Five sourced lines from a video, podcast, article or PDF, and a check that a quote is in its source
Scored across 7 tools
Each tool has a fairly distinct target: check_quote (single claim), check_memory (batch of up to 10), get_yank (by slug), get_lines_for_url (by URL), recent_yanks (listing), search_lines and search_transcripts (two different corpora). The riskiest pair is search_lines vs search_transcripts, which both full-text search the same library but over different representations; the descriptions do clarify this, though an agent could still misselect. Overall mostly distinct with one foreseeable confusion.
Six of seven names follow a clean verb_noun snake_case pattern (check_quote, check_memory, get_yank, get_lines_for_url, search_lines, search_transcripts). recent_yanks is the sole outlier, a noun-phrase listing tool with no verb, though its intent is clear. Minor deviation only.
Seven tools is well within the ideal 3-15 range and each earns its place: verification (single and batch), retrieval (by slug, by URL), discovery (recent, two search modes). Nothing looks redundant or padded.
The surface covers checking, retrieving and searching, but there is no tool to actually create a yank from a URL or source, even though check_quote and search_transcripts repeatedly tell the agent to 'yank it first' for audio or transcripts. That is a notable dead end in the core lifecycle: an agent can audit and read the library but cannot grow it. Listing/cleanup operations are also absent.
check_quote over a batch, for auditing what an agent wrote down earlier. Give it up to 10 remembered claims, each with the source it is attributed to, and it returns the same verdict and receipt for every one: FOUND, CLOSE MATCH, DISTILLED MATCH, NOT FOUND or CAN'T ESTABLISH, with where, the speaker and the class of that claim, the passage, which representation was read and how much of it, and when. Reach for it when a note, a memory store or a shared workspace has accumulated claims nobody re-checked, and you want the ones no source supports. The summary counts verdicts and lists the claims that came back NOT FOUND; it is deliberately not a score, a percentage or a trust rating, because a NOT FOUND is bounded to the representation Yank read and is never a claim that the words were not said. A claim whose source is audio Yank would have to transcribe returns CAN'T ESTABLISH rather than being paid for. No model is consulted anywhere in this: every verdict resolves against the representation itself, and none of them is a judgement on whether a claim is true.
Inputs: claims (required).
Yank Check: a statement and the source it is supposed to come from, in; a verdict and a receipt, out. FOUND (the words are in the representation Yank inspected), CLOSE MATCH (substantially the same wording, with the difference shown), DISTILLED MATCH (a passage shares most of the statement's content words but not its literal words; shared words can say the opposite, so read the passage), NOT FOUND (Yank inspected an adequate representation and could not find support), or CAN'T ESTABLISH (Yank could not obtain or inspect enough evidence; the reason says why). With where (a moment for media, the publisher's page opened on the exact words for an article), speaker with the class of that claim (file-stated, representation-stated, not-established), the passage around it, which representation was read (captions, transcript, article text, filing, document) and how much of it (inspected: words, duration, and sha256, the fingerprint of the exact text read), the checker's version (checker), and when. A NOT FOUND is a bounded claim scoped to the representation read, never a claim that the words were not said; a CLOSE MATCH lists the words that differ (difference.tokens), never a similarity number; every verdict carries one plain sentence (finding). No model is consulted: the verdict resolves against the representation itself. Sources Yank already holds check immediately; an article, a public PDF (located by page) or a captioned video is read for the check; audio that would need transcribing is not spent on here and returns CAN'T ESTABLISH with acquisition_required, so yank it first. None of the verdicts is a judgement on whether the statement is true.
Inputs: source (required), statement (required).
Ask before you extract: everything yank.info already holds from one source URL. Any form of the link works (a youtu.be short link with a timestamp, a watch URL with a playlist, an article with tracking parameters, an Apple Podcasts or Overcast episode link); equivalent URLs return the same answer. Returns real counts (yanks, lines, verbatim lines) and, for up to five yanks, every line with its verbatim flag, timestamp, moment link, context and permanent card URL. Zero yanks is an answer, not an error. Reads the library only; never fetches the source.
Inputs: url (required).
Fetch a yank (five extracted lines from one source) by its slug, the last path segment of a yank.info/y/ URL. Includes every line's verbatim flag, timestamp, moment link, surrounding context, permanent card URL, share image, plus the whole yank as Markdown.
Inputs: slug (required).
The newest yanks in the library, one card each (rank-1 line, author, source type, URL).
Inputs: limit.
Full-text search across every line yank.info has ever extracted. Returns quotable lines with a verbatim flag (true = literal substring of the source transcript; false = labeled paraphrase), the speaker/author, the source, a timestamp and a moment link where the source is a video, and the permanent card URL to cite. Supports quoted phrases and a leading minus to exclude a word.
Inputs: query (required), limit.
Full-text search inside the whole transcripts of every podcast and video yank.info has yanked, not only their five lines (§10.424). Returns moments: the source, where in it (startSeconds, endSeconds), the words around the match (snippet, the match between « and »), a moment link that opens the source at that second, and yankThisPartUrl, which yanks the few minutes around it. The words are a machine transcript (representation: transcript), not a checked line: to quote one, yank that part or use check_quote. Supports quoted phrases and a leading minus.
Inputs: query (required), limit.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Checks that every quote in an AI-written script appears word for word in its source.
Check that a source really says what your claim says, before you publish. Verbatim, multi-path.
Roboto videos, client interviews, articles and social posts with source citations.
Exact source text, not summaries: check citations and quotes in one call; read pages and PDFs.
Add one secure layer between your agents and this server.
