What it does
Connect once, and your AI assistant can pull YouTube transcripts on its own — as many as the conversation needs. Paste five talks and ask for a comparison; the assistant fetches all five. Paste a channel and ask a question: the assistant lists the channel's videos, picks the relevant ones by title, reads those transcripts, and answers with quotes. No copy-paste, no tab-switching, no account, no API key.
Claude on the web or the Claude desktop app
Click Add custom connector, paste the URL, save. Done — no login step for the connector.
Claude Code
Easiest path: paste the first snippet into any Claude Code chat and Claude configures itself; the tools appear in your next conversation. The terminal command does the same thing for CLI users — and if your terminal says claude is not recognized, just use the ask-Claude prompt instead.
Cursor and other clients
The button installs it in Cursor directly. Any other client that reads an mcpServers config takes the JSON snippet; the server speaks the Streamable HTTP transport, so anything MCP-compatible from 2025 onward connects.
Fair-use limits
Transcripts already in ScriptHaul's permanent cache serve instantly and generously. Videos nobody has fetched before count against a small daily allowance per connection (20 uncached fetches a day) — plenty for research and note-taking, deliberately useless for scraping. Whole channels and playlists don't run over MCP at all: ask for a bulk link and finish the job on scripthaul.com, where jobs of up to 500 videos run free.
Good prompts to try
"Get the transcript of this video and pull out every book recommendation." "Fetch these three talks and compare their positions on X." "Here's a channel — what has she said about interest rates, with quotes?" "Get the SRT for this video so I can translate the subtitles." The assistant chooses the tools and formats; you just paste links.
Related: JSON for AI pipelines, channels into NotebookLM, or downloading a whole channel.