Quick install
Server URL
https://transcriptfetch.com/mcpAdd the server to your client
Claude
Add custom connector
Name TranscriptFetch
URL https://transcriptfetch.com/mcpConnect in Claude opens TranscriptFetch in Claude's connector directory; or add it as a custom connector with the details above. Sign in to TranscriptFetch and approve access when Claude asks. No API key needed.
Connect in ClaudeClaude Code
claude mcp add --transport http transcriptfetch https://transcriptfetch.com/mcp \
--header "Authorization: Bearer tf_live_YOUR_KEY"Run /mcp in Claude Code to check that the server is connected.
Codex
codex mcp add transcriptfetch --url https://transcriptfetch.com/mcpCodex opens your browser to sign in to TranscriptFetch, no key needed. Check with codex mcp list; if the browser did not open, run codex mcp login transcriptfetch.
Claude Desktop
{
"mcpServers": {
"transcriptfetch": {
"command": "npx",
"args": [
"mcp-remote",
"https://transcriptfetch.com/mcp",
"--header",
"Authorization: Bearer tf_live_YOUR_KEY"
]
}
}
}Settings → Developer → Edit Config, then restart Claude Desktop. mcp-remote bridges the hosted server, so every tool works.
Cursor
{
"mcpServers": {
"transcriptfetch": {
"url": "https://transcriptfetch.com/mcp",
"headers": {
"Authorization": "Bearer tf_live_YOUR_KEY"
}
}
}
}Restart Cursor and the tools appear in Agent mode.
VS Code
{
"servers": {
"transcriptfetch": {
"type": "http",
"url": "https://transcriptfetch.com/mcp",
"headers": {
"Authorization": "Bearer tf_live_YOUR_KEY"
}
}
}
}Open Copilot Chat in Agent mode and enable the transcriptfetch tools.
Run it on your machine
Runs on your machine over stdio with your API key. Five of the hosted tools.
Claude Desktop
{
"mcpServers": {
"transcriptfetch": {
"command": "npx",
"args": ["-y", "transcriptfetch-mcp"],
"env": { "TRANSCRIPTFETCH_API_KEY": "tf_live_YOUR_KEY" }
}
}
}Settings → Developer → Edit Config, then restart Claude Desktop. Needs Node 18 or later.
Cursor
{
"mcpServers": {
"transcriptfetch": {
"command": "npx",
"args": ["-y", "transcriptfetch-mcp"],
"env": { "TRANSCRIPTFETCH_API_KEY": "tf_live_YOUR_KEY" }
}
}
}The same block as Claude Desktop, under mcpServers. Needs Node 18 or later.
Terminal
TRANSCRIPTFETCH_API_KEY=tf_live_YOUR_KEY npx -y transcriptfetch-mcp
# or install it globally
npm install -g transcriptfetch-mcpNeeds Node 18 or later.
Docker
git clone https://github.com/TranscriptFetch/youtube-tiktok-instagram-transcript-mcp
cd youtube-tiktok-instagram-transcript-mcp
docker build -t transcriptfetch-mcp .
docker run --rm -i -e TRANSCRIPTFETCH_API_KEY=tf_live_YOUR_KEY transcriptfetch-mcpKeep -i: without stdin the server exits at once.
Nine tools work with Instagram
get_transcript1 creditFetch the full transcript for a video or media file.
Accepts a YouTube video ID or URL, TikTok and Instagram URLs, and direct media file URLs. If no transcript comes back, the result says whether captions definitively do not exist (aiFallback.captionsUnavailable) and whether transcribing the audio would still work (aiFallback.available). When it does, call this tool again with ai_fallback: true.
videostringrequired- Video ID or URL: YouTube (dQw4w9WgXcQ, youtu.be/…), TikTok, Instagram, or a direct media file URL.
ai_fallbackboolean- Skip captions and use AI Fallback Transcription of the audio instead. Use this only after a previous call reported
aiFallback.available. Charged 1 credit per started 5 minutes of audio, minimum 1 credit, on delivery; short media may return inline, while longer work returns a processing result to collect with a later call. timestampsboolean- Include the timestamped segments in the structured result (default true). Set false when only the text is needed: the segments repeat the transcript's words with timings, so leaving them out keeps the response smaller.
markdownboolean- Return the transcript as Markdown (default false): a title line, then paragraphs that each open with their time, linked to that moment on YouTube. Use it when you will quote or cite the video.
chunksboolean- Also return the transcript split into pieces of about 500 tokens for an embedding model or vector store (default false), each with its start and end time and a link to that moment. In the structured result only.
search_videos1 creditSearch for videos matching a query on YouTube (default), TikTok, or Instagram.
Returns titles, ids, durations, and URLs that get_transcript accepts directly. On YouTube, upload_date, duration, sort and captions narrow the results; search_channels finds channels or playlists instead. To read what the videos say rather than only their titles, set transcripts to N: the caption transcripts of the first N rows (at most 5) come back in the same call.
querystringrequired- Search keywords.
platformyoutube | tiktok | instagramdefault youtube- Where to search. Defaults to youtube.
limitinteger- Max results to return (1–50). Defaults to 5.
transcriptsinteger- Also return the caption transcripts of the first N videos (1-5), each cut at 12,000 characters. Captions only: a video without captions (most Instagram reels) comes back with outcome no_transcript, and
get_transcripttranscribes its audio. 1 credit per caption transcript delivered, on top of the listing's; failures are free.
list_channel_videos1 creditList a creator's items, newest first by default.
Accepts a YouTube channel (@handle, UC… id, or URL), a TikTok profile URL (tiktok.com/@user), or an Instagram profile URL. On YouTube, tab picks the videos, shorts or live tab, sort orders it, and query searches inside the channel. Every row's url can be passed straight to get_transcript. To read what the videos say rather than only their titles, set transcripts to N: the caption transcripts of the first N rows (at most 5) come back in the same call.
channelstringrequired- YouTube handle/ID/URL, or a TikTok or Instagram profile URL.
limitinteger- Max results to return (1–50). Defaults to 5.
transcriptsinteger- Also return the caption transcripts of the first N videos (1-5), each cut at 12,000 characters. Captions only: a video without captions (most Instagram reels) comes back with outcome no_transcript, and
get_transcripttranscribes its audio. 1 credit per caption transcript delivered, on top of the listing's; failures are free.
get_creditsFreeCheck the caller's remaining credit balance.
Free; costs no credits.
No parameters.
list_monitors1 creditList the monitors on this account: what each one watches, how often it checks, whether it delivers transcripts, when it last checked and its last error, and whether a webhook is set.
Also the account's monitor limits, and last_event_id: pass it to get_new_videos as since to read only what is found after it. Free.
sincestring- An event id (mev_…): the last_event_id of an earlier answer. Only what was found after it counts as new. Sets each monitor's hasNew.
get_new_videos1 creditRead the new videos the account's monitors have found, newest first: titles, links and publish times, with their transcripts for monitors that deliver them.
Leave monitor_id out for every monitor. To ask again later, pass this answer's lastEventId as since and only videos found after it come back. Monitors find videos when they check, on their own schedule (check_monitor_now checks at once). Found videos are kept for 30 days. Free.
monitor_idstring- The monitor's id (mon_…), from list_monitors or create_monitor. Leave out for every monitor.
sincestring- An event id (mev_…): the last_event_id of an earlier answer. Only what was found after it counts as new.
cursorstring- nextCursor from the previous answer, for older finds.
limitinteger- Checks to read, newest first (1-50, default 10). One check can report several videos.
timestampsboolean- Include each transcript's timestamped segments (default false: the text only, which is smaller).
create_monitor1 creditWatch a YouTube channel or a TikTok or Instagram profile for new videos.
The monitor checks on a schedule, every interval_minutes, and the new videos each check finds are read with get_new_videos (and posted to webhook_url when one is set). A monitor costs credits for as long as it exists: creating it reads the channel once for 1 credit, and every scheduled check costs 1 credit whether or not it finds anything (over a 30-day month, daily checks come to about 30 credits and hourly checks to about 720). With transcripts: true, each new video's caption transcript costs 1 credit more, and AI Fallback Transcription of one without captions 1 credit per started 5 minutes of audio. Limits: 5 monitors checked at most daily on the free plan; 100 monitors checked as often as every 15 minutes on any paid plan. Tell the user what it will cost and get their go-ahead before calling this. Channels and profiles only: a playlist or a search cannot be watched.
channelstringrequired- What to watch: a YouTube channel (@handle, UC… id or URL), or a TikTok or Instagram profile URL.
platformyoutube | tiktok | instagram- Optional: read from channel. When sent, it must match.
interval_minutes15 | 60 | 360 | 1440- Minutes between checks. Defaults to daily on the free plan, hourly on a paid plan. Each check costs 1 credit.
transcriptsboolean- Also fetch each new video's transcript (default false). Costs 1 credit per caption transcript; AI transcription when there are none is billed by length.
namestring- A name for the monitor.
webhook_urlstring- Optional https URL that receives each find as a signed POST. When set, the answer carries the signing secret, shown only this once.
check_monitor_now1 creditRun a monitor's check now instead of waiting for its schedule, and return the new videos it found.
Costs 1 credit, like a scheduled check (free when the listing fails), plus transcripts for a monitor that delivers them. The next scheduled check counts from now. Takes up to about a minute and a half.
monitor_idstringrequired- The monitor's id (mon_…), from list_monitors or create_monitor.
timestampsboolean- Include each transcript's timestamped segments (default false).
delete_monitor1 creditDelete a monitor, and the new videos it has found.
It stops checking and stops costing credits. This cannot be undone: confirm with the user first. Free.
monitor_idstringrequired- The monitor's id (mon_…), from list_monitors or create_monitor.
A caption fetch or a listing call costs one credit, and AI transcription 1 credit per started 5 minutes of audio. Failed and empty calls are free, and get_credits costs nothing. A monitor costs 1 credit to create and 1 per check, whether or not the check finds anything; reading and deleting monitors is free.
What to know about Instagram
- Every transcript comes from the audioInstagram never exposes captions as a file, so the server transcribes the audio of the Reel or video post; the result's source is audio. It holds the call for a short clip, so the transcript usually comes back in the same response, billed at 1 credit per started 5 minutes of audio, which is 1 credit for a typical Reel.
- The post caption is not the transcriptThe text written under a post comes back as title. The transcript is what was said in the video, which is what an assistant working with a Reel almost always needs.
- Profiles and search
list_channel_videostakes an Instagram profile URL and returns up to 50 recent posts per call, andsearch_videoswithplatform: instagramfinds posts by keyword. Theupload_date, duration, sort and captions filters are YouTube only. - Media links expire
thumbnailUrlon Instagram is a signed link that stops working after a while. Store the post url, not the image link, if you need to come back to it. - What does not return textPrivate accounts, Stories that have expired, and Reels whose message was typed into the frames rather than spoken. Empty and failed fetches are never billed.
Other platformsYouTube transcript MCP serverTikTok transcript MCP server