vid2context

The full transcript of one video or local media file, with a header (title, channel, URL, duration, how the transcript was made, chapters). With timestamps (the default) every line starts with [m:ss] so you can cite moments. Uses creator captions, then platform auto-captions, then speech-to-text. Cost: speech is about 12k tokens per hour of video. For videos longer than about 20 minutes, call get_outline first and read only the sections you need with get_section, or set maxTokens to receive the transcript in pages (each page ends with a cursor for the next one). Formats: text (Plain transcript text. With timestamps, one "[m:ss] line" per caption.); json (The full VideoContext as JSON: metadata, chapters, timed segments, and how it was produced.); markdown (Markdown transcript: title and metadata, chapters as headings, paragraphs prefixed with timestamp links. Options: linkTimestamps, includeComments, includeVisual.); srt (SubRip (.srt) captions: at most 2 lines of 42 characters per cue, no overlaps, for players and editors.); vtt (WebVTT (.vtt) captions for the HTML <track> element, with speaker voice tags when known.); pack (Context pack: one Markdown file for one or many videos, written for an AI model to read, with citation instructions. Options: maxTokens, includeComments, includeVisual, linkTimestamps.)

Title, channel, description, duration, publish date, chapters, tags and language for a video URL or local file, without fetching the transcript. Cheap: a few hundred tokens, mostly the description.

The output formats get_transcript and get_section accept (text, json, markdown, srt, vtt, pack, ...), with what each is for. combinesVideos marks formats that can merge several videos into one document (export_context_pack). Cheap: under 500 tokens.

无需输入参数。

Start here for any video longer than about 20 minutes. Returns the metadata and the video's chapters (or sections of about five minutes when it has none), each with its start time, a timestamp link and its estimated token count, plus the total. Then read only what you need with get_section (by chapter index or title, or by time). Fetches the transcript once (saved for later calls); the outline itself costs about 300 to 1,500 tokens.

One chapter or time range of a video's transcript, with the same header as get_transcript. Pass chapter (an index or title from get_outline) or from/to (seconds, m:ss or h:mm:ss); from alone reads to the end, to alone reads from the start. Cost: the section's token count from get_outline (about 1k tokens per 5 minutes of speech).

Top comments of a video, with creator replies and pinned comments flagged. Useful for corrections, follow-up questions and audience consensus that the video itself lacks. Cost: about 50 tokens per comment; limit defaults to 50. Only sources whose provider can read comments support this.

Expand a playlist, channel, podcast feed or page into its videos (a single video URL returns itself). Use it to see what a list contains before reading, or pass the list straight to add_to_collection. Cost: about 40 tokens per item; limit defaults to 50.

One Markdown document for many videos (the pack format): a header per video with title, channel, URL and chapters, then each transcript in timestamped paragraphs. Set maxTokens to fit a budget: the pack trims transcripts to fit and says what it left out. Cost: about 12k tokens per hour of video unless maxTokens caps it. Prefer get_outline, get_section and search_transcript when you only need parts.

The models (rungs) this server can use for speech-to-text, LLM, embedding, video and visual work, in free-first order: tier (local, free, keyless, paid), quality 1 to 5, capabilities (json, tools, vision, stream, ...), context window, whether it is enabled, and the names of the env vars that would enable it. Never shows key values. Cost: about 80 tokens per model; filter with kind.

Find where something is said across saved video transcripts, instead of reading whole transcripts (about 12k tokens per hour). Hybrid search: keyword matching for exact names and model numbers, plus meaning-based matching for paraphrases (when an embedding model is available). Returns the best time ranges with title, [m:ss], highlighted text and a timestamp link. Syntax: "exact phrase" must appear, -word excludes. Scope with collection (a collection id) or urls (videos are fetched and saved first if needed).

Build a cited brief across videos on the server with a free LLM (map-reduce for large collections). Every claim cites a video, timestamp and quote that were checked against the transcript; points are marked consensus (two or more videos agree), disputed (both sides cited) or single-source, and claims that failed the check are listed separately, uncited. If you can read the transcripts yourself, the `brief` prompt is cheaper and usually better; use this for collections too large for your context.

Extract structured JSON from videos with a template (buyer-guide, tutorial-steps, comparison, summary, brief) or your own JSON Schema. The answer is validated against the schema (one repair attempt), and every citation ({videoId, startSec, quote}) is checked against the transcript: verified ones gain a title and timestamp url, failed ones are removed and listed in "unverified".

List research collections with how many videos and sources each has. A collection groups videos on one topic and can follow playlists, channels and podcast feeds.

无需输入参数。

Create a research collection. sources are playlists, channels or podcast feeds the collection follows: add_to_collection or a refresh ingests their videos, and watch lists pick up new uploads. Returns the collection with its id.

Ingest videos into a collection now: each URL is a video or a list (playlist, channel, feed), and lists are expanded. Transcripts come from captions first, speech-to-text only when needed. Videos already in the collection are skipped and per-video failures are reported without stopping. Returns the updated collection and a summary (added, skipped, failed with hints).

A collection with a summary of each video (id, title, channel, duration, URL, transcript origin), the sources it follows, inputs waiting for the next refresh and its watch list. Read transcripts with get_transcript, get_outline or search_transcript.

Answer a question from the transcripts of one or more videos, or a saved collection, with server-side retrieval and the free-first LLM chain. Every factual sentence cites a numbered source; each source is a timestamp link to the moment that supports it. Says so plainly when the videos do not cover the question. Use it when you want a cited answer without reading whole transcripts.

Search 1 to 10 videos (or a playlist, capped at 10) for a question and get only the relevant excerpts: ranked by BM25, each with the surrounding context, [m:ss] and a timestamp link, under a token cap. Use it instead of reading whole transcripts (about 12k tokens per hour of speech). Put exact phrases in double quotes to boost them. Lexical: if nothing matches, it lists words the videos do use so you can retry with a synonym.

List the caption tracks a video has (language, name, creator or auto-generated) and, for YouTube, the languages YouTube can machine-translate them into. Call it before get_transcript to pick a language.

The moments of a video worth jumping to, picked by the free-first LLM chain from the timestamped transcript. Every timestamp is checked against the real transcript (snapped within 5 seconds, dropped otherwise); the result says how many were dropped and which rung answered.

Community-marked SponsorBlock segments of a YouTube video: sponsor reads, self-promotion, subscribe reminders, intros, outros, previews and filler, each with its [m:ss] range, a timestamp URL and how much of the transcript it covers (lines, word share and the first words). Use it to leave ad copy out of summaries and briefs: never cite text inside these segments as a claim about the topic. Default categories: sponsor, selfpromo, interaction, intro, outro, preview, filler. All: sponsor, selfpromo, interaction, intro, outro, preview, hook, filler, music_offtopic, poi_highlight, exclusive_access, chapter. poi_highlight comes back as `highlight` (a point, not a skip); a whole-video label comes back as `fullVideo`. Other platforms return no segments. Results carry the SponsorBlock attribution (CC BY-NC-SA 4.0).

Write one Markdown note per video into a folder of an Obsidian vault on this machine: frontmatter properties, chapters as headings, paragraphs prefixed with [m:ss] timestamp links, plus an index note with [[wikilinks]] when a collection is exported. Notes vid2context wrote before are updated; other existing notes are never replaced unless overwrite is true.

Create one Notion page per video under a parent page (chapters as headings, paragraphs with timestamp links, a bookmark to the video). Re-exporting updates the same pages instead of duplicating them. Needs NOTION_TOKEN; the parent page comes from parentPageId or NOTION_PARENT_PAGE_ID.

Compare the transcripts of several videos and group reuploads (the same video uploaded again) and clips (an excerpt of a longer video, with the [m:ss] where it starts in the original and the range it covers). Use it before writing a brief across many videos: count each group once (representativeOf maps every video to its original), or a clip agreeing with its own source looks like consensus. Related pairs share passages without being copies. Works on transcripts only; re-dubbed or translated reuploads are not detected.

Transcribe a live stream (a live YouTube video or an HLS .m3u8 URL) while it runs. New lines arrive about every 30 seconds as progress notifications; the result is the full transcript with [stream time | UTC] stamps when the stream ends or maxMinutes pass. Default 10 minutes, at most 60. For recordings use get_transcript.

Show a video transcript in an interactive viewer (chapters, search, clickable [m:ss] timestamps, one-click citations, send selected lines to the chat) in clients that support MCP Apps, such as ChatGPT and Claude. The text result starts with a header (title, channel, URL, chapters); clients without MCP Apps get the full timestamped transcript as text.

Build one cited digest of several videos (a playlist, a collection, or a list of URLs) using MCP sampling: the server asks the client's model to take notes on each part of each transcript, then combine them. Every claim cites [Title @ m:ss](link), and every timestamp is checked against the real transcript. No server key is needed. If this client cannot sample, the result is a chunk plan with the exact instructions, so you can do the same work yourself.

List a video's caption tracks (languages, creator or auto-generated) and ask the user to pick one, then return that transcript with [m:ss] timestamps. Without a picker in this client, or when the user declines, returns the track list so you can call get_transcript with a language.