One-command tool to download a YouTube video, generate a transcript, and produce an LLM-powered summary — all organized into a clean per-video directory.
# Simplest usage — prompts for URL, auto-detects everything
./get-yt
# Pass URL directly
./get-yt "https://www.youtube.com/watch?v=VckmK-ZCpAU"Required:
- yt-dlp —
brew install yt-dlp - ffmpeg —
brew install ffmpeg - python3 — included with macOS / Homebrew
At least one LLM backend (auto-detected in this order):
- LM Studio — lmstudio.ai (local, Apple Silicon optimized, no API key)
- ollama —
brew install ollama(local, no API key needed) - llm CLI —
pipx install llm(Simon Willison's multi-provider CLI) - Anthropic API — set
ANTHROPIC_API_KEYenv var - OpenAI API — set
OPENAI_API_KEYenv var
Optional (for --method whisper):
- mlx_whisper —
pipx install mlx-whisper(Apple Silicon optimized)
| Flag | Description | Default |
|---|---|---|
-u, --url URL |
YouTube video URL. Prompted interactively if omitted. | — |
-m, --method METHOD |
Transcript source: auto, youtube, whisper |
auto |
-w, --whisper-model MODEL |
Whisper model identifier | mlx-community/whisper-large-v3-turbo |
-o, --output-dir DIR |
Base directory for all output | . (current dir) |
-p, --llm-provider PROVIDER |
LLM backend: lmstudio, ollama, llm, anthropic, openai |
auto-detected |
-l, --llm-model MODEL |
Model name for the chosen provider | provider-specific |
-a, --audio-only |
Download audio only (m4a) — no video | off |
-t, --transcript-only |
Get transcript only — no media download | off |
-d, --diarize |
Add speaker labels via LLM post-processing | off |
-n, --no-summary |
Skip the LLM summarization step | off |
-k, --keep-scratch |
Preserve the .scratch/ working directory |
off |
-h, --help |
Print usage and exit | — |
-v, --version |
Print version and exit | — |
Tries to download YouTube subtitles first (manual preferred, then auto-generated). If none are available, falls back to local Whisper transcription.
Uses only YouTube-provided subtitles. Fails with a clear error if the video has no English subtitles. This is the fastest option — no audio extraction or model download needed.
Extracts audio from the downloaded video and transcribes locally using mlx_whisper. Produces higher-quality transcripts than YouTube auto-subs but takes longer (roughly 1 minute per 10 minutes of video on Apple Silicon). Requires mlx_whisper to be installed.
The script makes exactly one LLM request per video — sending the full transcript with a structured summarization prompt. With --diarize, a second request is made for speaker attribution.
Connects to LM Studio's local server at localhost:1234 via its OpenAI-compatible API. Auto-selects the currently loaded model. No API key required.
LM Studio runs Apple Silicon-optimized models (MLX, GGUF) and is the recommended local provider for macOS. Install from lmstudio.ai, load a model, and enable the local server.
# Auto-detected when LM Studio server is running
./get-yt "$URL"
# Explicitly select LM Studio
./get-yt -p lmstudio "$URL"
# Use a specific loaded model
./get-yt -p lmstudio -l "qwen/qwen3-vl-8b" "$URL"Connects to the local Ollama server at localhost:11434. Auto-selects the first installed model if --llm-model is not set.
# Start ollama if not running
ollama serve &
# Use a specific model
./get-yt -p ollama -l gemma3:4b "$URL"Uses Simon Willison's llm CLI, which supports dozens of backends via plugins.
pipx install llm
llm keys set openai # or install a plugin: llm install llm-claude-3
./get-yt -p llm -l gpt-4o "$URL"Direct API call via python3. Requires ANTHROPIC_API_KEY in the environment.
export ANTHROPIC_API_KEY="sk-ant-..."
./get-yt -p anthropic "$URL" # default: claude-sonnet-4-20250514
./get-yt -p anthropic -l claude-opus-4-20250514 "$URL"Direct API call via python3. Requires OPENAI_API_KEY in the environment.
export OPENAI_API_KEY="sk-..."
./get-yt -p openai "$URL" # default: gpt-4o
./get-yt -p openai -l gpt-4o-mini "$URL"output-dir/
├── .scratch/ # temporary working space (auto-cleaned)
│ └── VIDEO_ID/
│ ├── video.mp4 # downloaded video (moved to final dir)
│ ├── video.en.vtt # raw YouTube subtitles
│ ├── audio.m4a # extracted audio (whisper method)
│ └── llm_prompt.txt # the prompt sent to the LLM
│
└── Video_Title_VIDEO_ID/ # one directory per video
├── video.mp4 # the video file (h264/mp4)
├── transcript.txt # plain-text transcript
└── summary.md # LLM-generated markdown summary
The .scratch/ directory is automatically deleted after a successful run unless --keep-scratch is used.
The script checks for existing output files at each step and skips work that's already done. If a run fails partway through (e.g., network error during LLM call), just re-run the same command — it picks up where it left off.
Adds speaker labels to the transcript using an LLM post-processing pass. The raw transcript is kept as transcript.txt and the labeled version is saved as transcript.diarized.txt. The summary (if generated) uses the diarized version for better context.
The LLM infers speaker names from the video title and dialogue context (e.g. "DICK CAVETT:", "BETTE DAVIS:"). Works with any LLM provider. Can be combined with --no-summary to get just the labeled transcript.
Note: Diarization quality depends heavily on model capability and context window size. Small local models may struggle with long transcripts. For best results, use a large-context model (Claude, GPT-4o, or a larger ollama model like llama3:70b).
Downloads only the audio track as audio.m4a instead of the full video. Useful for podcasts, interviews, or saving disk space. The transcript and summary steps work the same — Whisper can transcribe directly from the audio file.
Skips all media downloads and only produces transcript.txt. Uses YouTube subtitles (or subtitle-only fetch). Implies --no-summary. Ideal for quickly grabbing what was said without storing any media.
Skips the LLM summarization step. Produces the media file and transcript but no summary.md. Useful when you don't have an LLM provider configured or just want the raw transcript.
# Basic — auto-detect everything
./get-yt "https://www.youtube.com/watch?v=VckmK-ZCpAU"
# Add speaker labels to transcript
./get-yt --diarize "$URL"
# Audio only — no video file
./get-yt --audio-only "$URL"
# Just the transcript — no media, no summary
./get-yt --transcript-only "$URL"
# Download + transcript, skip the LLM summary
./get-yt --no-summary -o ~/archive "$URL"
# Use Whisper for higher-quality transcript
./get-yt --method whisper "$URL"
# Use Anthropic API with a specific model
./get-yt -p anthropic -l claude-sonnet-4-20250514 "$URL"
# Batch process a list of URLs
while IFS= read -r url; do
./get-yt -o ~/archive "$url"
done < urls.txt- Context window limits: Very long videos (3+ hours) may produce transcripts that exceed the context window of smaller models. Use a model with a large context window (e.g., Gemini, Claude, GPT-4o) for long content.
- YouTube subtitle quality: Auto-generated subtitles have no punctuation or speaker labels. Whisper transcripts are generally more readable.
- Video format: Videos are downloaded with h264 video + m4a audio in an mp4 container for broad compatibility.
- LLM calls: The script sends one LLM request per video for summarization. With
--diarize, a second request is made for speaker attribution.
"No LLM provider found"
Install LM Studio (lmstudio.ai), ollama (brew install ollama && ollama serve), the llm CLI (pipx install llm), or set an API key (export ANTHROPIC_API_KEY=...).
"mlx_whisper not installed"
Run pipx install mlx-whisper. Only needed when using --method whisper.
"No YouTube subtitles available"
The video has no English subs. Use --method auto (to fall back to Whisper) or --method whisper.
LLM timeout or empty output
For local models (ollama), ensure the server is running (ollama serve). For cloud APIs, check your API key and network connection. The script allows up to 10 minutes for the LLM response.
Video download fails
Update yt-dlp: brew upgrade yt-dlp or pipx upgrade yt-dlp. YouTube frequently changes its format negotiation.