GitHub - fugue-labs/brainrot-detection: LLM-powered YouTube brainrot detector. Monitors your TV, warns over Sonos, and turns it off if your kid doesn't stop watching garbage.

A single Go binary that monitors YouTube on a PS5 (or any YouTube TV app), detects brainrot content using an LLM, plays a warning over Sonos, and turns off the TV if the content isn't changed within 10 seconds.

How it works

PS5 YouTube App
      |
      | YouTube Lounge API (real-time event stream)
      v
  brainrot binary
      |
      |-- Fetches video title, channel, description from YouTube
      |-- Classifies content via GPT (free, uses ChatGPT subscription)
      |
      v
  BRAINROT? ──yes──> Play warning over Sonos (volume +30%)
                          |
                          | 10 second countdown
                          |
                     Still brainrot? ──yes──> Turn off TV via LG WebOS API
                          |
                          no ──> Crisis averted

Key behaviors:

• Pausing, seeking, or buffering does not cancel the warning
• Switching to another brainrot video does not cancel the warning
• The only way to cancel is to switch to content the LLM confirms is not brainrot
• If the TV is turned off and back on, the agent automatically reconnects

What gets flagged

The LLM classifies content with a bias toward flagging. Things it catches:

• YouTube Shorts and TikTok compilations
• Roblox clickbait (Blox Fruits, Brookhaven, "BANNED" videos)
• Minecraft clickbait/drama (not genuine tutorials)
• MrBeast-style challenge/stunt content
• Skibidi toilet, sigma/rizz/gyatt meme content
• Satisfying/oddly satisfying compilations
• Reaction videos, Gacha Life, story time animations
• Content in any language matching these patterns

Things it allows:

• Quality kids' shows (Bluey, Numberblocks, Octonauts, Peppa Pig, Sesame Street, Daniel Tiger, etc.)
• Educational content (PBS Kids, Khan Academy Kids, SciShow Kids, Crash Course Kids)
• Calm Minecraft building/survival content
• Drawing, craft, LEGO, and cooking tutorials
• Nature documentaries and animal videos
• Children's music, read-alouds, audiobooks
• Family movies and Pixar/Disney/Ghibli content
• Real music, sports, and genuine gaming tutorials

Requirements

• Go 1.25+ to build
• Sonos soundbar connected to your TV (tested with Arc Ultra)
• LG WebOS TV (for power-off control)
• PS5 (or any device running YouTube with "Link with TV code" support)
• ChatGPT Plus/Pro/Team subscription (for free LLM classification via OAuth)

Setup

1. Build

git clone https://github.com/trevorprater/brainrot-detection
cd brainrot-detection
go build -o brainrot .

2. Configure

Copy the sample config:

cp config.sample.json config.json

Edit config.json with your device IPs:

{
  "sonos_ip": "192.168.0.XXX",
  "lg_tv_ip": "192.168.0.XXX",
  "lg_tv_key": "",
  "warning_delay_seconds": 10
}

Finding your device IPs:

• Sonos: Open the Sonos app → Settings → System → your speaker → IP address
• LG TV: Settings → Network → Wi-Fi → your network → IP address

The lg_tv_key is populated automatically on first run — you'll see a pairing prompt on your TV. Accept it once, and the key is saved.

3. Run

First run does two things:

1. ChatGPT OAuth — opens your browser to log into your ChatGPT account. This is a one-time login that gives the agent free access to GPT for classification. Credentials are saved to ~/.golem/auth.json.

2. PS5 YouTube pairing — on your PS5, open YouTube → Settings → "Link with TV code". Enter the 12-digit code when prompted. Auth tokens are saved to config.json for future runs.

After setup, the agent runs silently and reconnects automatically if the TV is power-cycled.

4. Customize the warning audio

The warning audio is embedded in the binary at build time. To change it:

1. Replace warning.mp3 with your own audio file
2. Rebuild with go build -o brainrot .

The default warning uses an ElevenLabs-generated British broadcaster voice:

"Attention. Brainrot content has been detected. You have ten seconds to change what you are watching, or the TV will be turned off."

Built with gollem

The brainrot classification is powered by gollem, a Go agent framework with compile-time type safety and structured outputs. It made the LLM integration trivial — define a struct, get back typed results:

type BrainrotClassification struct {
    IsBrainrot bool   `json:"is_brainrot"`
    Confidence string `json:"confidence" jsonschema:"enum=high|medium|low"`
    Reason     string `json:"reason"`
}

agent := core.NewAgent[BrainrotClassification](provider,
    core.WithSystemPrompt[BrainrotClassification]("You are a parental content filter..."),
)

result, _ := agent.Run(ctx, "Title: Skibidi Toilet Episode 73...")
if result.Output.IsBrainrot {
    // play warning, turn off TV
}

Gollem also handles the ChatGPT subscription OAuth (no API key or billing needed), token refresh, and structured output schema generation from Go structs. Zero runtime dependencies in core.

Architecture

Single binary, no runtime dependencies. Everything is embedded or handled via HTTP:

Adapting for other setups

Different TV brand? The LG WebOS turnOffTV() function in main.go is ~15 lines. Replace it with your TV's API:

• Samsung: samsungctl
• Sony Bravia: HTTP API with pre-shared key
• Roku TV: POST http://<ip>:8060/keypress/PowerOff

No Sonos? Replace playWarningOnSonos() with any audio playback method. The warning MP3 is served on localhost:8769/warning.mp3.

Not a PS5? The YouTube Lounge API works with any device that supports "Link with TV code": smart TVs, Chromecast, Xbox, Fire TV, Roku.

License

MIT