Youty — Transformador de Vídeos em Base de Conhecimento Local
legetdev/youty
Extraia, transcreva e indexe vídeos localmente para busca com IA
Ver no GitHubMCP server que converte vídeos do YouTube, Instagram e TikTok em base de conhecimento pessoal com Markdown + frames, 100% local e sem telemetria. Funciona integrado ao Claude via MCP, CLI ou app macOS, com busca semântica on-device e transcrição automática.
Instalação
npx convez-ai legetdev/youty1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Pesquisa e curação de conteúdo de vídeo para projetos
- Construção de base de conhecimento pessoal sem nuvem
- Busca semântica em bibliotecas de vídeos já baixados
- Integração com fluxos de trabalho Claude/Cursor para análise de conteúdo
Veredito

Ferramenta sólida para quem quer dominar vídeos como dados estruturados. O suporte MCP é bem pensado e a privacidade on-device é genuína — não há bait-and-switch de telemetria. Requer macOS 26+ para transcrição nativa; CLI + MCP funcionam em Python 3.11+.
README
A tradução para o português deste README ainda está na fila.
Youty
Save YouTube, Instagram, and TikTok videos to a local folder of Markdown notes + frame snapshots — searchable from your Mac, your shell, and any MCP-compatible AI.
Everything runs on your device. No accounts. No telemetry. Your vault is
just a folder of .md files and JPEGs you fully own.
Install
| How | Need | |
|---|---|---|
| Mac app | Download the DMG from Releases, drag to /Applications | macOS 26 (Tahoe) |
| CLI | brew install legetdev/youty/youty or ./Scripts/install-cli.sh from a clone | Xcode 26 to build |
| MCP server | uv tool install youty-mcp | Python 3.11+, uv |
The Mac app auto-updates via Sparkle once installed — Check for Updates… lives in the app menu (or Settings → About). New releases are EdDSA-signed; an unsigned binary will be rejected even if the appcast URL is hijacked.
The Mac app, the youty CLI, and the youty-mcp server all read and
write the same vault folder.
Using it
Mac app. First launch walks you through a four-card setup (pick a vault; AI search works on-device with no key; optional CLI install; optional MCP wiring). The cards stay reachable from Settings → Onboarding. After that: paste a video URL, hit Load → Save to Vault. The Mac app also shows up in:
- Safari / Notes / Mail's Share menu (Share → Save to Youty)
- The macOS Services menu (right-click any URL → Save to Youty Vault)
- Shortcuts.app + Siri (
Save URL to Youty Vaultand two read intents) - An optional menu bar icon (Settings → Integrations)
CLI.
youty save https://www.youtube.com/watch?v=...
youty list --platform youtube
youty search "ai influencers"
youty transcript <url-or-id>
youty reindex --text-only # re-embed the vault for on-device search
youty login instagram # one-time, only for Instagram saves
MCP. Add youty-mcp to your Claude Desktop / Cursor / Claude Code
config and your AI can semantically search the vault, retrieve
transcripts, and find frames. See youty-mcp/README.md.
What lands in the vault
{vault}/youtube/Channel - Title/
video.md # YAML frontmatter + transcript with timestamps
00000000.jpg # frame snapshots, filename = ms offset
00007560.jpg
...
manifest.json # one-line-per-video corpus index
YouTube transcripts are scraped from YouTube's caption panel. TikTok
uses platform auto-captions. Instagram and TikToks-without-captions are
transcribed on-device via Apple's SpeechAnalyzer + SpeechTranscriber
(macOS 26 only; audio never leaves your Mac).
Privacy
See docs/privacy.md. The short version: vault is
local, transcripts and frames are local, AI search is 100% on-device,
no telemetry. Text search uses Google's EmbeddingGemma converted to Core ML
and runs entirely on your Mac — no API key, no provider option, nothing
leaves the device. The only network calls are (a) platform fetches to
scrape the video, and (b) Sparkle's anonymous once-a-day check for a newer
Youty release.
Terms
See docs/terms.md. Same posture as yt-dlp and IINA:
Youty is a personal user-agent that loads pages you're already entitled
to view; you're responsible for whether your use of the saved content
complies with each platform's terms of service.
Building from source
The on-device Core ML models (~415 MB) aren't committed — they're fetched from
the matching GitHub Release. Run this once before building (it's a checksummed,
idempotent no-op once they're in place); install-cli.sh and release-app.sh
call it for you:
./Scripts/fetch-models.sh # models → Vendor/ (one-time)
xcodebuild -scheme youty -configuration Release build # Mac app
xcodebuild -scheme youty-cli -configuration Release build # CLI binary
./Scripts/make-dmg.sh # DMG installer (versioned filename)
./Scripts/build-mcp-wheel.sh # MCP wheel + twine check
./Scripts/smoke-test-extractors.sh # full smoke
Release pipeline (requires a Developer ID certificate + a notarytool keychain profile):
export DEVELOPER_ID_APPLICATION_CERT="Developer ID Application: ... (TEAMID)"
export NOTARY_KEYCHAIN_PROFILE="youty-notary"
./Scripts/release-app.sh # build → sign → notarize → staple
./Scripts/make-dmg.sh # signed versioned DMG
./Scripts/sparkle-sign-and-cut.sh build/Youty-*.dmg # emits the appcast <item>
One vetted Swift Package dependency in the Mac app (Sparkle, MIT,
auto-update — pinned in Package.resolved). FFmpeg ships statically
linked from Vendor/ffmpeg/. The MCP server is a separate Python
package with its own pinned dependencies.
Security
Found something concerning? See SECURITY.md.
Open source notice
Youty's Mac app links against FFmpeg 7.1.1 (libavcodec, libavformat,
libavutil, libswscale), built from source as statically-linked
archives via Scripts/build-ffmpeg.sh.
FFmpeg is licensed under the GNU Lesser General Public License v2.1+
(LGPL-2.1+) — verbatim text in
Vendor/ffmpeg/licenses/COPYING.LGPLv2.1.
The app also bundles Google's SigLIP-Base-Patch16-224 image encoder
(Apache License 2.0) converted to Core ML for frame embedding — verbatim
text in Vendor/siglip/licenses/LICENSE,
with attribution + conversion provenance in
Vendor/siglip/licenses/NOTICE.
The app also bundles Google's EmbeddingGemma-300m text encoder, converted
to Core ML for on-device AI search (the default — no API key). It is
distributed under the Gemma Terms of Use (not an OSI license;
commercial use + redistribution permitted with attribution) — the required
NOTICE + modification record is in
Vendor/embeddinggemma/licenses/EmbeddingGemma-NOTICE.txt,
and the canonical terms live at https://ai.google.dev/gemma/terms. A
from-scratch Swift BPE tokenizer reproduces Gemma's tokenizer with no
third-party dependency.
The app also bundles Sparkle (MIT, the auto-update framework). Pinned
at version 2.9.2 in
Package.resolved;
license text ships inside the framework at
Sparkle.framework/Versions/Current/Resources/LICENSE. Vetted against
the third-party checklist (MIT, 20-year track record, EdDSA-signed
updates, single-purpose).
To relink the app against a modified FFmpeg: edit the FFmpeg source
unpacked by Scripts/build-ffmpeg.sh, re-run that script, then re-run
xcodebuild -scheme youty -configuration Release. To rebuild the SigLIP
Core ML artifact: edit and re-run
Scripts/convert-siglip-coreml.py.
Pinned versions for both also appear in the app's About panel
(Settings → About).
Other third-party components, their licenses, and where to find their
verbatim license texts:
THIRD_PARTY_LICENSES.md.
License
MIT. Copyright © Bent Eisheuer.
Contributing
See CONTRIBUTING.md for the project layout, build steps,
and ground rules.