Add one URL to Claude and approve in your browser. No API key to copy. Your assistant can then pull the transcript for any video that has captions, and search the channels you have TranscriptHook watching. Nothing to install. It runs on the free plan.
Yes, a few. TranscriptHook is a hosted one at https://transcripthook.com/mcp. Add that URL to Claude Code, claude.ai, Cursor, VS Code or ChatGPT, approve once in your browser, and the assistant can fetch the transcript of any YouTube video that has captions. It can also read the channels you monitor. Nothing runs on your machine.
/mcp to approve. The Claude Code steps are below.TranscriptHook speaks the Model Context Protocol at https://transcripthook.com/mcp over Streamable HTTP, protocol version 2025-03-26. Point any MCP client at it and 11 tools show up in the assistant's tool list.
Seven of them read: fetch a transcript, list what your channels posted recently, search your history by keyword, list your channels, check your plan usage, read a channel's keyword filter, and list the videos that matched your keywords. Four write: add a channel, pause one, delete one, set a keyword filter. The write tools respect the same plan limits as the dashboard, because they run the same code behind it.
The server runs on our side, so setup is a URL and a browser approval rather than a package and a runtime.
Pick your client. Each one connects through your browser: log in to TranscriptHook, check the approval screen, which names the app, the account and the address you will be sent back to, and approve. Revoke any app in Dashboard → Connect & API keys.
claude mcp add --transport http transcripthook https://transcripthook.com/mcp
Then run /mcp in Claude Code and choose Authenticate. Your browser opens; log in and approve. This is the setup for a YouTube transcript in Claude Code. After that, paste any YouTube link and ask for what you need: "Get the transcript for https://youtu.be/… and summarize the key points."
If the form doesn't open pre-filled: Settings → Connectors → Add custom connector, name it TranscriptHook and paste https://transcripthook.com/mcp. Claude Desktop uses the same connectors as your claude.ai account.
In ChatGPT, open Settings → Apps & Connectors, turn on Developer mode under Advanced, then create a connector with this URL and OAuth authentication: https://transcripthook.com/mcp
Install in VS Code VS Code Insiders
VS Code opens, adds the server, and asks you to sign in to TranscriptHook in your browser.
Cursor desktop signs in through your browser. If sign-in fails (older Cursor versions or Cloud Agents), use the API key method under Other clients.
Apps connected through the browser can read your data and add, pause or resume channels. Removing channels is never granted to them. Manage them under Connected apps in Dashboard → Connect & API keys.
Once connected, try "what did my channels post in the last 24 hours?" That is get_recent_videos, and it costs nothing. Transcripts come back as full text, so the assistant can summarise or quote from them directly.
Anything without OAuth support, and Cursor versions where browser sign-in fails, send an API key as a header instead. Create one in Dashboard → Connect & API keys (the same kind of key as the REST API; you can create several named keys, and the full key is shown only once because keys are hashed at rest). Claude Desktop's config file is ~/.config/claude/claude_desktop_config.json. Cursor keeps its config under Settings → MCP.
{
"mcpServers": {
"transcripthook": {
"url": "https://transcripthook.com/mcp",
"headers": {
"Authorization": "Bearer tr_live_xxxxxxxxxxxx"
}
}
}
}
Only get_transcript spends anything. Everything that reads your own monitored history is free on every plan, so an assistant can look around your account as much as it likes without touching your credits.
| Tool | What it does | Cost |
|---|---|---|
get_transcript |
Full transcript and the video description for any YouTube video, by ID or URL. The video does not have to be one you monitor | 1 credit per successful fetch |
get_recent_videos |
Everything your monitored channels published inside a time window (since: "24h", "7d"), transcripts and descriptions attached |
Free |
search_videos |
Keyword search across your subscription history. Matches titles and the transcript text itself. Results include each video's description, but the description is not searched | Free |
list_channels |
Your monitored channels with their latest video. Each row carries the subscription_id the write tools take |
Free |
get_account_status |
Plan, channels and webhooks used against their limits, credits left and when they reset, rate limit, poll interval | Free |
add_channel |
Start monitoring a channel by @handle, URL or UC… ID. Ask for one you already monitor and you get the existing subscription back, not a duplicate. Counts against your plan's channel limit |
Free |
set_channel_active |
Pause or resume a channel. Paused channels are skipped by the poller but keep every stored video and transcript. A paused channel still holds its slot against the channel limit | Free |
get_keyword_filter |
The keyword filter on each channel (or one), and whether your plan actually applies it | Free |
set_keyword_filter |
Set, change or clear a channel's keyword filter. With one, the channel's webhook fires only for videos whose title or full transcript contains a keyword (whole words, case-insensitive). Available on every plan | Free |
get_keyword_matches |
Videos that matched your keywords since a point in time, with the matching quotes and each video's description. Built for a scheduled routine to poll: pass the last response's next_since. Uses each channel's saved filter, or keywords you pass in |
Free |
remove_channel |
Permanently deletes a channel and every video, transcript and delivery record under it. Off unless you enable destructive tools for the key — see below | Free |
remove_channel is off by default. Until the account owner enables destructive tools in Dashboard → Settings (an account setting for API keys, never for browser-approved apps), the tool is stripped out of the tool list, so your assistant is never told it exists. With the flag on it still needs an explicit confirm. The setting changes only when you toggle it; revoking a key does not switch it back on or off.
The reason for a dashboard toggle rather than just a confirmation prompt: transcripts are written by the channel owner, not by you. A video whose captions carry instructions aimed at your assistant is a real thing, and the confirm flag is something the model decides to set. A toggle in your account is the part of that gate no text your assistant reads can flip. Deletion cascades and we keep no per-account backup, so it is worth the extra click.
Search npm or GitHub for a YouTube MCP server and you will find several good ones, such as kimtaeyoon83/mcp-server-youtube-transcript, jkawamoto/mcp-youtube-transcript, and projects from hancengiz and ergut. They install in a minute, they cost nothing, and for grabbing the transcript of a video someone just sent you, they do the job. If that is your whole use case, use one of those. We would rather say so than pretend otherwise.
What they have in common is where they stop.
A local server asks YouTube for captions from wherever it happens to be running, and YouTube blocks addresses it does not like — data-centre ranges especially. Nothing in the package can do anything about that. TranscriptHook is built to keep working where DIY scrapers get blocked.
Every one of them takes a video and returns its transcript. None of them watch a channel, so none can answer "what did these twelve channels publish this week" or "which of my channels mentioned pricing changes". TranscriptHook keeps that history for you, and search_videos runs over the transcript text, not just titles.
A local server means Node or Python on that box, a config entry on that box, and doing it again on your laptop, your desktop and anywhere else. It also rules out Claude on your phone and Claude on the web, which cannot run a local process at all. TranscriptHook is one URL and a browser approval, the same everywhere.
| Capability | Open-source transcript MCP servers | TranscriptHook MCP |
|---|---|---|
| Getting a transcript | ||
| Transcript by video ID or URL | ✓ Core feature | ✓ get_transcript |
| Survives a YouTube IP block | ✗ Your address, your problem | ✓ Built for where DIY scrapers get blocked |
| Price per transcript | ✓ Free, unmetered | 1 credit 10/mo free, 500/mo on Starter |
| Working across channels | ||
| Monitors channels for new uploads | ✗ | ✓ Checked about every 10 minutes, with stored history |
| "What did my channels post this week?" | ✗ | ✓ get_recent_videos, free |
| Full-text search over past transcripts | ✗ | ✓ search_videos, free |
| Assistant can add or pause channels | ✗ | ✓ Same plan limits as the dashboard |
| Running it | ||
| Install per machine | ✗ Node or Python on every box | ✓ Nothing to install |
| Works on Claude mobile and web | ✗ Local process only | ✓ Remote server over HTTP |
| You control the code | ✓ Fork it, patch it, audit it | ✗ Hosted service |
| Destructive actions gated outside the model | n/a Nothing to destroy | ✓ Per-key toggle you set in the dashboard |
Each of these maps to one or two tool calls. The assistant works out which. New to this? See how to get a YouTube transcript into Claude and summarize videos. Prefer plain HTTP? The same transcripts come from the YouTube transcript API.
"Summarise everything my channels posted in the last 24 hours, grouped by topic." One get_recent_videos call, transcripts included, no credits spent.
"Search my channels for anything mentioning interest rates this month and give me the exact wording." search_videos reads the transcript body, so it finds things no title contains.
"Check this week's videos from the channel that reviews AI tools, pull out the GitHub repos in the descriptions and tell me which are worth a look." The description comes back with each video; your assistant reads the links out of it.
"Pull the transcript for this link and tell me if it says anything about the pricing change." get_transcript takes the bare URL and spends one credit.
"Which channels haven't posted in 60 days? Pause those." list_channels then set_channel_active. Paused channels keep their stored transcripts.
"How many transcript credits do I have left, and when do they reset?" get_account_status answers without spending one, which is the point of it being free.
"Start monitoring @Bloomberg, tag it finance." add_channel takes a handle, a URL or a UC… ID, and hands back the existing subscription if you already had it.
https://transcripthook.com/mcp, and there are free open-source ones you install yourself. The hosted one needs no install, works on phone and web clients, and can also watch channels. The open-source ones are fine for fetching one transcript at a time.
claude mcp add --transport http transcripthook https://transcripthook.com/mcp, then /mcp and choose Authenticate. Your browser opens, you log in and approve. After that, paste a YouTube link in Claude Code and ask for the transcript or a summary.
get_transcript takes a video ID or URL and returns the full text. If you want plain HTTP from your own code instead, the YouTube transcript API returns the same transcripts.
https://transcripthook.com/mcp, not a local process. Any client that accepts a remote MCP URL can reach it: Claude Code, Claude Desktop, claude.ai, ChatGPT, VS Code and Cursor. These connect with a one-time browser approval (OAuth). Clients without OAuth use an API key header. Same URL everywhere, no per-machine setup.
add_channel treats @Bloomberg, youtube.com/@Bloomberg and the bare UC… ID as the same channel, and a unique constraint in the database catches anything the matching misses. Two subscriptions to one channel would mean every new video delivered twice, so this one is worth being strict about.
get_recent_videos anyway, so there is little reason to hand that surface to a tool call.
The free plan includes the full MCP server, 1 monitored channel and 10 transcript lookups a month.
No credit card · Free tier forever