mcp-books
Search and read records from the Internet Archive, Library of Congress, and BnF through one MCP server, without API keys.
Search inside scanned text for phrases, with excerpts, page numbers when available, and separate per-archive status.
Search catalogues by title, creator, subject, or plain words; filter by media type and year; sort within each archive.
Read one record by id in a uniform shape, selecting description, subjects, copies, context, and paged text windows.
Combine results across archives while keeping counts and failures per source, with derived queries, pagination, and attribution/provenance.
Allows searching the machine-read text of digitised books, periodicals, and documents; searching catalogue records; and retrieving item details from the Internet Archive.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@mcp-books Search for the phrase 'space race' in books and newspapers"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
mcp-books
Three great archives hold the scanned record of what was published, and each describes it in its own words. The Internet Archive keeps books, films, recordings and software deposited by anyone, and has run millions of them through optical character recognition. The Library of Congress publishes the national collections of the United States, one catalogue per kind of material. data.bnf.fr publishes the authority records of the Bibliothèque nationale de France, which describe works and the people who wrote them rather than copies.
This server reads all three with one question. You can search the words inside the scanned documents, search the catalogues, and read one record in a single shape whichever archive holds it. It needs no API key and no account.
Install
One-click install
Claude Code
claude mcp add books -- npx -y mcp-booksClaude Desktop, Cursor, and any client using the standard config format
{
"mcpServers": {
"books": {
"command": "npx",
"args": ["-y", "mcp-books"]
}
}
}Node 24 or later is required, and no environment variable has to be set.
With Docker
{
"mcpServers": {
"books": {
"command": "docker",
"args": ["run", "-i", "--rm", "ghcr.io/smeet666/mcp-books:2.1.0"]
}
}
}-i keeps stdin open, which is where the protocol travels, and -t is left out
because a TTY rewrites the stream. The container needs outbound HTTPS to
archive.org, openlibrary.org, www.loc.gov and data.bnf.fr, and nothing
else: no volume, no port, no credential.
Bundle, without npm
Download mcp-books-2.1.0.mcpb from
the latest release and
open it. A client that supports MCP bundles installs it on its own, with no npm
and no configuration file to edit. The bundle carries its dependencies, so
nothing is fetched at install time.
Related MCP server: anna-book-search
What you can ask
"Which books mention the Beaumont light-house?"
"Find me anything on the 1906 San Francisco earthquake."
"Read me that record and tell me who holds the original."
"What does the BnF have on that author?"
"Search the photographs rather than the books."
An answer takes several seconds: three archives are asked, each at its own pace.
The three sources
Source | Archive | What it describes |
| the Internet Archive | deposited copies, of every kind |
| the Library of Congress | the national collections, one catalogue per kind |
| the Bibliothèque nationale de France | works and the people who wrote them |
A row's id names its archive, so an identifier read from one answer goes back
to the right one. Counts are never added across archives, and an archive that
failed is reported as having failed rather than as having found nothing.
Tools
Tool | What it does |
| Searches the words inside the scanned documents. |
| Searches the catalogues by title, creator, subject or plain words. |
| Reads one record in a single shape, whichever archive holds it. |
search_inside
Searches the text inside the scanned documents, which came off the page through optical character recognition.
Argument | Type | Required | What it does |
| string, 2 to 300 characters | yes | The phrase to look for inside the documents. |
| integer, 1 to 25, default | no | Matches to keep from each archive. |
| integer, 1 to 100, default | no | Which page of matches. |
| integer, 80 to 1200, default | no | How much of a passage to serve. |
| integer, 1 to 10, default | no | Passages served per matching document. |
| boolean, default | no | Ask every archive rather than stopping at the first that answers. |
| array of source ids | no | Ask these archives alone. |
In return: hits, each carrying id, which get_item takes and which names
its archive; source and source_name; the archive's own identifier without
the prefix; title, creator and year; page_number where the archive states
one; excerpts; and excerpt_kind.
excerpt_kind decides what an excerpt is worth. A passage is the text
around the words that matched. A page_opening is the start of the page, sent
because the machine-read text the archive returned stops before those words
appear: it does not carry the match, so quoting it quotes something else. All the
excerpts of one match are of one kind.
search_items
Searches the catalogues.
Argument | Type | Required | What it does |
| string, 1 to 300 characters | yes | A title, a creator, a subject, or plain words. |
| a kind one of the archives holds | no | Which kind of material to search. |
| integer, 1000 to 2100 | no | Earliest year. |
| integer, 1000 to 2100 | no | Latest year. |
|
| no | How the rows are ordered. |
| integer, 1 to 25, default | no | Rows to keep from each archive. |
| integer, 1 to 100, default | no | Which page of rows. |
| boolean, default | no | Ask every archive. |
| array of source ids | no | Ask these archives alone. |
The three archives divide their material differently. The Internet Archive
searches every kind at once when none is named; the Library of Congress is one
route per kind, so a search naming none is told which one was read; and the BnF
search reads works. A media_type one archive has no notion of leaves that
archive out, and the answer says so.
In return: rows in the shape a hit carries, with per_source giving one
report per archive: its status, the count it contributed, its
reported_total and reported_total_means, which says what that number counts
there.
get_item
Reads one record in a single shape, whichever archive holds it.
Argument | Type | Required | What it does |
| string, 1 to 500 characters | yes | The |
| array of | no | Which parts to return. |
| integer, 1 to 50, default | no | Copies to list. |
| integer, 0 to 1000000, default | no | Where to resume the text. |
| integer, 200 to 8000, default | no | How much text to serve. |
In return: the record with its id, source and source_name, the
archive's own identifier, title, creator, date exactly as published, and
year beside year_means, which says what that year is the year of, since the
three archives date a record differently. attribution is what that archive asks
to be credited with, and identifier_provisional says when the identifier was
built rather than read, so a caller knows it may not resolve.
What an answer states about the archives
Every answer accounts for each archive separately. One that failed, one nobody asked, and one that answered with nothing are three different things, and they are reported as three. A total stays beside the archive that published it, with what that archive counts when it says it: one counts documents, another counts newspaper leaves.
What scanned text is worth
The words inside a scanned document came off the page through optical character recognition. An excerpt carries the misreadings of that process, and it is served as it was read rather than corrected. Quote it as scanned text, and link the record so a reader can look at the page.
Configuration
Every variable is optional. Set them in the env block of your client config.
Variable | Default | What it does |
| the project identity | Names your application to the three archives, with an address where a person can be reached. |
| each archive's own pace | Widens the gap between two requests to one archive, from 500 to 60000. Left unset, every archive keeps the pace it publishes, and a figure set here applies only where it is wider. |
|
| Deadline for one request, from 1000 to 120000. |
|
| Attempts after a transient failure, from 0 to 8. |
|
| How long an answer stays in memory, from 0 to 86400000. |
|
| Answers held in memory at once, from 1 to 5000. |
|
|
|
A value outside its range falls back to the default, and the reason is written to stderr.
Errors
Every failure carries one of six codes, a message, and where it helps a hint naming the next move.
Code | What happened | What to do |
| An archive answered, and holds no such record. | Check the identifier with |
| The arguments were refused before any request went out. | Read the message, which names the argument. |
| An archive asked this client to slow down. | Wait, then call again with the same arguments. The record is still there. |
| An answer arrived in a shape this client cannot read. | Report it at the issue tracker. |
| The request did not complete. | Try again shortly. |
| The request passed its deadline. | Raise |
An archive that failed is reported per archive rather than failing the whole answer, so one silent archive never hides the others.
As a library
The layer reading the three archives is published on its own, with its pacing, its cache and its errors, and with no protocol attached.
import { BooksClient } from "mcp-books/client";
const client = new BooksClient();
const read = await client.searchItems({ query: "beaumont light-house", limit: 3 });
console.log(read.data.rows.length);Each read answers { data, cached }, and throws an error carrying one of the six
codes. Each archive keeps its own pace, and its floor holds here as well.
Pacing and attribution
Each archive is paced on its own, one request at a time, and the widest of its
own floor and the configured interval governs: the Library of Congress publishes
the slowest, and asking all three at once therefore costs each of them one
request rather than three. The User-Agent always ends with the project identity
and an address where a person can be reached.
Every record carries the address of its page and the attribution its archive
asks for. The Internet Archive items belong to their depositors, the Library of
Congress records state their own rights, and the BnF asks that the source and the
date of retrieval be stated wherever its metadata are shown.
This MCP server is an unofficial project, with no affiliation to any of the archives it reads.
Privacy
This server collects nothing about you and sends nothing to its author. It runs
on your machine, contacts archive.org, openlibrary.org, www.loc.gov and
data.bnf.fr and nothing else, holds its answers in memory while it runs, and
writes nothing to disk. PRIVACY.md states what a request carries
and which settings change any of it.
Development
npm install
npm run build:fixtures
npm test
npm run checkTests run against generated fixtures and make no network request. The live suite,
npm run test:live, makes one request per route and runs nightly against the
archives themselves.
Contributing
Bugs, questions and ideas belong in the issue tracker. Pull requests are welcome; opening an issue first helps agree on the shape of the change. See CONTRIBUTING.md.
License
MIT, see LICENSE. The records belong to the archives that published them and to their depositors.
mcp-books (français)
Trois grandes archives conservent la trace numérisée de ce qui a été publié, et chacune la décrit dans ses propres mots. L'Internet Archive garde les livres, les films, les enregistrements et les logiciels que chacun y dépose, et en a passé des millions par la reconnaissance optique de caractères. La Library of Congress publie les collections nationales des États-Unis, un catalogue par type de document. data.bnf.fr publie les notices d'autorité de la Bibliothèque nationale de France, qui décrivent des œuvres et ceux qui les ont écrites plutôt que des exemplaires.
Ce serveur lit les trois avec une seule question. On peut chercher dans les mots contenus dans les documents numérisés, chercher dans les catalogues, et lire une notice sous une forme unique quelle que soit l'archive qui la détient. Aucune clé d'API, aucun compte.
Installation
Installation en un clic
Claude Code
claude mcp add books -- npx -y mcp-booksClaude Desktop, Cursor, et tout client au format de configuration standard
{
"mcpServers": {
"books": {
"command": "npx",
"args": ["-y", "mcp-books"]
}
}
}Node 24 ou plus récent est nécessaire, et aucune variable d'environnement n'est à renseigner.
Avec Docker
{
"mcpServers": {
"books": {
"command": "docker",
"args": ["run", "-i", "--rm", "ghcr.io/smeet666/mcp-books:2.1.0"]
}
}
}-i garde l'entrée standard ouverte, qui est le canal du protocole, et -t est
omis parce qu'un TTY réécrit le flux. Le conteneur a besoin d'un accès HTTPS
sortant vers archive.org, openlibrary.org, www.loc.gov et data.bnf.fr, et
de rien d'autre : aucun volume, aucun port, aucun identifiant.
Bundle, sans npm
Téléchargez mcp-books-2.1.0.mcpb depuis
la dernière publication
et ouvrez-le. Un client qui gère les bundles MCP l'installe seul, sans npm et
sans fichier de configuration à modifier. Le bundle emporte ses dépendances, donc
rien n'est téléchargé à l'installation.
Ce qu'on peut demander
« Quels livres mentionnent le phare de Beaumont ? »
« Trouve-moi ce qu'il y a sur le tremblement de terre de San Francisco en 1906. »
« Lis-moi cette notice et dis-moi qui conserve l'original. »
« Qu'est-ce que la BnF a sur cet auteur ? »
« Cherche dans les photographies plutôt que dans les livres. »
Une réponse prend plusieurs secondes : trois archives sont interrogées, chacune à son rythme.
Les trois sources
Source | Archive | Ce qu'elle décrit |
| l'Internet Archive | les exemplaires déposés, de tout type |
| la Library of Congress | les collections nationales, un catalogue par type |
| la Bibliothèque nationale de France | les œuvres et ceux qui les ont écrites |
L'id d'une ligne nomme son archive, donc un identifiant lu dans une réponse
retourne vers la bonne. Les comptes ne sont jamais additionnés entre
archives, et une archive qui a échoué est rapportée comme ayant échoué plutôt
que comme n'ayant rien trouvé.
Les outils
Outil | Ce qu'il fait |
| Cherche dans les mots contenus dans les documents numérisés. |
| Cherche dans les catalogues par titre, auteur, sujet ou mots simples. |
| Lit une notice sous une forme unique, quelle que soit l'archive. |
search_inside
Cherche dans le texte contenu dans les documents numérisés, texte issu de la reconnaissance optique de caractères.
Argument | Type | Requis | Ce qu'il fait |
| chaîne, 2 à 300 caractères | oui | La phrase à chercher dans les documents. |
| entier, 1 à 25, défaut | non | Correspondances à garder de chaque archive. |
| entier, 1 à 100, défaut | non | Quelle page de correspondances. |
| entier, 80 à 1200, défaut | non | La longueur de passage à servir. |
| entier, 1 à 10, défaut | non | Passages servis par document correspondant. |
| booléen, défaut | non | Interroger chaque archive plutôt que s'arrêter à la première qui répond. |
| tableau d'identifiants de source | non | N'interroger que ces archives. |
En retour : hits, chacun portant id, que get_item reprend et qui nomme
son archive ; source et source_name ; l'identifier propre à l'archive, sans
le préfixe ; title, creator et year ; page_number là où l'archive en
indique un ; excerpts ; et excerpt_kind.
excerpt_kind décide de ce que vaut un extrait. Un passage est le texte
autour des mots trouvés. Un page_opening est le début de la page, envoyé parce
que le texte lu par machine que l'archive a rendu s'arrête avant que ces mots
apparaissent : il ne porte pas la correspondance, donc le citer cite autre chose.
Tous les extraits d'une correspondance sont d'un seul type.
search_items
Cherche dans les catalogues.
Argument | Type | Requis | Ce qu'il fait |
| chaîne, 1 à 300 caractères | oui | Un titre, un auteur, un sujet, ou des mots simples. |
| un type que l'une des archives détient | non | Le type de document à chercher. |
| entier, 1000 à 2100 | non | Année la plus ancienne. |
| entier, 1000 à 2100 | non | Année la plus récente. |
|
| non | L'ordre des lignes. |
| entier, 1 à 25, défaut | non | Lignes à garder de chaque archive. |
| entier, 1 à 100, défaut | non | Quelle page de lignes. |
| booléen, défaut | non | Interroger chaque archive. |
| tableau d'identifiants de source | non | N'interroger que ces archives. |
Les trois archives découpent leurs fonds différemment. L'Internet Archive cherche
dans tous les types à la fois quand aucun n'est nommé ; la Library of Congress a
une route par type, donc une recherche qui n'en nomme aucun se voit dire lequel a
été lu ; et la recherche de la BnF lit des œuvres. Un media_type dont une
archive n'a pas la notion l'écarte de la réponse, et la réponse le dit.
En retour : des lignes dans la forme d'un hit, avec per_source qui donne
un rapport par archive : son status, le count qu'elle a fourni, son
reported_total et reported_total_means, qui dit ce que ce nombre compte
là-bas.
get_item
Lit une notice sous une forme unique, quelle que soit l'archive qui la détient.
Argument | Type | Requis | Ce qu'il fait |
| chaîne, 1 à 500 caractères | oui | L' |
| tableau de | non | Les parties à rendre. |
| entier, 1 à 50, défaut | non | Exemplaires à lister. |
| entier, 0 à 1000000, défaut | non | Où reprendre le texte. |
| entier, 200 à 8000, défaut | non | La longueur de texte à servir. |
En retour : la notice avec son id, source et source_name,
l'identifier propre à l'archive, title, creator, date exactement telle
que publiée, et year accompagné de year_means, qui dit de quoi cette année
est l'année, les trois archives datant une notice différemment. attribution est
ce que cette archive demande qu'on lui crédite, et identifier_provisional dit
quand l'identifiant a été construit plutôt que lu, pour qu'un appelant sache
qu'il peut ne pas résoudre.
Ce qu'une réponse dit des archives
Chaque réponse rend compte de chaque archive séparément. Une qui a échoué, une que personne n'a interrogée et une qui a répondu vide sont trois choses différentes, et elles sont rapportées comme trois. Un total reste à côté de l'archive qui l'a publié, avec ce que cette archive compte en le disant : l'une compte des documents, une autre des feuillets de journaux.
Ce que vaut un texte numérisé
Les mots contenus dans un document numérisé sont issus de la reconnaissance optique de caractères. Un extrait porte les erreurs de lecture de ce procédé, et il est servi tel qu'il a été lu plutôt que corrigé. Citez-le comme un texte numérisé, et liez la notice pour qu'un lecteur puisse regarder la page.
Configuration
Chaque variable est facultative. Elles se posent dans le bloc env de la
configuration du client.
Variable | Défaut | Ce qu'elle fait |
| l'identité du projet | Nomme votre application auprès des trois archives, avec une adresse où joindre une personne. |
| le rythme propre à chaque archive | Élargit l'écart entre deux requêtes vers une même archive, de 500 à 60000. Non posée, chaque archive garde le rythme qu'elle publie, et une valeur posée ici ne s'applique que là où elle est plus large. |
|
| Délai d'une requête, de 1000 à 120000. |
|
| Tentatives après un échec passager, de 0 à 8. |
|
| Durée pendant laquelle une réponse reste en mémoire, de 0 à 86400000. |
|
| Réponses gardées en mémoire à la fois, de 1 à 5000. |
|
|
|
Une valeur hors de sa plage retombe sur le défaut, et la raison est écrite sur la sortie d'erreur.
Erreurs
Chaque échec porte un des six codes, un message, et quand cela aide une indication du geste suivant.
Code | Ce qui s'est passé | Que faire |
| Une archive a répondu, et n'a pas cette notice. | Vérifiez l'identifiant avec |
| Les arguments ont été refusés avant toute requête. | Lisez le message, qui nomme l'argument. |
| Une archive demande à ce client de ralentir. | Attendez, puis rappelez avec les mêmes arguments. La notice est toujours là. |
| Une réponse est arrivée dans une forme illisible ici. | Signalez-le sur le suivi d'incidents. |
| La requête n'a pas abouti. | Réessayez sous peu. |
| La requête a dépassé son délai. | Augmentez |
Une archive qui échoue est rapportée archive par archive plutôt que de faire échouer toute la réponse, donc une archive silencieuse n'en cache jamais d'autres.
Comme bibliothèque
La couche qui lit les trois archives est publiée seule, avec son rythme, son cache et ses erreurs, sans protocole attaché.
import { BooksClient } from "mcp-books/client";
const client = new BooksClient();
const read = await client.searchItems({ query: "beaumont light-house", limit: 3 });
console.log(read.data.rows.length);Chaque lecture répond { data, cached }, et lève une erreur portant un des six
codes. Chaque archive garde son propre rythme, et son plancher tient également
ici.
Rythme et attribution
Chaque archive est cadencée pour elle-même, une requête à la fois, et c'est le
plus large de son propre plancher et de l'intervalle configuré qui gouverne : la
Library of Congress publie le plus lent, et interroger les trois à la fois coûte
donc à chacune une requête plutôt que trois. Le User-Agent se termine toujours
par l'identité du projet et une adresse où joindre une personne.
Chaque notice porte l'adresse de sa page et l'attribution que son archive
demande. Les documents de l'Internet Archive appartiennent à ceux qui les ont
déposés, les notices de la Library of Congress énoncent leurs propres droits, et
la BnF demande que la source et la date de récupération soient indiquées partout
où ses métadonnées sont montrées.
Ce MCP est un projet non officiel, sans affiliation à aucune des archives qu'il lit.
Confidentialité
Ce serveur ne collecte rien sur vous et n'envoie rien à son auteur. Il tourne sur
votre machine, ne joint que archive.org, openlibrary.org, www.loc.gov et
data.bnf.fr, garde ses réponses en mémoire le temps qu'il tourne, et n'écrit
rien sur le disque. PRIVACY.md dit ce qu'une requête emporte et
quels réglages changent cela.
Développement
npm install
npm run build:fixtures
npm test
npm run checkLes tests s'exécutent sur des fixtures engendrées et n'émettent aucune requête.
La suite en direct, npm run test:live, émet une requête par route et tourne
chaque nuit contre les archives elles-mêmes.
Contribuer
Les anomalies, les questions et les idées ont leur place dans le suivi d'incidents. Les propositions de modification sont bienvenues ; ouvrir un ticket d'abord aide à s'accorder sur la forme du changement. Voir CONTRIBUTING.md.
Licence
MIT, voir LICENSE. Les notices appartiennent aux archives qui les ont publiées et à ceux qui les y ont déposées.
Available Tools
3 toolsget_itemRead one record, routed by the archive its identifier namesARead-onlyIdempotent
Read one record in full from the archive its identifier names: what it is, who made it, when, what the archive says about it, and what a reader can open. 'identifier' must come from search_inside or search_items. It names the archive, so this reads the right one without guessing; a string no archive would have minted is refused, and a shape more than one archive mints is refused rather than sent to a guess, because sending it anywhere answers about the wrong thing. Terms of reuse come back on every read and belong to that record alone. A record stating none has granted nothing, and silence is never read as permission. 'sections' decides what else comes back, and the answer names what was left out and what this archive files nothing under: a field empty because nobody asked for it is a different thing from a field the archive never fills. 'description_means' says what the archive files under the field this reads a description out of, which holds an account of the thing on one record and a line of the catalogue entry on the next. Long prose is returned one window at a time: 'text_offset' says where to resume, and a window ends at a line boundary. An offset past the end says so rather than answering with an empty description. An answer can take several seconds, because one of the archives publishes a request ceiling this server keeps to.
Inputs: identifier (required), sections, max_copies, text_offset, +1 more. Returns: item, notes, id_read_as, text_window, +3 more
search_insideFind a phrase in the scanned text of every archiveARead-onlyIdempotent
Search the machine-read text of every archive this server reads that holds any, at the same time, and get one list of where a phrase was printed. The archives behind this tool hold different material, so the list is additive: it is the places a phrase appears in each corpus, put together, rather than the same question answered twice. 'per_source' names what each corpus is. A match carries 'page_number', which is a real leaf on an archive whose index holds one and null on an archive whose index holds none. That null is the index having no leaf, never a page this server dropped, and no page is ever invented. A match also carries 'excerpt_kind'. 'passage' is the text around the words that matched. 'page_opening' is the start of the page, sent because the machine-read text that came back stops before the searched words appear, so it does not carry the match. The notes say how many excerpts are openings. Every count is that archive's own and counts something of its own: documents in one place, leaves in another. They are never added together, and there is no total across archives. Rows are interleaved one archive at a time. Nothing ranks them against each other and nothing orders them by date, because a year is measured on different things in each archive. Matches whose excerpt carries the searched words are placed before matches whose excerpt is a 'page_opening' and carries them nowhere. That rests on what each row states about its own excerpt rather than on any score, no match is ever dropped for it, and the interleaving holds inside each of the two groups. The archives read the words given in different ways, and 'per_source' says which each one does. An index that answers only where every word appears returns nothing for a question written as a sentence, even on a work it holds several copies of; an index that scores the words instead answers such a question with the pages it ranks highest, which can carry only some of them. Either way it is the words: a character that is neither a letter nor a digit is no word to an index, and 'non_word_characters' lists any the query carried. Shorter and differently spelled wordings are therefore derived from the query and asked for their union, which costs nothing extra when the words as asked already answer. Every wording sent is named in 'per_source' with what it returned, every match carries the wording that returned it in 'found_by_query', and 'fan_out' turns the derivation off. Use search_items for a work by its title, its creator or its subject: this tool reads the text on the pages and knows nothing of a catalogue, so a title given here finds every book that happens to mention it and misses the book itself. Answers take several seconds, because one of the archives publishes a request ceiling this server keeps to. A slow answer is the pacing, not a stall.
Inputs: query (required), page, limit, fan_out, +3 more. Returns: hits, page, notes, order, +6 more
search_itemsSearch every catalogue at onceARead-onlyIdempotent
Search the catalogue of every archive this server reads, at the same time, for a title, a creator or a subject, and get one merged list. Each row carries the id get_item takes, and that id names the archive it came from, so nothing has to be guessed afterwards. 'media_type' keeps one name across the archives and a vocabulary per archive, because the same word does not name the same set of things twice. An archive that files nothing under the name you give is not asked and is named as absent, with its own names listed, rather than asked under a translation. An archive that keeps one catalogue per kind of material is asked for its default when you name none, and the answer says which catalogue that was. Rows are interleaved one archive at a time. No score orders them against each other, and 'sort' is applied inside each archive rather than across them: a year is the date of an edition in one place and the date on a catalogue record in another, so there is no date order that spans the answer. 'oldest' and 'newest' order on a date field carrying a year and no era, so a date before the common era is filed there as a year of this one, and a record stating no date is placed by a stand-in rather than by its age. The first row of a date order is therefore not established as the oldest or newest thing an archive holds, the notes count the rows carrying no year, and this server orders nothing itself. Every count in 'per_source' is that archive's own and counts something of its own. They are never added together, and there is no total across archives. The catalogues read the words given in different ways, and 'per_source' says which each one does. One answers only where every word appears, so a question written as a sentence comes back empty; another scores the words and answers with the records it ranks highest, so a row of its can carry only some of them. Either way it is the words: a character that is neither a letter nor a digit is no word to an index, and 'non_word_characters' lists any the query carried. A name is also filed under more than one spelling, so further wordings are derived from the query and asked for their union. It costs nothing extra when the words as asked already answer. Every wording sent is named in 'per_source' with what it returned, every row carries the wording that returned it in 'found_by_query', and 'fan_out' turns the derivation off. A row's 'media_type' is the word that record carries for the kind of thing, which is often none of the names this argument takes: those are the divisions of a catalogue, and 'media_types' publishes them per archive. A row states no terms of reuse. Read the record with get_item for what that record itself says, and read silence as silence. Use search_inside for a phrase printed on a page: this tool reads catalogue records and knows nothing of what a book says, so a sentence given here matches only where a catalogue happens to carry it. Answers take several seconds, because one of the archives publishes a request ceiling this server keeps to.
Inputs: query (required), page, sort, limit, +5 more. Returns: page, items, notes, order, +6 more
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 3 tools
The three tools target clearly distinct layers: catalogue metadata search (search_items), full-text page search (search_inside), and full record retrieval by identifier (get_item). Their descriptions explicitly warn against using one where the other is appropriate, especially catalogue title search versus page-phrase search. An agent should be able to select correctly despite the verbose descriptions.
All names are snake_case and begin with a clear verb: search_items, search_inside, get_item. The pattern is mostly consistent, though search_inside uses an adverb/preposition rather than a noun object, which is a minor deviation from the verb_noun shape.
Three tools are well matched to a read-only archive search server: one for catalogue discovery, one for full-text discovery, and one for record retrieval. Each has a distinct role, and there are no redundant or filler tools.
The surface covers discovery through catalogue search, full-text search, and detailed record retrieval, which is the core lifecycle for this read-oriented domain. Minor gaps remain, such as explicit archive enumeration or browsing, but the search responses expose source context and identifiers that make the core workflow workable.
Maintenance
Related MCP Connectors
Unofficial Archive MCP — wraps the Internet Archive APIs (free, no auth)
Internet Archive (archive.org) item search & metadata MCP.
MCP server for Project Gutenberg: 78,000+ public-domain ebooks with full plain-text retrieval.
One MCP server for all web data. Scrape, search, crawl, and AI research in one install.
Related MCP Servers
- AlicenseAqualityDmaintenanceMCP server for the Internet Archive's Wayback Machine. Search archived snapshots, extract page text from a specific date, track how a site has changed over time, check if broken links are recoverable, and perform research across Internet Archive collections.63MIT
- AlicenseNot gradedqualityCmaintenanceSelf-hosted MCP server for searching and discovering books using Anna's Archive and Goodreads datasets, enabling full-text search, ISBN/md5 lookup, similarity matching, and optional download URL retrieval.12 npm4MIT
- AlicenseAqualityCmaintenanceMCP server for searching and retrieving full-text pages from the Library of Congress, including newspapers, books, and manuscripts, via the loc.gov API.3Apache 2.0
- AlicenseNot gradedqualityBmaintenanceFull-coverage MCP server for Internet Archive, enabling search, metadata lookup, collection browsing, and Wayback Machine snapshot retrieval via 13 tools.2BSD Zero Clause