Mindown

A native macOS video and audio downloader with a deliberately gaudy, classic-Winamp UI and a built-in AI assistant.

Powered by yt-dlp, a static ffmpeg, and any OpenAI-compatible chat model you point it at.

Highlights

• Paste a video or audio URL → pick format & quality → hit DOWNLOAD.
• Single-line playlist queue with an inline progress fill, up to 5 parallel downloads.
• AI ASSIST sidebar — chat-driven downloads. Ask "anti-hero by taylor swift as mp3" or "top 5 radiohead songs" and the model finds the right URL, format, and quality, then asks you to approve before queuing.
• Auto ID3 / iTunes tagging for every audio download — title, artist, album, album_artist, composer, year, genre, track number, copyright, full lyrics (USLT), and a 1200×1200 cover, sourced from Apple's iTunes Search API and LRCLib.
• SITES window — searchable list of every site yt-dlp supports (1900+), bundled offline so it works without a network round trip.
• Bundle ships with yt-dlp and a static ffmpeg matching your Mac's architecture, so end users do not need to install anything from the terminal.
• Custom Winamp-style chrome: beveled panels, green LCD readouts, dark grey gradients, classic monospaced bitmap-feel typography.

Supported formats and qualities

• Video — MP4 / WEBM / MKV at BEST / 2160p / 1440p / 1080p / 720p / 480p / 360p.
• Audio — MP3 / M4A / OPUS / WAV at BEST / 320 / 256 / 192 / 128 kbps.

AI assistant

The right sidebar runs an OpenAI-compatible chat completions session. Configure once with your API key, base URL, and model, and it becomes a music-download front-end for the rest of the app:

• The model is allowed to call two tools — search_youtube(query, limit) and propose_downloads(items[]).
• search_youtube runs the bundled yt-dlp ytsearch: and returns up to 10 candidates. The model picks the official version (skipping covers and reactions).
• propose_downloads surfaces an inline approval card. You tick or un-tick individual items and click APPROVE before anything is queued — nothing downloads automatically.

Works with anything that exposes /chat/completions — OpenAI, OpenRouter, Groq, Together, Anthropic via gateway, local Ollama in OpenAI mode, etc.

Music metadata enrichment

Every audio download (mp3 / m4a / opus / wav) automatically goes through a post-download tagging pass so files land in your library already labelled and ready for Music.app, Plex, Doppler, etc. Video downloads are skipped.

Field	Source	ID3 / iTunes mapping
Title	iTunes Search · trackName	TIT2 / ©nam
Artist	iTunes Search · artistName	TPE1 / ©ART
Album	iTunes Search · collectionName	TALB / ©alb
Album artist	= artist	TPE2 / aART
Composer	falls back to performing artist	TCOM / ©wrt
Year	iTunes Search · releaseDate (year prefix)	TYER / ©day
Genre	iTunes Search · primaryGenreName	TCON / ©gen
Track number	iTunes Search · trackNumber/trackCount	TRCK / trkn
Disc number	iTunes Search · discNumber/discCount	TPOS / disk
Copyright	iTunes album lookup · copyright	TCOP / cprt
Cover art	iTunes (1200×1200 upscale of artworkUrl100)	APIC / covr
Lyrics	LRCLib · plainLyrics	USLT (lang=eng) / ©lyr
Bitrate / sample rate / duration / channels / codec	embedded by ffmpeg from the audio stream itself	container-intrinsic

Implementation:

• The lookups (album-copyright, lyrics, cover download) run in parallel via Swift async let, so total wall time is dominated by whichever is slowest — typically ~1s.
• The matched track and a 1200×1200 cover are passed to the bundled ffmpeg, which rewrites the file in place with -id3v2_version 3 -write_id3v1 1 for mp3 and the equivalent atom-based metadata for m4a / opus / wav.
• The post-download status badge in each playlist row goes from TAG… (amber) to ♪ (green) when tagging succeeds, or TAG✗ (red) on failure. Hover the row for the full match summary or error message.
• Toggle the whole feature with the iTunes TAG switch in PREFS.

iTunes Search and LRCLib are both free, public, no-auth APIs.

Build & run

./build.sh
open Mindown.app

The build script:

1. Downloads yt-dlp_macos from the official yt-dlp release.
2. Downloads a static ffmpeg matching your Mac's architecture (darwin-arm64 on Apple Silicon, darwin-x64 on Intel).
3. Downloads the latest supportedsites.md so the SITES window works offline.
4. Caches all three under .bin-cache/ so subsequent builds skip the downloads.
5. Compiles the Swift app, assembles Mindown.app, copies the binaries into Contents/Resources/bin/, and ad-hoc-signs everything.

The resulting .app is fully self-contained — drop it anywhere, no terminal install needed for end users.

Develop

swift build           # debug
swift run Mindown     # run from the terminal (no .app bundle)
swift build -c release

Standard SwiftUI macOS app, single executable target, no third-party Swift dependencies.

Project layout

Sources/Mindown/
├── App.swift                 // @main entry, window styling
├── ContentView.swift         // Main UI — input row, queue, footer, sidebar
├── SettingsView.swift        // PREFS sheet (download dir, binary paths)
├── SupportedSitesView.swift  // SITES window (offline list + filter)
├── WinampStyle.swift         // Palette, beveled panels, LCD text, etc.
├── DownloadModels.swift      // MediaFormat / Quality / DownloadItem
├── DownloadManager.swift     // yt-dlp Process orchestration + progress
├── MetadataEnricher.swift    // iTunes Search + LRCLib → ID3 / iTunes tags
├── AppSettings.swift         // UserDefaults-backed preferences
├── AISettings.swift          // API key / base URL / model / sidebar
├── ChatModels.swift          // ChatMessage / Proposal / ProposedDownload
├── AIClient.swift            // OpenAI-compatible chat completions client
├── AITools.swift             // Tool definitions + host-side runners
├── AIChatView.swift          // Sidebar UI (form + chat + approval cards)
└── ChatViewModel.swift       // @MainActor orchestrator + tool-call loop

Configuration

• PREFS sheet (footer button): override download directory, yt-dlp path, ffmpeg path, and toggle the iTunes TAG metadata enrichment pass. Bundled binaries are picked up automatically on fresh installs.
• AI ASSIST sidebar: API key, base URL, model. Defaults to OpenAI's gpt-4o-mini against https://api.openai.com/v1.
• All values persist via UserDefaults.

Notes

• This is not sandboxed or notarised — it is a personal-use tool. On first launch on another Mac, right-click → Open if Gatekeeper grumbles.
• The app uses yt-dlp --progress-template to parse percent / speed / ETA / total size from a structured pipe-delimited line, which is more reliable than scraping [download] console output.
• Output files are named %(title).200B [%(id)s].%(ext)s so duplicate titles do not clobber each other.

License

Personal-use utility, no warranty. yt-dlp and ffmpeg are redistributed under their respective licenses (Unlicense and LGPL/GPL respectively).