Video Data API

Turn a TikTok video URL into structured data: metadata (author, duration), engagement insights, a Whisper transcript of the spoken audio, parametrically-sampled frames, and the on-screen text burned into them — through one API call or an MCP tool.

Read the docs Pricing

What you get

For TikTok, FrameFetch returns metadata, insights, Whisper transcript, parametric frames, and on-screen text (OCR) per frame. One JSON response, billed per call — every response includes a cost block.

Quickstart

curl -X POST https://framefetch.net/v1/extract \
  -H "Authorization: Bearer <your-key>" \
  -H "Content-Type: application/json" \
  -d '{
    "url": "https://www.tiktok.com/@user/video/1234567890",
    "fields": ["metadata", "transcript"]
  }'

Get a key with POST /v1/keys (free credit). Full reference in the docs. Agents can pay per call with x402 (USDC) — no account.

Same call from Node.js, using the framefetch npm client:

import { FrameFetch } from 'framefetch';

const ff = new FrameFetch({ apiKey: process.env.FRAMEFETCH_API_KEY });

const r = await ff.transcript('https://www.tiktok.com/@user/video/1234567890');
console.log(r.transcript.source);   // 'whisper' — TikTok has no caption track to prefer
console.log(r.transcript.text);
for (const seg of r.transcript.segments ?? []) {
  console.log(`[${seg.start.toFixed(1)}s] ${seg.text}`);
}

npm install framefetch — a zero-dependency client; ff.transcript() is a scoped shortcut for fields: ["transcript"] on /v1/extract. Source on GitHub.

How TikTok actually serves video

Every downloadable format TikTok's CDN exposes is muxed — video and audio interleaved into one h264/aac file. There is no audio-only rendition to ask for, the way YouTube offers a small opus or m4a track for exactly this purpose. That single difference is why TikTok transcription can never be as bandwidth-cheap as YouTube's: getting the audio out requires downloading a video-shaped file first, every time — there's no way to fetch "just the sound."

FrameFetch's format selector accounts for this with a five-step fallback chain, tried in order: the smallest audio-only stream that's still speech-sufficient (≤64kbps), then any audio-only stream, then the best audio-only stream available — those three cover YouTube and Instagram, which do serve real audio-only formats — and only then, for a muxed-only platform like TikTok, the smallest stream that still carries an audio track, falling back to whatever's available as a last resort. On a real TikTok video that chain resolves to its 540p rendition (1.95 MiB) rather than its 1080p (3.79 MiB) — the audio track is the identical aac stream either way, so there's no reason to pull almost twice the data for pixels that get thrown away immediately after. On a video that does offer genuine audio-only formats, the same logic measured a 49kbps m4a track at 2.78 MiB for an 8-minute video, against 7.78 MiB for yt-dlp's own default (137kbps opus) pick — a 2.8x saving TikTok's muxed-only catalog can never reach.

This chain exists because a simpler, audio-only-first version of it once broke TikTok completely: for a stretch on July 19, 2026, every TikTok transcript request matched no format at all and came back with no transcript and no visible error — a paid field silently returning nothing. It was caught the same day by re-running the exact code path against a platform it hadn't been verified against, and fixed with the muxed fallback described above.

No captions, so it's Whisper end to end

Before spending anything on transcription, FrameFetch always checks first whether a video has a usable caption track — captions are free to fetch and exact, against a per-minute charge for Whisper. TikTok essentially never gives it one to find: unlike YouTube, where creator or auto-generated captions are common and get read directly with no audio download at all, TikTok's metadata reports no usable subtitle or automatic-caption track, so captionsAvailable comes back false and every TikTok transcript takes the Whisper path — which, per the above, means a muxed video download first, every time. transcript.source reads "whisper" on TikTok; it isn't an occasional fallback, it's the only path that exists.

Accuracy & limitations

TikTok transcripts are produced by whisper-large-v3-turbo running on Groq — the same model and provider behind every non-caption transcript FrameFetch produces. It's a faster, distilled variant of OpenAI's full Whisper large-v3, trading a small amount of accuracy for significantly quicker inference; it's multilingual, and reliably strong on one clear speaker talking straight into a phone mic, which covers a large share of TikTok content. The trade-offs are the ones any Whisper model has: a loud trending-audio track underneath the speech, several people talking over each other, or a stretch with no real speech at all (music only, or silence) can produce a misheard word or a short hallucinated phrase rather than a clean gap — a real consideration on a platform where an audio overlay is closer to the norm than the exception. Because there's no caption track to cross-check against (see above), transcript quality on TikTok depends entirely on how cleanly the speech separates from whatever's playing underneath it, with none of the ground-truth captions sometimes available on YouTube.

Sample response

For fields: ["metadata", "insights", "transcript"] — field names and shape match the real schema exactly; the values below are illustrative, not a captured call:

{
  "platform": "tiktok",
  "url": "https://www.tiktok.com/@user/video/7123456789012345678",
  "captionsAvailable": false,
  "metadata": {
    "title": "wait for it",
    "uploader": "user",
    "durationSec": 34,
    "uploadDate": "2026-06-11",
    "sourceFps": 30,
    "thumbnail": "https://p16-sign-va.tiktokcdn.com/....jpeg"
  },
  "insights": { "views": 482100, "likes": 61200, "commentCount": 843 },
  "transcript": {
    "text": "okay so watch what happens when ...",
    "source": "whisper",
    "lang": "en",
    "segments": [ { "start": 0, "end": 2.1, "text": "okay so watch what happens when" } ]
  },
  "cost": { "totalMicros": 2000 },
  "warnings": []
}

A 34-second clip like this one lands on the $0.002 per-call floor rather than the per-minute transcript rate — most TikTok videos are short enough that a metadata+transcript call hits the floor, not a duration-scaled price. Full breakdown on pricing.

Use it from an AI agent (MCP)

FrameFetch ships an MCP server at POST https://framefetch.net/mcp with the tools framefetch_extract, framefetch_search, framefetch_account and framefetch_platform_capabilities — point your agent at a TikTok URL directly. See the MCP setup guide for a working Claude Desktop / Cursor config.