Transcripts, search, channels and playlists, plus summaries, chat, comments and images. Over REST or MCP, on one key, spending the same credits as the app.
Let your AI agents fetch the transcript and work with it. One MCP server, the same key.
Ten endpoints, priced by what each one costs to run. The AI ones are not an add-on tier, they are on the same key and the same balance. Each one is also an MCP tool, so an agent can reach it without you writing the call.
The official caption track for any video, watch, Shorts, mobile or embed URL, cleaned, with or without timestamps, in 125+ languages.
Point any MCP client at the server, paste the same key once, and the model decides when to pull a transcript, sweep a channel or read an hour of video. No wrapper code, no second bill.
Add it in the agent.
claude mcp add --transport http youtubescribe http://mcp.localhost:3000/v1 --header "Authorization: Bearer yts_live_..."
Nothing to install. Add the block to your client's config, restart it, and the ten tools appear.
The REST key is the MCP key. Same balance, same credit costs, one dashboard.
Each schema carries its credit price, so a model reaches for the free tool before the expensive one.
A video with no captions returns a named error to the agent and draws down nothing.
Summarise TED's three newest uploads and pull the sharpest line from each.
All three circle the same argument, that attention is the scarce input now. The medicine talk puts it most plainly:
Full caption track, timestamps optional
TL;DR plus timestamped key points
Videos, channels or playlists by query
A question answered with citations
One creator's catalogue, by topic
Comments, with optional sentiment
Every upload from a channel, in order
Story frames keyed to the transcript
Whole courses and lecture series
Newest uploads, safe to poll
The server is model-agnostic. Anything that speaks MCP can call it, and the tool schemas carry their own credit cost so a model can pick the cheap route first.
Sign up and copy it from the dashboard. 50 credits are on the account before you have entered a card.
One header, one query string. Every response is JSON with the same envelope, so error handling is written once.
Works from anything that speaks HTTP, your own code, n8n, Make, Zapier, a cron job on a Pi.
no_captions, private_video, region_blocked, live_in_progress. Never a bare 500, and never a bill.
Transient YouTube failures are retried inside the request. You see one response, not a backoff loop.
Transcript text is passed through, not stored. Turn on the library if you want it kept.
Credits bought on Pro or Studio spend either way. A transcript call costs one credit, an AI summary costs twelve, and the free endpoints never draw down the balance.
50 credits, once. Enough to wire it up and see the shape of the data.
Billed monthly. Cancel any time.
Secure checkout by Stripe
Billed monthly. Cancel any time.
Failures do not bill. A video with no caption track, a private video, an upstream timeout, none of them draw down your balance.
Credits bought on Pro or Studio spend either way, in the browser or through your key. There is no separate API subscription and no separate top-up rate.
Past roughly 250K credits a month the rate comes down and the rate limit is negotiable. Tell us the shape of the workload.
Drag to your monthly credit volume. The figures follow the monthly cycle you picked above.
Roughly 5,000 transcripts, or 416 AI summaries, at this volume.
Credits are weighted by what the call costs to run. A transcript is 1 credit, an AI summary is 12, a story image is 20, and endpoints like tracking new uploads are free. Failures never bill, so no captions, a private video or an upstream timeout do not draw down the balance.