Nagare (流れ)

Pronounced "Nah-gah-reh" — subtitle mining companion for Emby, Jellyfin, Plex, and AudioBookShelf.

🎬 See it in Action

The subtitle timeline — mine any line from the current or past session directly in the browser. Highlighting via https://jiten.moe/reader

---

Anki Enhancement Dialogue allowing for tight control over what we mine

---

card_example.mp4

Finished Card (Kiku Notetype)

---

What does it do?

Nagare watches your active media server playback sessions, displays a live subtitle timeline in the browser, and enriches Anki cards with sentence audio, screenshots, and source metadata — without interrupting your immersion.

Note: This project is my most vibe-coded yet, so YMMV. It's really a problem that I sought out to solve for myself, but I believe/hope it can be useful for others.

---

Features

• Live subtitle timeline synced to playback
• Sentence audio extraction and animated AVIF screenshot clips
• AnkiConnect integration with automatic card matching
• Playback controls (seek, pause, resume) from the browser.
• Yomitan-aware pause behavior. (Must turn off Secure Popup in Yomitan)
• Watch history for mining after playback ends
• Session card review with saved SRT snapshots, review progress, context expansion, and audio previews
• Multi-server support (Emby + Jellyfin + Plex + AudioBookShelf simultaneously)
• Manual-review, daily, or automatic Tadoku listening-log sync, grouped by show with duplicate protection

Roadmap

• Initial prototype with Emby support
• Add Jellyfin support
• Add Plex support
• Add AudioBookShelf support for local MP3/M4B sidecar subtitles
• AnkiConnect integration
• Support for subtitles even when player has none (listening practice while maintaining mineability).
• Mining History, allowing you to touch up cards after the fact, or add more context.
• Session History, allowing you to load past sessions and mine from them.
• Manual Subtitle Offset
• Automatic Subtitle Sync? IDK if this is even feasible, the ability to press a button, Nagare syncs with alass or subplz, and then sends the updated sub to the media server would be the idea.
• More Active Subtitle Sync? If you change subtitle timing in media player, Nagare will not adjust. I doubt this is possible...
• More options for audio/ss formats

Installation

Docker (recommended)

1. Run with Docker Compose:

# docker-compose.yml
services:
  nagare:
    image: ghcr.io/bpwhelan/nagare:latest
    container_name: nagare
    ports:
      - "9470:9470"
    volumes:
      - ./data:/app/data
      # Optional: mount media library for disk-mode access
      # - /path/to/anime:/media/Anime:ro
    extra_hosts:
      - "host.docker.internal:host-gateway"
    restart: unless-stopped

docker compose up -d

2. Open http://localhost:9470 and configure Nagare from the web UI Config page.

Binary release

Download the latest binary for your platform from GitHub Releases.

Requirements:

• ffmpeg on PATH
• Anki with AnkiConnect

./nagare

The web UI is served at http://localhost:9470.

Build from source

cd frontend && npm ci && npm run build && cd ..
cargo build --release

Configuration

All configuration is managed through the web UI Config page and stored in data/nagare.sqlite. On first run, configure:

1. Media server — URL and API key (Emby/Jellyfin), token (Plex), or admin token (AudioBookShelf)
2. AnkiConnect — URL and field mappings (Sentence, SentenceAudio, Picture)
3. Media access — auto, disk, or api mode; add path mappings if server and Nagare see different file paths
4. Tadoku (optional) — save your Tadoku username and password, then choose manual review, daily sync, or automatic sync. Nagare signs in and refreshes the browser session automatically. Manual review lets you approve or permanently decline individual ready episodes; daily sync defaults to 8 PM Eastern. Automatic sync checks every five minutes and immediately after a playing item unloads. Recent completed episodes, including short anime episodes, sync promptly when they have audio in the configured target language. AudioBookShelf items longer than two hours use the separate high-density long-form path, which requires at least 30 minutes of new playback followed by 30–60 minutes of inactivity for automatic sync. Manual review also offers the current uncredited playback of these long audiobooks as a checkpoint when automatic sync misses them; other media use standard logs. This keeps uninterrupted listening together and prevents stale or wrong-language history from being exported. Items excluded by these automatic rules remain available for manual review. When the review workflow is first enabled, episodes completed after the previous successful sync are queued. Tadoku tags can also be assigned from case-insensitive file-path matches; by default, paths containing anime receive the anime tag.

AudioBookShelf downloaded playback

AudioBookShelf does not expose an open playback session while its Android app plays a downloaded book. Instead, the app writes a Local listening-session row when playback is paused. Nagare polls that listening history every three seconds, resolves the row's library item to the exact server-side MP3/M4B track, and loads a same-basename sidecar subtitle such as Book 01.srt through the configured path mappings. When no same-basename subtitle exists, Nagare adds the other SRT files in that directory to the track selector in fuzzy filename-match order. The detected paused session remains available in Nagare for 15 minutes after its last AudioBookShelf update.

This discovery uses the administrator-only /api/sessions endpoint, so the configured AudioBookShelf token must belong to an administrator. Nagare monitors the newest listening row for each selected user; a newer streamed row replaces an older downloaded one.

How it works

1. Nagare polls your media server(s) for active playback sessions
2. Select a session or let Nagare choose automatically and stay with that playback
3. Create a card in Anki — Nagare matches it to the exact subtitle context
4. Confirm the match, preview audio/screenshot, and enrich the card

When browser tabs share a media-server session ID, Nagare keeps their different items as separate entries in the session picker. Each item remains available for up to 30 seconds between check-ins, so paused tabs do not repeatedly replace one another. Choose an entry by its title to keep mining that item. Newly started playback can take over in automatic mode, and disconnected or unloaded sessions are removed. Plex playback-instance IDs also keep separate tabs playing the same item from mixing their progress updates.

Remote playback controls are unavailable while several items share the same server control ID, because that ID cannot reliably address an individual tab. Controls return when only one entry remains.

Review a mining session later

Open History → Card sessions, or choose Review cards on a watch-history item. Each session keeps the subtitle snapshot and audio track used when its cards were detected. You can search or filter cards, inspect their saved Anki fields, add surrounding subtitle lines, edit the sentence/translation, preview audio and frames, and save a revised clip to Anki. Reviewed & next saves progress across reloads and restarts. Unsaved edits stay available when switching between cards in the workspace.

Automatically enhance new cards uses the existing browser preference and keeps working while that browser is open, including on the review page. Cards are recorded before enhancement completes, so skipped cards, failed attempts, and cards that already had media remain available for review. Failures require a deliberate retry. Changing the media, subtitle/audio track, playback session, or returning after 30 minutes without mining starts another card session. Earlier enhanced notes are imported by title with the subtitle history available at upgrade time; their original session boundaries cannot be reconstructed.

AnkiBeacon full payloads take the fastest notification path: they do not wait for fallback polling or optional card-ID lookups. ID-only payloads still need AnkiConnect metadata, but those lookups run separately. The UI distinguishes Card received from the confirmation that enhancement finished.

Project structure

src/            Rust backend (Axum + Tokio)
frontend/       Svelte frontend (Vite)
Dockerfile      Multi-stage container build

Data is stored in data/nagare.sqlite. Generated Anki media files are prefixed with nagare_.