my-yt

A clean and minimal youtube frontend supported by yt-dlp, and optionally your local AI model, to make your youtube experience local, mindful and succint.

Features

• Channel management/Subscriptions
• Download videos from YouTube, using yt-dlp behind the scenes
• Ignore videos you don't want to watch
• Remove sponsors automatically thanks to SponsorBlock
• Offline media playback
• Native Google Chromecast support
• Disable clickbait thumbnails
• Play videos in background
• Summarize video content using your local AI model (e.g. Ollama/LMStudio) or hosted provider (e.g. OpenAI)
• Native Picture-in-Picture support
• No dependencies (except for nano-spawn, which itself has no transient deps)
• HTML/CSS only, no JS frameworks on client/server side
• Subtitles using <track> element and WebVTT API
• Automatic transcoding to h264 for max compat (especially for iOS devices)
• Host it in your home network to playback videos on all your devices
• Just JSON files for persistence, stupid simple management and backup

Check out the todo list

Application runs on http://localhost:3000

Installation (node.js)

git clone https://github.com/christian-fei/my-yt.git
cd my-yt
npm i
# install yt-dlp, https://github.com/yt-dlp/yt-dlp
# install ffmpeg, https://ffmpeg.org/

npm start

Installation (docker)

git clone https://github.com/christian-fei/my-yt.git
cd my-yt
docker compose up --build -d

or

docker run -p 3000:3000 -v /path/to/your/data/folder/for/persistence:/app/data christianfei/my-yt:latest
# ok calm down, here is an actual example command to copy-n-paste, jeez
docker run -d -p 3000:3000 -v $HOME/my-yt-data:/app/data christianfei/my-yt:latest

Environment variables for LLM integration

The default values are

AI_MODEL       meta-llama-3.1-8b-instruct
AI_HOST        http://127.0.0.1:1234
AI_ENDPOINT    /v1/chat/completions
AI_APIKEY

Simply set the env variables to your needs, by following the format above (e.g. url starting with "http", no ending slash, AI_ENDPOINT with leading slash and path)

Some examples:

AI_MODEL=gpt-4o-mini AI_HOST=https://api.openai.com AI_APIKEY=sk-proj-123 npm start

# or with docker in background
docker run -e AI_MODEL=gpt-4o-mini -e AI_HOST=https://api.openai.com -e AI_APIKEY=sk-proj-123 -d -p 3000:3000 -v $HOME/my-yt-data:/app/data christianfei/my-yt:latest
# running in foreground
docker run -e AI_MODEL=gpt-4o-mini -e AI_HOST=https://api.openai.com -e AI_APIKEY=sk-proj-123 --rm -it -p 3000:3000 -v $HOME/my-yt-data:/app/data christianfei/my-yt:latest

Environment variable to skip transcoding to h264

If your system is under stress when downloading a video, that could be because of the automatic transcoding that's happening behind the scenes.

This is a feature to make the video compatible with most devices (especially iOS)

If you don't have the need, you can skip the transcoding process by setting SKIP_TRANSCODING=true.

Project

Here are some links to help you understand the project better:

server.js

Bare HTTP server

Handles SSE for client updates

Implements HTTP Ranged requests for video playback

llm.js

Makes requests using the chat completions API of your favorite LLM provider that has a "standard" /v1/chat/completions endpoint (e.g. Ollama/LMStudio/OpenAI and others)

sse.js

Utility functions for Server-sent events

subtitles-summary.js

Summarizes video transcript using LLM provider

youtube.js

yt-dlp wrapper to download videos, get channel videos and video information and transcript

repository.js

Handles persistence of video information (set video as downloaded, summary, ignored, upserting videos, etc.)

client

dependency less, bare HTML5, CSS3 and JS for a basic frontend

Handles SSE updates, interacting with the API

Why?

• wanted to get back my chronological feed, instead of a "algorithmically curated" one
• no distractions
• no clickbait thumbnails (using mq2 instead of mqdefault thumbnail, thanks @drcheap)
• no comments
• no related videos, or any algorithmically determined videos pushed in your face
• no sponsors, thanks to SponsorBlock
• just videos and a clean UI
• wanted to dabble with Server-sent Events
• wanted to try integrate the so much hyped AI in a personal project
• wanted to try out yt-dlp
• wanted to experiment with the HTML5 <track> element and WebVTT API
• feedback was awesome on HackerNews, so here we are
• I am even paying for YouTube Premium, so it's not a matter of money, but a matter of control over my attention and enhanced offline experience

development

run tests

npm t

build docker image

docker build . -t my-yt:latest