video-transcripts

Let your model watch the video for you. Give it a YouTube link and it reads what's said: to summarise a talk, find the part about one topic, or answer questions instead of guessing from the title.

Project home · Report a problem

Try it

"Summarise this talk in five bullet points: https://www.youtube.com/watch?v=…"

"What does the speaker say about caching?"

"Give me the recipe from this video as a shopping list."

Setup

This plugin uses yt-dlp, a free tool for reading video sites. Install it once:

Windows	macOS	Linux
`winget install yt-dlp`	`brew install yt-dlp`	your package manager, or `pip install yt-dlp`

Then turn the plugin off and on in LM Studio, or restart LM Studio. The tool only appears once yt-dlp is found, so it never shows up as a tool that can only fail.

The tool

Tool	What it does
`video_transcript`	Returns the video's title, channel, length and transcript.

Works with YouTube and most other video sites that yt-dlp supports.
Prefers subtitles written by a person, and falls back to automatic captions, saying so when it does.
Other languages: ask with language (default English), if the video has subtitles in it.
Plain text, without timestamps, which keeps it short; it can say what was said, not exactly when.
Long transcripts come in parts. The model asks for the next part when it needs it.

Settings

Setting	Default
Max Characters	15000	How much of a transcript comes back at once.

Privacy and safety

Only the subtitles are downloaded, never the video, into a temporary folder that's deleted straight away.
Nothing goes anywhere except the video site itself. No accounts, no API keys.
yt-dlp runs directly, never through a shell, and anything that isn't an http or https link is refused.

Using agent-toolkit? Its Web group already includes video_transcript, so you don't need this plugin as well. Enabling both in one chat gives the model the same tool twice.

MIT licensed · Part of the LM Studio Agent Toolkit