Summary

Ask the video a question instead of parsing its transcript.

The answer path is what earns the install: a grounded reply with timestamped quotes, charged only when an answer is actually produced, and cached transcripts so a follow-up question is fast without re-downloading. When there is no transcript at all, the answer is grounded in sampled frames instead — and the response says so, with confidence capped and no quotes to give.

What it is

A hosted video-data API with an MCP server in front of it, plus this open-source client and docs. A model cannot watch a video; this converts one into an answer, text and images first. It covers YouTube, YouTube Shorts, TikTok, Instagram Reels, Pinterest and Reddit.

What you get

`framefetch_extract` — the main call: metadata, insights, transcript, frames, on-screen text, digest, spoken audio digest, structured chapters and entities, comments and comment sentiment
`framefetch_search` for finding a video when you do not have a URL yet, `framefetch_platform_capabilities` for what each platform supports, `framefetch_account` for your key's state
Question answering that returns a short grounded answer with verbatim timestamped quotes and a confidence level, instead of a transcript you have to parse
Frames sampled the way you ask — every Nth, one per second, or a time range — at any width, with OCR over them
Transcript translation into one of 25 languages, and subtitle export as SRT or VTT
Batch extraction of up to 10 URLs in one call, where a single failing URL carries its own error rather than failing the batch

Requirements

A FrameFetch API key, free to create, or an agent wallet that can pay per call with x402 in USDC on Base — no account in that case. Connect remotely by pointing a client at https://framefetch.net/mcp with the key in an Authorization header, or run the bundled stdio bridge with `npx -y framefetch-mcp` and FRAMEFETCH_API_KEY set. Node 18+ for the client package. Two details worth knowing: this repo is at 0.4.0 while npm currently serves 0.3.0, where the extra helpers are reached through extract with a fields list; and on-screen text uses `text_overlay` on the request but comes back as textOverlay in the response. MIT licensed.

Setup effort

One command plus a key — claude mcp add --transport http framefetch https://framefetch.net/mcp --header "Authorization: <YOUR_FRAMEFETCH_KEY>", then supply credentials