The MCP server is live for every plan.Connect it in a minute →
MCP

Your AI can read. Give it something to watch.

Connect Claude, Cursor and any other MCP client to transcripts from YouTube, TikTok, Instagram and direct media files. One server. No code.

Connect your AIRead the docs50 FREE CREDITS / NO CARD

Your assistant stops at the play button

Ask an assistant about an article and it answers. Send it a video and it guesses from the title, the description, whatever the page happened to expose. The words spoken in the clip are the part it never sees.

TranscriptFetch MCP hands it the transcript instead, with timestamps attached.

The shape of it

One server. Every platform. Any client.

transcriptfetch/mcp
ClaudeChatGPTCursorAny MCP clientYouTubeTikTokInstagram+ media files

From asking about a video to reading it

Whether you are one person with a Claude tab open or a team shipping agents to production.

For individuals and teamsAdd the server once in your client's settings. Then paste a link into the chat and ask about it the way you would ask about a page. No key handling, no code, no terminal.MCP CLIENTCONNECTEDClaude DesktopLIVEChatGPTLIVECursorADDVS CodeADDSet it up →
For teams building agentsYour agent calls one tool and gets structured segments back. Every call is logged against a key you control, and a failed fetch costs nothing.TOOL CALLget_transcript(video)188 msRESPONSEsegments[] · source: captions4 rowsBILLED1 credit, on success onlyfrom $0.005Read the docs →
In practice

What people ask it for

Three things that were a morning of scrubbing through video, now a sentence in a chat window.

RESEARCHTwelve interviews, one afternoonDrop a playlist into the chat and ask which guests mentioned pricing. The answer comes back with timestamps you can jump to.
CONTENTA newsletter from a two-hour interviewAsk for the three arguments a two-hour YouTube episode actually makes, quoted verbatim rather than paraphrased from the show notes.
ENGINEERINGAn agent that watches the feedYour agent pulls new clips from a creator each morning, reads them, and files anything that matches a phrase you care about.

Everything the server handles for you

Fallback on every fetchCaptions come back immediately. When a video has none, AI transcription starts as a job and the assistant calls the tool again to collect the text.
Timestamps on every segmentEach line carries its start time, so an assistant can cite the moment instead of summarising the whole thing.
One shape, every platformA TikTok link and a YouTube link return the same fields. Nothing downstream has to branch on where the link came from.
Failures cost nothingA credit is spent when text comes back. Failed fetches, private videos and dead links are never billed.
A key per clientSeparate keys per client or per agent, each with its own log, so you can see what asked for what and revoke one alone.
No infrastructure to runNo queue to babysit, no proxy pool, no scraper to repair the week a platform changes its markup.

One server URL over streamable HTTP: https://transcriptfetch.com/mcp

SET UP YOUR CLIENT

Claude MCP server setup →

All three Claude paths on one page: the claude.ai one-click connector (OAuth, no key), Claude Code's one-liner, and the Claude Desktop mcp-remote bridge.

Cursor MCP server setup →

The ~/.cursor/mcp.json entry, and the same block for VS Code, so the tools appear in Agent mode after a restart.

Codex MCP server setup →

Codex configures MCP over stdio, so this is the mcp-remote bridge and the ~/.codex/config.toml it belongs in.

BY PLATFORM

YouTube MCP server →

Ask about any YouTube link in the chat. Captions where the video publishes them, AI transcription where it does not.

TikTok MCP server →

Short-form clips read as text, with timestamps, so an assistant can quote what was said rather than the caption.

Instagram MCP server →

Reels and video posts transcribed on demand, returned in the same shape as every other source.

The MCP setup guideAdding the server in Claude, Cursor and VS Code, with the exact fields each one asks for.Read it →Tool and response referenceEvery field the tool returns, the error codes, and how the fallback decides which source to use.Read it →

Frequently asked questions

A standard way for an AI client to call outside tools. You add our server once and your assistant gains the ability to fetch transcripts, the same way it might gain a calendar or a file system.

No. Adding the server is pasting a URL and a key into your client's settings. Code is only involved if you are building your own agent, in which case the same server works through the API.

AI transcription kicks in automatically - short videos come back in the same call, typically in ~30 seconds. Long media becomes a background job the assistant collects on its next call, or it retries with ai_fallback: true when the response suggests it.

Caption fetches cost one credit; AI transcription is billed by audio duration, only on delivery. Both draw from the same allowance as the API. Failures, private videos and blocked or unavailable links cost nothing.

Yes. Each key is logged separately, so you can give one client its own key, watch what it calls, and revoke it without touching the others.

If it speaks MCP, it works. Nothing in the server is specific to a particular client, so anything that ships support later will connect with the same URL.

Give your assistant the whole recording

50 free credits a month. Failed fetches are never billed.

Get a keySee the setup