside it and every one can be rewritten. The timings stay the show's, so a changed word still lands on the frame where it was spoken — change one name across three clips, drop a line by emptying it, and leave the rest of the dialogue as recorded. Edits belong to the clip and survive switching caption modes back and forth. Re-indexing the title can move its cue timings, which is the one thing that will orphan an edit; the original line comes back if it does.

Still images

Choose PNG or JPG and a frame picker appears. The frame is grabbed from the finished cut, so any burned caption is baked in exactly as it appears in the video.

Configuration

Configuration

All optional.

Variable	Default	Purpose
JELLYMEME_DATA	./data	SQLite database and cached model weights
JELLYMEME_RENDERS	./renders	Finished exports
JELLYMEME_MODEL_CACHE	$JELLYMEME_DATA/models	Embedding model weights (~90 MB, downloaded once)
JELLYMEME_CAPTION_FONT	DejaVu Sans	Font family for burned captions
JELLYFIN_URL / JELLYFIN_API_KEY	—	Pre-seed the connection instead of using the UI
AUTH_PASSWORD	—	Ask for this password before any page or API call

Docker

One command, no clone (Compose 2.34+), amd64 or arm64:

docker compose -f oci://ghcr.io/armanckeser/jellymeme-stack:latest up -d

Then open http://<server>:8080 and connect Jellyfin in Settings. It needs no volume mounts for your media, only a volume for its own database and renders.

From your server's app store: add the catalog once, then install Jellymeme like any other app.

Server	Where	Paste
Cosmos	Market → Sources → Add Source	https://armanckeser.github.io/selfhost-apps/servapps.json
Umbrel	App Store → ⋯ → Community App Stores	https://github.com/armanckeser/selfhost-apps
CasaOS / ZimaOS	App Store → Add Source	https://github.com/armanckeser/selfhost-apps/archive/refs/heads/casaos.zip
Runtipi	Settings → App Stores → Add App Store	https://github.com/armanckeser/selfhost-apps
Portainer	Settings → App Templates → URL	https://armanckeser.github.io/selfhost-apps/portainer/templates.json

Unraid's Community Apps doesn't list it yet; selfhost-apps has the template to add by hand.

Set AUTH_PASSWORD to ask for a password before anything else, if anyone else on your network can reach the server. docker compose up -d from a checkout builds the image from source instead.

Performance, testing and known limitations

Performance

On four CPU cores, embedding runs at roughly 600 passages/second. A 22-minute episode produces ~200 dialogue windows, so a 200-episode series indexes in about a minute of embedding plus however long Jellyfin takes to serve the subtitles. No GPU required.

Rendering is dominated by ffmpeg encode time. A ten-clip montage at 720p takes a few seconds; the clips are read over HTTP but only the bytes around each timestamp are transferred.

Testing

npm test                      # unit tests: parsing, windowing, paste splitting
RUN_FFMPEG_TESTS=1 npm test   # adds the real ffmpeg pipeline and a full end-to-end run

The end-to-end suite stands up a mock Jellyfin server implementing endpoint shapes verified against a real 10.11.7 install, then indexes, searches, builds a montage and renders it — every layer except Jellyfin itself runs for real. The mock's limits, and which assertions they weaken, are in docs/ARCHITECTURE.md.

Known limitations

• Dialogue only. Search matches what characters say. A wordless visual gag or a reaction shot with no line cannot be found. Visual search would need frame embeddings (CLIP) and a GPU; the schema and pipeline are shaped so that could be added alongside rather than replacing this.
• No transcription. Content without a text subtitle track is skipped rather than transcribed. Adding Whisper for the gaps is the natural next step if your library has holes.
• Single user. There is no authentication. Don't expose it to the internet as-is.
• Seven npm audit advisories, all high, all transitive: Next.js and its own tree (next, postcss, sharp, nanoid) and transformers.js (onnxruntime-node → adm-zip). They cannot be resolved without forcing breaking upgrades, and onnxruntime-node is deliberately pinned below 1.24 for a crash reason described in docs/ARCHITECTURE.md.

Legal

Bring your own library. Jellymeme searches and cuts media you already have on your own Jellyfin server. It does not download, host, index or distribute anything, and nothing it reads leaves your machine except requests to your own Jellyfin server.

What you export is yours to answer for. A clip or a still is still part of someone else's film or show. Short excerpts for commentary, criticism or parody may be fair use (US) or fair dealing (UK, Canada and others), but that depends on the use and the jurisdiction and is not something this software can decide for you. Share what you have the right to share. This is not legal advice.

Not affiliated with Jellyfin. Jellymeme talks to Jellyfin's public HTTP API. It is not made, endorsed or supported by the Jellyfin project, and does not use the Jellyfin logo.

Third-party components

Component	Licence	How it arrives
FFmpeg	LGPL-2.1+ / GPL-2.0+ depending on build	Debian package in the Docker image; ffmpeg-static (GPL-3.0-or-later) for local dev
all-MiniLM-L6-v2	Apache-2.0	downloaded on first run, not in this repo
transformers.js	Apache-2.0	npm
sqlite-vec	MIT / Apache-2.0	npm
better-sqlite3, Next.js, React	MIT	npm

Media in this README comes from the Blender Foundation open movie Tears of Steel, © Blender Foundation, CC BY 3.0, run through Jellymeme unmodified apart from the trim and the burned-in official subtitles.

Test fixtures. The two comment threads under src/lib/montage/*.fixture.txt keep a real Reddit thread's layout — usernames, timestamps, vote counts, ads, reply buttons — because the parser has to survive it, but every name and comment in them is invented.

Licence

GNU AGPL v3, see LICENSE.