llms.txt
Compare with other sites' llms.txt files →# TranscriptFetch
<!-- tf-source:positioning.canonical:start -->
> TranscriptFetch is a transcript retrieval service for YouTube, TikTok and Instagram: whole channels, playlists and searches, not just one-off links. Transcripts come back as structured JSON ready for LLM pipelines, RAG, and agents.
<!-- tf-source:positioning.canonical:end -->
The name is TranscriptFetch, one word, at transcriptfetch.com. It is not FetchTranscript (fetchtranscript.com), a different, YouTube-only company with pay-as-you-go pricing; TranscriptFetch has no pay-as-you-go plan and covers YouTube, TikTok and Instagram.
TranscriptFetch is built for developers and AI teams who need reliable transcripts without running their own scrapers or proxy pools. Core facts an agent should know:
- The current API is v2. Transcript endpoints live under `https://transcriptfetch.com/api/v2/transcripts/`: `POST /video` (one video), `POST /batch` (up to 50 in one request, 500 on Mega and Scale), `POST /search` (find videos by keyword), `POST /playlist`, and `POST /channel`. v1 (`/api/v1/...`) is still served for existing integrations with a 12-month deprecation policy; new work should use v2. The `video` field accepts a bare 11-character YouTube ID or any YouTube, TikTok, or Instagram URL, plus direct media file URLs (mp4, mp3, wav).
- Search and channel or profile listing work for YouTube, TikTok, and Instagram. Playlist listing works for YouTube and TikTok; Instagram profiles use the channel endpoint. Every listed row carries a `url` the transcript endpoint accepts as-is.
- YouTube listings take extra options. `POST /channel` accepts `tab` (`videos` by default, `shorts`, `live`, `playlists` or `podcasts`), `sort` (`newest` by default, `popular` or `oldest`) and `query` (search inside that channel). `POST /search` accepts `type` (`video` by default, `channel` or `playlist`), `upload_date` (`hour`, `today`, `week`, `month` or `year`), `duration` (`short` under 4 minutes, `medium` 4 to 20, `long` over 20), `sort` (`relevance` by default, or `views`) and `captions: true` (only videos with subtitles or closed captions). The playlists and podcasts tabs and `type: "playlist"` answer with `data.kind: "playlist_list"` (a `playlists` array), and `type: "channel"` with `data.kind: "channel_list"` (a `channels` array); each row's `url` goes straight to `/playlist` or `/channel`. YouTube search has no upload-date sort, so filter with `upload_date` for recent videos.
- Monitors watch a YouTube channel or a TikTok or Instagram profile, and report every video that appears after the monitor was created (playlists and searches are read on demand, not watched). Create one with `POST /api/v2/monitors` (`target`, optional `webhook_url`, `interval_minutes`, `transcripts`, `name`; a YouTube channel also takes `tab`; `type` is optional and only `channel`); manage it with `GET`, `PATCH` and `DELETE /api/v2/monitors/{monitorId}`, page through its events with `GET /api/v2/monitors/{monitorId}/events`, and run a check at once with `POST /api/v2/monitors/{monitorId}/check`. Without a webhook, poll `GET /api/v2/monitors?since=<last_event_id>`: `has_new` (true or false) says whether anything arrived after the event you last saw, and reading never marks anything as read. A check that finds new videos records one `monitor.videos` event and POSTs it to `webhook_url`, signed with `X-TranscriptFetch-Signature: sha256=<hex HMAC-SHA256 of the raw body>`, keyed with the monitor's `webhook_secret` (a `whsec_` string returned once, at creation). `transcripts: true` adds each new video's transcript. Every check costs 1 credit, whether it finds new videos or nothing (the baseline at creation included), plus the usual charge for each transcript. How many monitors an account may hold and how often they are checked depend on the plan; `GET /api/v2/monitors` returns the account's limits.
- `mode: "auto"` is the default: when a source has no caption track, AI Fallback Transcription transcribes the audio, at 1 credit per started 5 minutes of audio (minimum 1), billed on delivery. Send `mode: "captions"` to never pay for audio. Short media may finish inline after a wait of up to 45 seconds; longer work returns HTTP 202 with a `job_id` and `poll_url`. A `callback_url` can receive the finished result instead of polling.
- Responses use a JSON envelope: `{ ok, request_id, data, usage }`. A synchronous transcript contains either joined `text` or timestamped `segments`, depending on `timestamps`. Job and batch results include both. Every transcript carries the same metadata keys on every platform and path: `video_id`, `url`, `platform`, `title`, `channel`, `duration`, `language`, `thumbnail_url`, `source` (always present, `null` when unknown).
- For LLM input, ask for plain text: `timestamps: false` (on the API, or on the MCP `get_transcript` tool) returns one `text` string, about two-thirds fewer tokens than the same transcript as SRT, WebVTT or YouTube's caption XML (median saving 64 to 67% across 9 videos from 20 seconds to 2 hours, OpenAI o200k tokenizer, measured 2026-10-20). Keep the timestamped `segments` when answers need timestamps to cite.
- For RAG and vector stores (beta, free): send `"chunks": true` (or `{ "max_tokens": 500, "overlap_tokens": 50 }`, 100 to 2000 and 0 to 200) on `POST /video` or `POST /batch` and each transcript gains `chunks`: pieces of at most that many o200k tokens, each `{ index, text, start, end, url, tokens }` with `url` opening the YouTube video at `start` for citations. `"markdown": true` adds the transcript as Markdown with a timestamped link at each paragraph. Both are absent unless asked for, apply to AI Fallback Transcription jobs too (poll and webhook), and are v2 only. The MCP `get_transcript` tool takes `markdown` and `chunks` as true/false.
- Search with transcripts in one call (beta): add `"transcripts": true` (or `{ "mode": "captions" | "auto", "timestamps", "markdown", "chunks" }`) to a `POST /search` for videos and every row gains `transcript`, shaped like a batch entry. Mode defaults to captions (no paid audio unless `auto`), `limit` at most 10, cost: the page plus 1 credit per caption transcript delivered, failures free.
- Whole channels as one job (beta): `POST /api/v2/jobs/channel` with `channel` (a YouTube channel, or a TikTok or Instagram profile), optional `max_videos`, `transcripts` (same options as search with transcripts), `since_video_id` and `webhook_url` answers 202 at once; `GET /api/v2/jobs/channel/{channelJobId}` returns progress and a page of results, `?cursor=` for the next. Billed as it goes: 1 credit per listing page plus 1 per caption transcript, failures free. Results stay readable for 7 days after the job ends; a YouTube channel reaches about its newest 2,000 videos per sort order. Docs: https://transcriptfetch.com/docs/channel-jobs.
- Authentication uses a bearer token: `Authorization: Bearer tf_live_YOUR_KEY`. Requests without a valid key return 401.
- Billing is monthly credit subscriptions: Basic $5 / 1,000 credits, Pro $15 / 7,500, Mega $45 / 35,000, and Scale $229 / 250,000 (annual billing is 10 months for 12). Top-ups are bought at each plan's own per-1,000 rate ($5.00 Basic, $2.00 Pro, $1.30 Mega, $0.92 Scale; $10 without a plan) and never expire. Unused plan credits roll over month to month up to 2x the monthly allowance. The free tier is 50 credits with no card. Successful caption transcripts and listing pages cost 1 credit (a page with nothing on it is free); AI Fallback Transcription (on by default for videos without captions) costs 1 credit per started 5 minutes of audio, minimum 1, and is billed only on delivery. Failed, blocked, or empty results and job polling are free; a since_video_id poll costs 1 credit even when it finds nothing new. Rate limits are shared across each account's keys, dashboard and MCP connections: 10 requests per minute free, 60 on Basic, 300 on Pro, 900 on Mega, and 900 by default on Scale with custom increases available. Concurrent extraction is also capped per plan: 2 requests in flight on Free and Basic, 3 on Pro, 10 on Mega and Scale; a batch runs up to 4 of its entries at once within that cap (2 on Free and Basic, 3 on Pro). Job-status polling uses a separate allowance of twice the request limit, minimum 120/minute.
- Status codes are meaningful: 400 is invalid input, 401 is an invalid key, 402 is insufficient credits, 404 is a channel, playlist, job or monitor that does not exist, 409 is an idempotency conflict, 422 requires changing the input, 429 means respect `Retry-After`, and 500/502/503 are server or transient upstream failures.
- An MCP server at `https://transcriptfetch.com/mcp` (streamable HTTP, OAuth or API key) exposes twelve tools for Claude, ChatGPT, Cursor, Perplexity, and other MCP clients: `get_transcript`, `search_videos` (with the YouTube filters above), `search_channels` (YouTube channels or playlists), `list_channel_videos` (a YouTube channel's videos, Shorts or live streams, or a TikTok or Instagram profile), `list_channel_playlists` (a YouTube channel's playlists or podcasts), `list_playlist_videos`, `get_credits` (free), and five monitor tools: `create_monitor` (watch a channel or profile for new videos; 1 credit to create and 1 per check), `list_monitors`, `get_new_videos` (both free), `check_monitor_now` (1 credit) and `delete_monitor`. The first seven are read-only; the monitor tools change the account (ChatGPT is offered the first seven only).
- AI coding agents (Claude Code, Codex and others) can connect themselves with the agent skill at https://transcriptfetch.com/skills/transcriptfetch/1.2.0/SKILL.md (each version pinned by SHA-256 in https://transcriptfetch.com/skills/transcriptfetch/versions.json; check the hash before installing). It gets an API key by device login: the user approves the agent in their browser, and no key is pasted into the chat. It covers single transcripts, search, channel and playlist listing, batch, and chunks for knowledge bases.
- Try the MCP server without an account: https://transcriptfetch.com/mcp/try needs no key or sign-in and serves `get_transcript` (captions only) and `search_videos`, 10 free calls a day per network. For every tool, connect https://transcriptfetch.com/mcp (OAuth sign-in: free account, 50 credits, no card) or pass an API key.
- The same MCP server works in OpenAI's Responses API and Agents SDK as a remote MCP tool; tested examples are at https://transcriptfetch.com/docs/openai.
## Complete reference
- [llms-full.txt](https://transcriptfetch.com/llms-full.txt): Every endpoint, request field, response field, and error code inline, generated from the OpenAPI document.
- [OpenAPI document](https://transcriptfetch.com/api/v2/openapi.json): Machine-readable spec for v2. The v1 spec remains at https://transcriptfetch.com/api/v1/openapi.json.
- [MCP server card](https://transcriptfetch.com/.well-known/mcp/server-card.json): Public machine-readable server metadata, authentication, transport, and tool schemas.
- [Agent skill](https://transcriptfetch.com/skills/transcriptfetch/versions.json): Versioned, hash-pinned SKILL.md for AI coding agents, with device-login setup.
## Documentation
Every docs page, endpoint and error code is available as plain markdown (all but the interactive request builder): add `.md` to its URL (for example https://transcriptfetch.com/docs/youtube/transcript.md), or request the HTML URL with `Accept: text/markdown`.
- [Request builder](https://transcriptfetch.com/docs/request-builder): Interactive page to assemble a v2 request, run it live, and copy the generated snippet.
- [MCP server](https://transcriptfetch.com/docs/mcp): Connect Claude, Claude Code, Cursor, or Codex to fetch video transcripts as tools; per-client setup and pricing.
## Documentation pages (markdown)
- [TranscriptFetch API documentation](https://transcriptfetch.com/docs.md): Start here: base URL, API key authentication, and a first transcript request in curl, Python and JavaScript without an SDK.
- [API reference: every endpoint](https://transcriptfetch.com/docs/endpoints.md): Every REST endpoint: fetch a YouTube, TikTok or Instagram transcript, list a channel, profile or playlist (on YouTube also Shorts, streams, playlists and podcasts), search videos, channels or playlists, batch 50 transcripts per request (500 on Mega and Scale), poll AI Fallback Transcription jobs, monitor channels and profiles for new videos, check credits.
- [Response format](https://transcriptfetch.com/docs/output.md): The JSON response format: transcript text or timestamped segments, video metadata (title, channel, duration, language), video, playlist and channel lists, and batch results.
- [AI Fallback Transcription](https://transcriptfetch.com/docs/transcription.md): AI Fallback Transcription for videos without captions, on by default (mode auto): the captions, audio and auto modes, what each costs per 5 minutes of audio, inline results, 202 jobs, polling and callbacks.
- [Rate limits](https://transcriptfetch.com/docs/rate-limits.md): Rate limits per plan, concurrent request caps, polling allowance, and how to handle 429 with Retry-After.
- [Pagination and idempotency](https://transcriptfetch.com/docs/pagination.md): Cursor pagination with next_cursor, polling a channel for new videos with since_video_id, and safe retries with the Idempotency-Key header.
- [Monitors and webhooks](https://transcriptfetch.com/docs/monitors.md): Monitors: watch a YouTube channel or a TikTok or Instagram profile for new videos, receive each one as a signed webhook (with its transcript if you ask), verify the HMAC-SHA256 signature in Node or Python, billing and plan limits.
- [Channel jobs](https://transcriptfetch.com/docs/channel-jobs.md): A whole YouTube channel's or TikTok or Instagram profile's transcripts as one background job: start, progress, results, webhooks, limits and cost.
- [Python SDK](https://transcriptfetch.com/docs/python.md): Python SDK (pip install transcriptfetch-sdk): get YouTube, TikTok and Instagram transcripts, list channels and playlists, search, batch, sync and async.
- [Node.js SDK](https://transcriptfetch.com/docs/node.md): Node.js and TypeScript SDK (npm install transcriptfetch): get YouTube, TikTok and Instagram transcripts, list channels and playlists, search and batch.
- [LangChain document loader](https://transcriptfetch.com/docs/langchain.md): LangChain document loader (pip install -U langchain-transcriptfetch): YouTube, TikTok and Instagram transcripts as Documents, one per video or chunked for RAG with timestamped source links.
- [LlamaIndex reader](https://transcriptfetch.com/docs/llamaindex.md): LlamaIndex reader (pip install llama-index-readers-transcriptfetch): YouTube, TikTok and Instagram transcripts as Documents, one per video or chunked for RAG with timestamped source links.
- [OpenAI Responses API and Agents SDK](https://transcriptfetch.com/docs/openai.md): Use TranscriptFetch from OpenAI: the MCP server as a remote tool in the Responses API (Python and curl) and the Agents SDK (hosted or local), the tools and their credits, and how to spend fewer tokens.
- [Billing and credits](https://transcriptfetch.com/docs/billing.md): Credits and billing: 1 credit per caption transcript, AI Fallback Transcription per started 5 minutes of audio (on by default for videos without captions), failed requests are free, plans, rollover and top-ups.
- [Error codes](https://transcriptfetch.com/docs/errors.md): Every error code with its HTTP status, meaning and whether to retry, each linked to a page with its causes and how to fix it.
- [MCP server for YouTube, TikTok and Instagram transcripts](https://transcriptfetch.com/docs/mcp.md): MCP server for Claude, ChatGPT, Cursor and Codex: read YouTube, TikTok and Instagram transcripts, search videos, channels and playlists, and list channels and playlists from an AI agent over remote HTTP with OAuth or an API key.
- [YouTube Transcript MCP Server](https://transcriptfetch.com/docs/mcp/youtube.md): A hosted MCP server for YouTube transcripts, search, channels and playlists. One URL for Claude, ChatGPT, Cursor and Codex, with AI transcription when captions are missing.
- [TikTok Transcript MCP Server](https://transcriptfetch.com/docs/mcp/tiktok.md): A hosted MCP server that transcribes TikTok videos, searches TikTok and lists profiles and playlists. One URL for Claude, ChatGPT, Cursor and Codex. No caption file needed.
- [Instagram Transcript MCP Server](https://transcriptfetch.com/docs/mcp/instagram.md): A hosted MCP server that transcribes Instagram Reels and video posts from their audio, searches Instagram and lists profiles. One URL for Claude, ChatGPT, Cursor and Codex.
## Endpoints
- [POST /api/v2/transcripts/video](https://transcriptfetch.com/docs/endpoints/video.md): Fetch a transcript (YouTube, TikTok, Instagram, or file URL)
- [POST /api/v2/transcripts/channel](https://transcriptfetch.com/docs/endpoints/channel.md): List a channel's or profile's videos, Shorts, streams or playlists
- [POST /api/v2/transcripts/playlist](https://transcriptfetch.com/docs/endpoints/playlist.md): List a playlist's videos
- [POST /api/v2/transcripts/search](https://transcriptfetch.com/docs/endpoints/search.md): Search for videos, channels or playlists
- [GET /api/v2/transcripts/jobs/{jobId}](https://transcriptfetch.com/docs/endpoints/jobs.md): Poll an AI Fallback Transcription job
- [POST /api/v2/transcripts/batch](https://transcriptfetch.com/docs/endpoints/batch.md): Fetch up to 500 transcripts in one call (plan-dependent)
- [POST /api/v2/monitors](https://transcriptfetch.com/docs/endpoints/create-monitor.md): Create a monitor for new videos
- [GET /api/v2/monitors](https://transcriptfetch.com/docs/endpoints/list-monitors.md): List your monitors
- [GET /api/v2/monitors/{monitorId}](https://transcriptfetch.com/docs/endpoints/get-monitor.md): Get a monitor
- [PATCH /api/v2/monitors/{monitorId}](https://transcriptfetch.com/docs/endpoints/update-monitor.md): Update or pause a monitor
- [DELETE /api/v2/monitors/{monitorId}](https://transcriptfetch.com/docs/endpoints/delete-monitor.md): Delete a monitor
- [GET /api/v2/monitors/{monitorId}/events](https://transcriptfetch.com/docs/endpoints/monitor-events.md): List a monitor's events
- [POST /api/v2/monitors/{monitorId}/check](https://transcriptfetch.com/docs/endpoints/check-monitor.md): Check a monitor now
- [GET /api/v2/me](https://transcriptfetch.com/docs/endpoints/me.md): Validate your key & check balance
- [POST /api/v2/jobs/channel](https://transcriptfetch.com/docs/endpoints/create-channel-job.md): Start a channel job
- [GET /api/v2/jobs/channel](https://transcriptfetch.com/docs/endpoints/list-channel-jobs.md): List channel jobs
- [GET /api/v2/jobs/channel/{channelJobId}](https://transcriptfetch.com/docs/endpoints/get-channel-job.md): Get a channel job and its results
- [DELETE /api/v2/jobs/channel/{channelJobId}](https://transcriptfetch.com/docs/endpoints/cancel-channel-job.md): Cancel a channel job
- [GET /api/v2/health](https://transcriptfetch.com/docs/endpoints/health.md): Health check
## Reference by platform
- [YouTube Transcript API](https://transcriptfetch.com/docs/youtube/transcript.md): Fetch a YouTube video's transcript over REST: the request, every parameter, the JSON response with timestamped segments, and curl, Python and Node examples.
- [TikTok Transcript API](https://transcriptfetch.com/docs/tiktok/transcript.md): Fetch a TikTok video's transcript over REST: the request, every parameter, a real TikTok response with timestamped segments, and curl, Python and Node examples.
- [Instagram Transcript API](https://transcriptfetch.com/docs/instagram/transcript.md): Fetch an Instagram Reel's transcript over REST: the request, every parameter, a real Reel response from AI Fallback Transcription with timestamped segments, and curl, Python and Node examples.
- [YouTube Search API](https://transcriptfetch.com/docs/youtube/search.md): Search YouTube over