# Source Library

> Search 15K rare pre-modern texts translated to English: philosophy, religion, science, literature.

- name: io.github.Embassy-of-the-Free-Mind/sourcelibrary
- version: 4.3.4
- connection class: R0 (autonomous)
- trust: 90/100
- quality: 69/100 (needs work)
- owner: not claimed
- descriptor: https://api.protogrid.dev/v1/servers/io.github.Embassy-of-the-Free-Mind%2Fsourcelibrary
- tools: https://api.protogrid.dev/v1/servers/io.github.Embassy-of-the-Free-Mind%2Fsourcelibrary/tools

**An agent can connect right now, no human step.** Remote endpoint, no authentication, reachable on the last probe.

## Connection (mcpServers)

```json
{
  "mcpServers": {
    "sourcelibrary": {
      "type": "http",
      "url": "https://sourcelibrary.org/api/mcp"
    }
  }
}
```


## Trust

- hygiene: 100
- liveness: 100
- freshness: 65
- provenance: 90
- drivers: +repository +namespace-matches-repo +website +reachable +uptime-100% ~updated-122d-ago
- Derived from observable signals (official registry feed and our own credential-free probes); no code audit performed.

## Quality 69/100 (needs work)

### protocol: 75 (weight 25)

- warn `protocol.modern`: newest supported version is 2025-11-25; 2026-07-28 not supported, older versions are deprecated until 2027-07-28
- n/a `protocol.stateless`: only applies to 2026-07-28 servers
- pass `protocol.transport`: streamable HTTP
- n/a `protocol.list_ttl`: only applies to 2026-07-28 servers

### authorization: n/a (weight 15)

- n/a `auth.prm`, `auth.as_metadata`, `auth.cimd`: no remote uses OAuth
- n/a `auth.secret_in_url`: no templated URL

### tool hygiene: 86 (weight 30)

- pass `tools.descriptions`: every tool has a description
- warn `tools.description_length`: 9 tools have descriptions over 1,024 characters, which crowds the context
- pass `tools.schemas`: every tool has an object input schema
- pass `tools.annotations`: 22 of 22 tools declare readOnlyHint or destructiveHint
- pass `tools.directory_hints`: every tool declares readOnlyHint, destructiveHint, idempotentHint and openWorldHint
- warn `tools.token_cost`: about 13,046 tokens to load every tool
- pass `tools.api_dump`: tools are not a one-to-one API dump

### stability: 25 (weight 15)

- warn `stability.changes`: 26 tool changes in 30 days
- fail `stability.rug_pull`: 1 tools kept their name but changed most of their description in 30 days; review before trusting them

### dependencies: n/a (weight 15)

- n/a `deps.known_vulns`, `deps.mcp_sdk_version`, `deps.resolvable`: no npm or PyPI package

- tools: 22, about 13,046 tokens to load them all
- tool set hash: 76af24382a95a4ef610ce889d4b6c6aa
- tools last changed: 2026-10-20T14:54:24.306Z
- badge: [![protogrid quality](https://protogrid.dev/badge/io.github.Embassy-of-the-Free-Mind/sourcelibrary.svg)](https://protogrid.dev/servers/io.github.Embassy-of-the-Free-Mind/sourcelibrary)
- Checks run on what our credential-free, read-only probes observe; tools are never called and no code audit is performed.

### Tool changes

- 2026-10-20 `search_images` changed (inputSchema)
- 2026-10-20 `search_within_book` changed (description)
  - before: SEARCHES INSIDE ONE BOOK (requires book_id). PRIMARILY KEYWORD: it runs a lexical search over the book's pages plus a narrow scoped-semantic pass (top ~10), interleaved by relevance. PICK THIS when you know the wording you are looking for, or want every page of one book mentioning a term. → IF YOU ARE SEARCHING FROM A PARAPHRASE, a half-remembered line, or a modern restatement, USE search_concept INSTEAD — it is the meaning-matching tool and it searches the whole corpus, including translations whose vocabulary differs completely from yours (Thomas Taylor writes "energies" for energeia and "felicity" for eudaimonia, so a sensible modern paraphrase can miss his pages entirely while matching semantically). → To find the book first, use search_library or search_concept, then pass its book_id here. Each result carries score (0-1, normalised within this book) and found_by ("keyword", "semantic", or "both" — both is the strongest signal). Results flagged is_front_matter are the translator's or publisher's words rather than the author's, and are ordered last. Returns OCR and translation snippets with page numbers, ready to cite.
  - after: SEARCHES INSIDE ONE BOOK (requires book_id). PRIMARILY KEYWORD: it runs a lexical search over the book's pages plus a narrow scoped-semantic pass (top ~10), interleaved by relevance. PICK THIS when you know the wording you are looking for, or want every page of one book mentioning a term. → IF YOU ARE SEARCHING FROM A PARAPHRASE, a half-remembered line, or a modern restatement, USE search_concept INSTEAD — it is the meaning-matching tool and it searches the whole corpus, including translations whose vocabulary differs completely from yours (Thomas Taylor writes "energies" for energeia and "felicity" for eudaimonia, so a sensible modern paraphrase can miss his pages entirely while matching semantically). → To find the book first, use search_library or search_concept, then pass its book_id here. Results are ordered by relevance within this book. Each carries score (0-1, ABSOLUTE, not relative to the other results, so a query the book cannot answer scores low everywhere): the higher of similarity (raw cosine from the meaning pass, calibrated like search_concept: 0.70+ strong, 0.55–0.70 verify, below 0.55 drift) and phrase_match (how much of a multi-word query the page prints in order at one place; 1 is all of it; a one-word query the page prints scores 1). Each also carries found_by ("keyword", "semantic", or "both" — both is the strongest signal). Results flagged is_front_matter are the translator's or publisher's words rather than the author's, and are ordered last. Returns OCR and translation snippets with page numbers, ready to cite.
- 2026-10-13 `propose_correction` changed (inputSchema)
- 2026-10-13 `submit_page_review` changed (inputSchema)
- 2026-10-13 `continue_review_shift` changed (description)
  - before: Next page of the book being reviewed, after the user says "continue" (or "next"). Pass next.book and next.after from the previous page, and volunteer_code. Returns the same shape as start_review_shift; follow how_to_run again. When page is null the book is finished: say so and offer start_review_shift with new_book true.
  - after: Next page of the book being reviewed. Call it straight after finishing a page until the stretch the user chose is done, and again when they ask for more. Pass next.book and next.after from the previous page, and volunteer_code. Returns the same shape as start_review_shift; follow how_to_run again. When page is null the book is finished: say so and offer start_review_shift with new_book true.
- 2026-10-13 `start_review_shift` changed (description)
  - before: Start a volunteer REVIEW SHIFT: the user lends their Claude (you) to Source Library to check one book, page by page, against its scans. Use when the user wants to help Source Library, says "start a shift" or "review pages", or asks how they can contribute. The user does NOT need to read the book's language: YOU do the checking (scan against transcription, transcription against English) and the user runs the shift, watches, asks questions and says "continue". Returns ONE page: its scan (inline image), our transcription, our English, the last lines of the previous page, and next (pass to continue_review_shift). Follow first_time once, then how_to_run on every page. With no book, it resumes the user's last run (pass volunteer_code) or picks a book that needs checking; book takes an id, slug or reader URL if the user names one; new_book true skips the resume. Shifts are for a person at the keyboard: one page per message, never run unattended or in a loop.
  - after: Start a volunteer REVIEW SHIFT: the user lends their Claude (you) to Source Library to check one book, page by page, against its scans. Use when the user wants to help Source Library, says "start a shift" or "review pages", or asks how they can contribute. The user does NOT need to read the book's language: YOU do the checking (scan against transcription, transcription against English) and the user starts the shift, chooses how many pages, watches and asks questions. Returns ONE page: its scan (inline image), our transcription, our English, the last lines of the previous page, and next (pass to continue_review_shift). Follow first_time once, then how_to_run on every page, working through the stretch of pages without stopping between them. With no book, it resumes the user's last run (pass volunteer_code) or picks a book that needs checking; book takes an id, slug or reader URL if the user names one; new_book true skips the resume. Run shifts only while the user is in this conversation; never set one up to run on a schedule or from a script.
- 2026-10-13 `search_concept` changed (inputSchema)
- 2026-10-13 `search_translations` changed (inputSchema)
- 2026-10-12 `search_translations` changed (description, inputSchema)
  - before: RETURNS QUOTABLE PASSAGES (page-level snippets + citation URLs), matched by KEYWORD/term. PICK THIS to find a quote or textual evidence on a topic across the whole library. → If the modern word won't literally appear in historical texts, use search_concept (matches by meaning); to list which BOOKS cover a topic use search_library; to dig inside one known book use search_within_book; if the user named an author/work, get_book first (its AI summary is usually the right first read). Query tips: single distinctive terms ("memory palace", "wax tablet") work best; multi-word natural-English queries ("unity of the intellect") may return fewer results because matching is term-based, not phrase-based. Each snippet has a snippet_type — "translation"/"ocr" means it is a verbatim extract from the source text; "summary" means it is AI-generated description (do not quote those as the author's words). Response includes total_matches, returned, and offset for pagination. Cross-cultural tip: for pre-modern or non-Western topics, search source-tradition vocabulary rather than modern English terms — e.g. for seminal economy search "jing" or "bindu" or "istimnāʾ", not "semen retention"; for female homoeroticism search "tribade" or "sahq", not "lesbian". The corpus is indexed via period translations that use tradition-internal terminology.
  - after: RETURNS QUOTABLE PASSAGES (page-level snippets + citation URLs), matched by KEYWORD/term. PICK THIS to find a quote or textual evidence on a topic across the whole library. → If the modern word won't literally appear in historical texts, use search_concept (matches by meaning); to list which BOOKS cover a topic use search_library; to dig inside one known book use search_within_book; if the user named an author/work, get_book first (its AI summary is usually the right first read). Query tips: a distinctive term ("wax tablet") or the sentence you are looking for both work; matching is by word, and a page that prints your words together and in order ranks above one that scatters them. Each passage carries phrase_match (0-1, absolute, not relative to the other results): 1 means the page prints your whole query in order at one place, 0 means your words occur only apart; when every passage is under about 0.2 the library probably does not print that sentence, so try search_concept. It is absent for a one-word query and for a passage found by meaning alone. found_by says which lane found the page: "keyword", "semantic", or "both" (the strongest signal). Each snippet has a snippet_type — "translation"/"ocr" means it is a verbatim extract from the source text; "summary" means it is AI-generated description (do not quote those as the author's words). Response includes total_matches, returned, and offset for pagination. Cross-cultural tip: for pre-modern or non-Western topics, search source-tradition vocabulary rather than modern English terms — e.g. for seminal economy search "jing" or "bindu" or "istimnāʾ", not "semen retention"; for female homoeroticism search "tribade" or "sahq", not "lesbian". The corpus is indexed via period translations that use tradition-internal terminology.
- 2026-10-12 `list_image_facets` added
  - List the legal values of search_images' type, subject, figure and symbol filters, with how many images carry each. Those filters are EXACT string matches on tags an AI extractor wrote, and each vocabulary runs to thousands of strings (e.g. "musical notes", "Jesus Christ", "medicinal plants"), so a guessed value usually matches nothing. Look it up here first. Call with no facet for the top values of all four; with facet + contains to find the exact spelling of a term ("ouro" → "ouroboros"). Counts cover book illustrations of visible books only (the museum-artwork collection is not counted), and only strings on 3 or more images are listed; the list is refreshed weekly. If contains finds nothing, the tag is absent or rare: use it as free text in search_images' query instead. No URL in this result is a browse page; link images from search_images results.

Full history: https://api.protogrid.dev/v1/servers/io.github.Embassy-of-the-Free-Mind%2Fsourcelibrary/changes

## Remotes

- https://sourcelibrary.org/api/mcp (streamable-http, auth none, reachable true, uptime30d 1)

## Packages

_none_

## Tools (22)

- `check_duplicate`: DOES THE LIBRARY HOLD THIS WORK? Call this BEFORE telling the user the library lacks a book, and before proposing one for acquisition. A miss from list_books or search_library is NOT evidence of absence: list_books matches words and search_library ranks by relevance, so a held book can sit behind dozens of better-ranked hits. Give a title + author (+ year if known), or a library URL (Internet Archive, Gallica, e-rara, BSB, a IIIF manifest). Matches the same scan, the same edition, other editions and the same work, plus near and cross-lingual titles (title in any language). Returns a verdict (same_object / same_edition / possible_same_edition / other_edition / related_title / new), confidence, matches with reasons, ids and URLs, held_not_public (copies held but not yet published: counted, not described), and limits (what the check could not search). A "new" verdict from a URL alone has not searched other editions; pass the title and author too.
- `continue_review_shift`: Next page of the book being reviewed. Call it straight after finishing a page until the stretch the user chose is done, and again when they ask for more. Pass next.book and next.after from the previous page, and volunteer_code. Returns the same shape as start_review_shift; follow how_to_run again. When page is null the book is finished: say so and offer start_review_shift with new_book true.
- `get_book`: READ PIPELINE step 1 — DISCOVER. START HERE for any named work or author. Returns the book's AI-generated summary, chapter list, edition metadata, DOI, page counts, IIIF manifest, and the cover image (inline, so you and the user can see the book). LANGUAGE: `language` is what is printed on THIS EDITION's leaves, which is frequently not the language the work was written in. Where the record knows they differ the response also carries `work_language` and a `translation_note`, and `text_role` (original / period-translation / modern-translation / unclassified) says what kind of witness this is — read them before describing a passage as the author's own words, because an edition can be a translation of a translation (de Slane's 1863 French Muqaddimah, read in English, is English←French←Arabic). Absent `work_language` means the work's language is NOT RECORDED on this record. It is not a statement that the edition is in the work's own language (most records have never been checked; Risner's 1572 Latin Alhazen carries none, and the work is Arabic). Only `text_role: "original"` asserts that these are the author's own words in the author's language; `text_role: "unclassified"` means nobody has examined it, so do not infer the work's language from `language`, and do not conclude the library lacks a tradition because a `language` filter came back empty. Use list_editions to find an original-language witness. The summary is typically a multi-paragraph orientation covering the book's argument, structure, and significance — often answering the question without further searching. Then: get_book_text to read a chapter or page range (step 2), get_quote / get_quotes to lock specific pages with full citation apparatus (step 3). search_within_book locates passages inside this book. MULTI-WORK VOLUMES: where the scans carry running heads, contains_works lists the works the volume ACTUALLY holds with their page spans, taken from the heads the printer put on each leaf. Trust it over the title — collected-works titles routinely name works the volume does not contain, and the volume holding a work often does not name it. If contains_works is absent the scans have no heads to read; status "insufficient-heads" means it was examined and could not be decided. EDITIONS: `editions` says which languages this book can be READ in and how many pages each covers ({ en: 357 } on most books, { en: 357, es: 357 } where a localized edition exists). Pass a code as `lang` to get_quote, get_quotes, get_book_text or search_within_book to read that edition; without it you get English.
- `get_book_text`: READ PIPELINE step 2 — READ. Read a book's text. Call get_book first (step 1) for the chapter list, then come here. Preferred: use the chapter param to read one chapter at a time (includes [Page N] markers for citation). Alternatively, use from/to for explicit page ranges (e.g. from=1 to=50). When you find passages worth quoting, hand the page numbers to get_quote / get_quotes (step 3) for verbatim text + a citation link. TRUNCATION: the response always includes truncated: true/false. When truncated=true, the truncation_note field gives the exact next from/to values to call — this means content was cut short by a page-budget limit, NOT that the book ended. An AI agent MUST NOT infer end-of-book from pages_returned alone; check truncated first. Daily page budgets apply across get_book_text/get_quote/get_quotes: anonymous 500 pages/24h, signed-in 1,000, free Explorer keys 2,000, paid keys uncapped — sign in at sourcelibrary.org/auth/signin or get a key at sourcelibrary.org/developers. Corpus-scale text belongs on the dataset API (sourcelibrary.org/dataset), not on this tool.
- `get_locus`: Turn a CANONICAL CITATION into the actual leaves that carry it. Aristotle is cited by Bekker number (1094a8, 1447a) and Plato by Stephanus number (Rep. 328b) — the references scholarship has used for centuries, which survive re-typesetting and are shareable in a way a scan page never is. USE THIS FIRST whenever a passage arrives as a canonical reference rather than a page: do not try to derive the page yourself from a book's pagination, which is what produced a wrong guess before this tool existed. Bekker numbers are unique across the whole Aristotelian corpus, so the number alone is enough and it also tells you WHICH WORK you are citing. Stephanus numbers restart in each of the three 1578 volumes, so pass work ("Republic", "Timaeus") — without it the response lists the candidate dialogues instead of choosing one. Returns every witness the library holds: the Greek reference edition and, where we have one, an English translation of the same lines, each with its scan page, a reader URL and a quote_api link — so you can compare the original against a translation at one reference. Then call get_quote with the returned book_id + page for the verbatim text and a citable shortlink. LIMITS, stated plainly: a witness is only returned where the reference is PRINTED on that leaf (or, in the two root editions, where a verified constant offset brackets it) — nothing is interpolated, so an empty result means this library holds no anchored leaf there, NOT that the citation is wrong; editions_searched shows what was consulted and the range each covers. Line numbers (the "8" of 1094a8) are not resolved — you get the right leaf and read the line off it. Two works can share a page where one ends and the next begins (Bekker 184 and 1447 are both such joins), and each leaf is filed by the running head printed on it, so a reference at the very start of a work may come back under its predecessor — always read other_works_at_this_reference before concluding a passage is absent. A bare number that exists in both systems returns Aristotle and Plato leaves together; check the system field on each.
- `get_quote`: READ PIPELINE step 3 — CITE. Get the exact verbatim text of a single page plus its citation apparatus. ALWAYS use before putting text in quotation marks. The response headline is citation_link (the stable sourcelibrary.org/q/… shortlink) — present it to the user alongside the quote. Render as:
> [exact translation text, verbatim]
> — [Author], [citation.locator]. [citation_link]
PAGE BREAKS: this corpus is paginated from physical leaves, and nearly one prose page-boundary in five has a sentence running across it — sometimes a word split by a hyphen ("…our move-" / "movements…"). A page that opens or breaks off mid-sentence still reads as complete prose and still carries a perfectly valid citation, so check the continuity field on every response BEFORE quoting: if continues_on_next or continues_from_previous is true, call again with context: true and quote the whole sentence. Quoting a fragment as though it were the author's complete thought is a misattribution even when the page number is right.
NON-LATIN SCRIPTS: where the page is Greek, Hebrew, Arabic, Sanskrit, Cyrillic and so on, the response also carries romanized — the romanization of the original — so the citation can be shown in three layers: original → romanized → translation → citation_link. It is AI-generated reading apparatus, not a transcription; quote the source from original or translation, never from romanized. Absent on Latin-script pages and on non-Latin pages not yet romanized.
ENGLISH ORIGINALS: where the leaf is already English there is no translation and none is needed — the response omits `translation`, sets `text_source: "ocr_original"`, and the verbatim text is `original` (with a `transcription_note`). Quote it as the source's own words, never as a translation, and expect period spelling and long-s (ſ) — it is an uncorrected transcription of the scan. `text_source` is on every response (`translation` otherwise), so branch on it rather than guessing from pages_translated, which is 0 for an English-original book by construction.
TRANSLATED EDITIONS: `original` means the text printed on this leaf, which on a translated edition is the TRANSLATOR's language, not the author's. When the response carries `translation_note`, the chain is stated there — attribute the wording to the translator and do not offer the passage as evidence of what the author wrote in their own tongue. Call list_editions to find an original-language witness of the same work.
For several pages of one book at once, use get_quotes.
- `get_quotes`: READ PIPELINE step 3 — CITE, in batch. Get verbatim text + citation_link for SEVERAL pages of a single book in one round-trip, to assemble a multi-passage dossier. Specify either pages (an explicit array, e.g. [12, 40, 41]) or an inclusive from/to range. Max 25 pages per call. Each entry carries its own citation_link to present alongside the quote, and — on non-Latin-script pages that have one — a romanized layer to show between the original and the translation (AI apparatus, not a transcription). Every entry also carries `text_source`: `translation` normally, or `ocr_original` on a leaf that is already English, where the verbatim text is `original` and must be attributed as the source's own words rather than as a translation. One batch can mix both — a Latin volume can hold an English preface. Batch pages count toward the shared daily page budget (see get_book_text); this is a citation tool, and corpus-scale extraction belongs on the dataset API (sourcelibrary.org/dataset).
- `list_books`: BROWSES/FILTERS THE CATALOG by metadata (canonical author via author_id, title/author/summary words via search, language, category, year range, translation recency) — no page-text matching. PICK THIS to see WHAT EXISTS by an author or in a tradition, or to enumerate a date range. Returns books with title, author (string + canonical author_id + author_url), language, year, and translation progress. → For a relevance-ranked topic search use search_library; for passages on a theme use search_translations (exact words) or search_concept (by meaning).
- `list_editions`: Every edition of one work that the library holds — the other witnesses to the same text, across languages and centuries. Give it a book_id (easiest: the id of any edition you already found) or a work_id. USE THIS when a quotation needs checking against more than one witness, when you want the original-language text behind a translation, or when comparing how a passage reads across editions — differences between witnesses are often the finding. Returns language, date, page count and translation coverage per edition, so you can pick the right one to read. Note: for multi-volume collected works the identifier names the SET rather than a single text, and the response says so explicitly when that applies.
- `list_image_facets`: List the legal values of search_images' type, subject, figure and symbol filters, with how many images carry each. Those filters are EXACT string matches on tags an AI extractor wrote, and each vocabulary runs to thousands of strings (e.g. "musical notes", "Jesus Christ", "medicinal plants"), so a guessed value usually matches nothing. Look it up here first. Call with no facet for the top values of all four; with facet + contains to find the exact spelling of a term ("ouro" → "ouroboros"). Counts cover book illustrations of visible books only (the museum-artwork collection is not counted), and only strings on 3 or more images are listed; the list is refreshed weekly. If contains finds nothing, the tag is absent or rare: use it as free text in search_images' query instead. No URL in this result is a browse page; link images from search_images results.
- `propose_collection`: Propose a themed collection of books to the Source Library team — a title, a rationale (why these books belong together and what thread connects them), and an ordered list of book ids. Get book ids from search_library / list_books / get_book. Like submit_feedback and share_findings, this goes to the team for REVIEW — it does NOT create a public collection instantly; a curator reviews and approves it. Use this when the user has identified a coherent set of books worth grouping and wants to contribute that curation back.
- `propose_correction`: During a review shift, propose a correction to a page's transcription (field "ocr") or English (field "translation"). A correction is a list of exact span edits: each "find" must be copied exactly from the text the shift returned and occur once on the page (add surrounding words if it repeats); "replace" is the corrected text ("" deletes). Fix only what the scan or the original shows is wrong: misread letters, dropped or added words, a sentence the English skips. Do not restyle or modernise. Propose transcription fixes before translation fixes, as separate calls. Pass base_hash = page.text_version for that field. A proposal does NOT change the page: it is applied only after a second reader agrees, and the volunteer is credited. If the call says the page changed since it was loaded, say so and move on. drafted_by: "assistant" if you found and worded it in a book run, "volunteer" if the user supplied the wording, "assistant_accepted" if you drafted it and the user approved it specifically, "mixed" otherwise.
- `search_concept`: RETURNS QUOTABLE PASSAGES matched by MEANING (cosine similarity on Gemini embeddings, 768d) — paraphrases and adjacent phrasings match even with zero keyword overlap. PICK THIS when the modern term won't literally appear in historical texts — e.g. "distributed cognition" maps to passages about active intellect, art of memory, wax tablet metaphors; "social contract" maps to pre-Hobbesian discussions of consent and authority. → For exact words/distinctive terms use search_translations (cheaper, more precise); to list which BOOKS cover a topic use search_library; if the user named an author/work, get_book first (semantic search is expensive — reserve it for cross-corpus discovery). Similarity calibration: 0.70+ is a strong match, 0.55–0.70 is worth reading but verify, below 0.55 is mostly conceptual drift. Set max_per_book to diversify results across many books rather than cluster on one source. Each passage carries a snippet_type — quote only "translation" snippets, never "summary". Cross-cultural tip: for pre-modern or non-Western topics, also try source-tradition vocabulary — e.g. for seminal economy try "jing preservation" or "bindu yoga" or "istimnāʾ"; for masturbation try "mollities" (Latin) or "hastamaithuna" (Sanskrit) or "shouyin" (Chinese). The corpus is indexed via period translations that use tradition-internal terminology, so adjacent/euphemistic terms often surface material that modern English keywords miss.
- `search_images`: Search 200,000+ historical illustrations, emblems, engravings, diagrams, AND 24,000+ artworks (paintings, prints, sculptures). Filter by type, subject, figure, symbol, year. Results interleave two collections: illustrations extracted from book pages (each with a page number and book link) and standalone museum artworks (type: "artwork"). The first few results also return as inline images YOU can see. Hosts that support MCP Apps render an in-chat image gallery for this tool automatically; on other clients images may sit inside the collapsed tool-result view, so never tell the user images are "rendered above" unless the gallery appeared — describe what you see and give each image's url link instead. Every image_url is public and stable — an HTML page that references them directly works in any online browser. If images.length is 0, read the note field — an empty result under a book_id filter means that book has no EXTRACTED images yet, not that the physical book has no plates. A broad query can match tens of thousands (read total): narrow with type/subject/symbol (exact tag strings; look them up with list_image_facets) or page with offset instead of raising limit. On museum-artwork results, a title_is_descriptive flag means the title is an AI description of the picture rather than a title the work was published under — cite such a record by its source_record_title, never by the descriptive one (#4288).
- `search_library`: RETURNS A LIST OF BOOKS (works on a topic) — NOT passages. PICK THIS to discover which works exist on a subject. → For quotable text use search_translations (exact words) or search_concept (by meaning); if the user already named an author/work, call get_book directly (or list_books to find the ID) — the AI summary + chapter outline is usually the right first answer. Searches titles, authors, subjects, and (as a secondary signal) translated text. Query tips: single distinctive words or short phrases work best ("memory palace", "ouroboros"); quoted phrases match exactly. Each result includes total_matches (full count) + returned (this page) + offset for pagination.
- `search_translations`: RETURNS QUOTABLE PASSAGES (page-level snippets + citation URLs), matched by KEYWORD/term. PICK THIS to find a quote or textual evidence on a topic across the whole library. → If the modern word won't literally appear in historical texts, use search_concept (matches by meaning); to list which BOOKS cover a topic use search_library; to dig inside one known book use search_within_book; if the user named an author/work, get_book first (its AI summary is usually the right first read). Query tips: a distinctive term ("wax tablet") or the sentence you are looking for both work; matching is by word, and a page that prints your words together and in order ranks above one that scatters them. Each passage carries phrase_match (0-1, absolute, not relative to the other results): 1 means the page prints your whole query in order at one place, 0 means your words occur only apart; when every passage is under about 0.2 the library probably does not print that sentence, so try search_concept. It is absent for a one-word query and for a passage found by meaning alone. found_by says which lane found the page: "keyword", "semantic", or "both" (the strongest signal). Each snippet has a snippet_type — "translation"/"ocr" means it is a verbatim extract from the source text; "summary" means it is AI-generated description (do not quote those as the author's words). Response includes total_matches, returned, and offset for pagination. Cross-cultural tip: for pre-modern or non-Western topics, search source-tradition vocabulary rather than modern English terms — e.g. for seminal economy search "jing" or "bindu" or "istimnāʾ", not "semen retention"; for female homoeroticism search "tribade" or "sahq", not "lesbian". The corpus is indexed via period translations that use tradition-internal terminology.
- `search_within_book`: SEARCHES INSIDE ONE BOOK (requires book_id). PRIMARILY KEYWORD: it runs a lexical search over the book's pages plus a narrow scoped-semantic pass (top ~10), interleaved by relevance. PICK THIS when you know the wording you are looking for, or want every page of one book mentioning a term. → IF YOU ARE SEARCHING FROM A PARAPHRASE, a half-remembered line, or a modern restatement, USE search_concept INSTEAD — it is the meaning-matching tool and it searches the whole corpus, including translations whose vocabulary differs completely from yours (Thomas Taylor writes "energies" for energeia and "felicity" for eudaimonia, so a sensible modern paraphrase can miss his pages entirely while matching semantically). → To find the book first, use search_library or search_concept, then pass its book_id here. Results are ordered by relevance within this book. Each carries score (0-1, ABSOLUTE, not relative to the other results, so a query the book cannot answer scores low everywhere): the higher of similarity (raw cosine from the meaning pass, calibrated like search_concept: 0.70+ strong, 0.55–0.70 verify, below 0.55 drift) and phrase_match (how much of a multi-word query the page prints in order at one place; 1 is all of it; a one-word query the page prints scores 1). Each also carries found_by ("keyword", "semantic", or "both" — both is the strongest signal). Results flagged is_front_matter are the translator's or publisher's words rather than the author's, and are ordered last. Returns OCR and translation snippets with page numbers, ready to cite.
- `share_findings`: Share a research dossier back to the Source Library team — a title, an optional summary, and an ordered list of citations (the passages your thesis rests on). Each citation is a reference { book_id, page, note }, NOT copied text: the library re-renders the canonical quote from the reference, so links stay authoritative. Use this when the user has assembled a thesis backed by passages across one or more books and wants to contribute it back. Like submit_feedback, this goes to the team for review (not an instant public page).
- `start_blind_check`: OPTIONAL lane, only for a user who reads the original language and asks to judge pages themselves: up to five random pages from different books, each with scan, transcription and English. Here the USER is the reviewer: ask for THEIR verdict first, before any opinion of yours, then give your reading with exact quotes, let them accept or reject each point, and call submit_page_review with mode "blind" and their final verdict. Never choose the verdict for them. For everyone else use start_review_shift.
- `start_review_shift`: Start a volunteer REVIEW SHIFT: the user lends their Claude (you) to Source Library to check one book, page by page, against its scans. Use when the user wants to help Source Library, says "start a shift" or "review pages", or asks how they can contribute. The user does NOT need to read the book's language: YOU do the checking (scan against transcription, transcription against English) and the user starts the shift, chooses how many pages, watches and asks questions. Returns ONE page: its scan (inline image), our transcription, our English, the last lines of the previous page, and next (pass to continue_review_shift). Follow first_time once, then how_to_run on every page, working through the stretch of pages without stopping between them. With no book, it resumes the user's last run (pass volunteer_code) or picks a book that needs checking; book takes an id, slug or reader URL if the user names one; new_book true skips the resume. Run shifts only while the user is in this conversation; never set one up to run on a schedule or from a script.
- `submit_feedback`: Submit feedback, bug reports, or feature requests to the Source Library team. Before proposing NEW functionality, read https://sourcelibrary.org/llms.txt and https://sourcelibrary.org/developers — several past submissions proposed building things that already exist (IIIF manifests, Content Search, DTS), which wastes reviewer time. State in the report which docs you checked. Bug reports with record IDs and reproducing queries are the most actionable kind.
- `submit_page_review`: Record the verdict on one page of a shift. mode "book-run" (start_review_shift / continue_review_shift): the verdict and findings are YOURS, from reading the scan and both texts; set user_looked true if the user checked anything on the page themselves. mode "blind" (start_blind_check): the verdict is the USER'S, given before your reading; put yours in assistant_verdict. Verdicts: both_sound (transcription matches the scan and the English matches the original), translation_drift (transcription right, English departs from it), transcription_off (the text does not match the scan), both_off, unclear (the scan or the text cannot be judged). Quote exact words in every finding. A review is evidence about the page; it does not change it.