heritage-research-mcp
Search six public archives and a Civil War newspaper from one MCP server, returning normalised records with rights statements and provenance.
Search every ready archive at once (
search) or one at a time, optionally filtered by materialkind(text, image, map, audio, video) and date range.Find words inside documents: Internet Archive OCR (
ia_fulltext_search,ia_grep_text,ia_read_text), NARA extracted text, newspaper page OCR, and Dispatch article text.Query individual archives: Internet Archive (metadata, items, files), Wikimedia Commons (files, categories, licences), DPLA, NARA Catalog (records, children, textual records), and Smithsonian Open Access (search, content, vocabulary, stats).
Search digitised US newspapers through Chronicling America — either the live loc.gov API (
chronam_search,chronam_get_page,chronam_page_text) or a local bulk-OCR index over whole years — and the Richmond Daily Dispatch via its local SQLite index (1860–1865).Fetch a single record in the normalised shape with its media list and rights (
get_record), including archival citations from NARA.Download an original file to the local cache with a
.provenance.jsonsidecar (source, URL, retrieval time, SHA-256, stated rights) viadownload_mediaor the CLI downloader.Inspect readiness and usage:
list_sourcesshows which sources are configured,usage_reportcounts requests and rate limits.DPLA, NARA and Smithsonian need free API keys; Internet Archive, Commons, Chronicling America and the Dispatch work without one. Virginia Chronicle is deliberately not included.
Provides tools for searching the Internet Archive's scanned books, memoirs, regimental histories and official reports, including searching inside them. Offers metadata search (ia_search), full-text search across OCR text (ia_fulltext_search), item file listings (ia_get_item), and reading or grepping OCR text (ia_read_text, ia_grep_text) to locate names and passages with page numbers and snippets.
Provides tools for finding freely licensed photographs, maps and prints, with the licence read from each file. Offers file search (commons_search), detailed file metadata including licence and image dimensions (commons_file_info), and category browsing (commons_category_members).
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., "@heritage-research-mcp search for a Civil War map of Fairfax County, Virginia"
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.
heritage-research-mcp
One MCP server for historical research across six public archives and a Civil War newspaper. Ask an AI assistant to find a period map, a photograph, a memoir passage or a federal record, and get back records in one shape, each with the holder's rights statement and a provenance trail for anything downloaded.
Source | Good for | Key |
Scanned books, memoirs, regimental histories, official reports, and the full text inside them | none | |
Freely licensed photographs, maps and prints, with the licence read from each file | none | |
Finding items held by US libraries, archives and museums | free, emailed to you | |
Federal records, military and census material, with OCR text for many scans | free, by email request | |
Objects, photographs, archives and library items; CC0 images with direct high-resolution files | free, self-service | |
Chronicling America (Library of Congress) | Digitised US newspaper pages, 1690 to the present, with raw OCR text | none; rate limited |
The full text of a Richmond newspaper, 1860 to 1865 (1,384 issues), searched from a local index | none; one 92 MB download |
It was built for sourcing reference material where "where did this come from, and may I reuse it?" matters as much as finding it.
Quick start
Install with the one command under Install. The Internet Archive and Commons work straight away, with no key.
Add the free keys you want (DPLA, NARA, Smithsonian). docs/KEYS.md has the steps for each, and a source with no key is simply skipped.
Check it:
heritage-research-mcp doctor --livelists each source and makes one small real request to it.Ask your assistant something like:
"Find 1860s maps of Fairfax County, Virginia, and tell me which I may reuse."
"Find the passage in Mosby's memoirs about the capture at Herndon Station, and quote it with the page."
"Find Brady photographs of Union cavalry in the National Archives and give me archival citations."
"Download the highest-resolution scan of that map, with its provenance."
Related MCP server: Arke Institute MCP Server
Which source for what
You want | Start with |
A name, place or phrase inside a book, memoir or official report |
|
A photograph, map or print you may reuse |
|
Federal records: military files, War Department maps, Brady's photographs, pension files, census |
|
What US newspapers printed about a name, place or event |
|
What a Richmond newspaper said, day by day, 1860 to 1865 |
|
Something held by a library, archive or museum near the place you care about |
|
Objects and photographs from the Smithsonian museums |
|
Not sure |
|
What it looks like
Ask for a Civil War map of Fairfax County, Virginia, and search with kind: "map" returns, among others:
{
"source": "commons",
"title": "A Civil War field map of Fairfax County, Virginia with Fort Coccoran. LOC 2014588392",
"date": "1861",
"description": "Shows paths and roads that no longer exist and names of some landowners. …",
"kind": "map",
"rights": { "reuse": "free", "label": "Public domain", "basis": "holder" },
"media": [{ "kind": "image", "mime": "image/jpeg", "bytes": 3960781, "width": 4615, "height": 6169 }],
"landing_url": "https://commons.wikimedia.org/wiki/File:A_Civil_War_field_map_of_Fairfax_County,_Virginia_with_Fort_Coccoran._LOC_2014588392.jpg"
}And ia_fulltext_search finds a name inside a scanned book, with the page and a snippet (matches in **bold**):
Mosby's War Reminiscences (1887): "…Capture of a Federal Picket at Herndon Station. The Dash and Excitement of a Cavalry Skirmish…"
ia_grep_text then finds every occurrence in that book and ia_read_text reads around one. download_media caches an original (a 7 MB high-resolution CC0 photograph, say) and writes a .provenance.json beside it.
Searching by type
search takes kind: text, image, map, audio or video. Each archive is asked in its own vocabulary, every result says how (kind_applied), and every record carries a best-effort kind of its own.
kind | Internet Archive | Commons | DPLA | Smithsonian | NARA |
|
|
|
| scanned books, full-text documents, books, manuscripts | Textual Records |
|
|
|
|
| Photographs and other Graphic Materials |
| the "maps" subject and the map collections |
| images with the subject "Maps" |
| Maps and Charts |
|
|
|
|
| Sound Recordings |
|
|
|
|
| Moving Images |
Do maps work? Yes, with honest limits. Commons, DPLA and the Smithsonian have solid map cataloguing and return real maps (the example above came from this filter). The Internet Archive has no map type at all, so map there is a subject-and-collection match and is noisy. Because archives catalogue differently, treat kind as a strong hint rather than a guarantee, and run a query both with and without it when something seems missing. Multi-word queries need every word to match at all five archives, so fewer words find more.
Text versus images. kind selects whole records. To search inside documents, use ia_fulltext_search (the OCR text of Internet Archive books), and nara_search with include_extracted_text or nara_extracted_text for NARA scans.
The National Archives
NARA holds the federal paper trail: Official Records, Adjutant General and Quartermaster files, Brady's photographs, War Department maps, pension files. Beyond a plain nara_search:
The first page of results carries
facets: how many hits fall in each record group and type of material, so a broad query ("Fairfax") can be narrowed withrecord_group=77ortype_of_materials=mapinstead of paging.Each record names its record group and series, keeps photographer or mapmaker credits, keeps "ca." on estimated dates, and puts places apart from subjects.
nara_get_recordadds a ready-made archivalcitation, the use and access restrictions in NARA's own words, related links (Fold3, microfilm publication numbers) and the creating office.nara_extracted_textreturns each scan's machine OCR and any partner or volunteer transcriptions (FamilySearch's, on pension files), labelled, with AI-generated ones flagged. Use them to find names, then read the scan.Restrictions are reported honestly: "Unrestricted" is
free; "Restricted - Fully/Partly" isrestricted; "Restricted - Possibly" and "Undetermined" areunknown, because NARA is flagging a possible copyright or donor issue, not stating one.Date filters need both bounds at NARA's end; this server fills the open end, so
start_datealone works. NARA also matches records whose parent series or file spans the range, so results include undated records and some outside it: the response says so, and each record's owndateis the one to read.NARA's file sizes are sometimes placeholders (1234, 123456, 5242880 on files that were really 5 to 8 MB); those are left out rather than repeated.
Several search words must all match, so quote a phrase (
"Fairfax County"). Parentheses next to AND or OR often make NARA's firewall answer with its website; the error says so.
Chronicling America
The Library of Congress's digitised US newspapers (1690 to the present), searched through the public loc.gov JSON API, which needs no key. Pages come back with the matching words in bold, the paper, place and date, and links to the page's PDF, raw OCR text and word-coordinate XML.
chronam_searchtakes words (or an exact phrase),date_from/date_to(year, month or day) and astate.chronam_get_pagereturns one page's metadata and file links;chronam_page_textreads its OCR.Ids look like
sn83030214/1863-06-11/ed-1/2(LCCN, date, edition, page).OCR is raw and uncorrected, so names and numbers are often garbled. Try spelling variants and read the scan before quoting.
Pages up to 1930 are marked
freewithbasis: date-heuristic; later pages areunknown, because the Library states no rights per page.Limits. The Library allows 20 requests a minute, blocks clients for about an hour beyond that, and says a bot challenge can appear even below the limit. This server waits 3 s between requests. If it gets a challenge or a 429 it reports it, makes no further requests to loc.gov for 10 minutes, and
doctorshowsBLOCKEDinstead of failing. A block on loc.gov does not affect the other sources. If you see one, wait; the server will not try to get around it.Whole years: the local index. For a research question like "everything Virginia papers said in 1863", the live API (20 requests a minute) is the wrong tool. The Library publishes the OCR of every page as one archive per digitisation batch, a route it recommends for text analysis, and this server can build a searchable local index from them for the years you pick:
heritage-research-mcp chronam-index --from 1863 --to 1863 --state virginia --plan # what and how big, no download heritage-research-mcp chronam-index --from 1863 --to 1863 --state virginia # build itor ask your assistant to use
chronam_bulk_plan,chronam_bulk_startandchronam_bulk_status. Thenchronam_local_searchsearches every page in the index with no network request, andchronam_page_textreads a page from it. Things to know:It is a big download. Archives hold whole batches (many papers, many years), so every year in a batch is streamed and only your range is kept. Virginia for 1863 is about 43 GB over 40 archives and every US paper of 1863 about 460 GB, so the plan step shows the size first and a job over
--max-gb(default 60) is refused. Run it in a terminal for anything large; it can be interrupted and run again, and finished batches are skipped.Narrowing by month (
--from 1863-06 --to 1863-06) narrows what is kept, not what is downloaded, because an archive holds a paper's whole run. What actually shrinks the download:Name the papers with
--lccn(repeatable). For June 1863, the Richmond Daily Dispatch alone is 7 archives and 9.9 GB; the Dispatch, Enquirer and Whig together are about 22 GB; all ten Virginia papers about 50 GB.Archives that start after your range are skipped. Single-paper archives are sorted by date, so their start date (read with a small range request of about 128 KB, then cached) tells whether they can contain your range.
--planreports how many were skipped.Reading stops early once every paper you asked for in an archive has moved past your end date, so an archive is not read to the end. Such an archive cannot be checked against the Library's SHA-256 and is listed under
batches_read_early_without_checksum; pass--verify(orverify=true) to read everything and check every archive. This relies on archives being written in path order, which the one archive checked in full was (lccn/yyyy/mm/dd/..., zero-padded, so path order is date order); the reader turns early stop off if it sees dates go backwards before the stopping point.
Each archive read to the end is verified against the SHA-256 the Library publishes, and a mismatch discards that archive. Only the OCR text is kept (the ALTO XML, 97% of the bytes, is dropped as it streams past), so the index is a small fraction of the download.
Which batches is decided from the Library's list of batches and a title directory (the Library's own
chronam.csv) giving each paper's first and last issue. A paper added to Chronicling America after that directory was published is missed until it is refreshed; name it with--lccnto include it.Page ids are the same as the live API's, so
chronam_page_textandget_recordread from the index when a page is in it and only call loc.gov when it is not.
Verification of the local index. The archive reader was run on a real Library archive (
vi_barbaro_ver01, the Abingdon Virginian 1862 to 1864): the member layout, the published SHA-256 and the 1863 page selection all matched, giving 196 pages and 22 mentioning Mosby. Reading the Library's batch list and title directory over the network is covered by tests built from real excerpts, and was rate limited by the Library during development, so it has been exercised less than the reader.Verification of the live API. The search request and result handling follow the Library's own published notebooks (LibraryOfCongress/data-exploration) and a real result they record. This adapter has not been run successfully against the live service: from the machine it was written on, loc.gov answered every request with a challenge.
get_pageandpage_textrest on fewer recorded details thansearch(file locations are derived from the page's storage path). If a response differs from what is described here, please open an issue.
The Richmond Daily Dispatch
The University of Richmond publishes its transcription of the Daily Dispatch (1 November 1860 to 30 December 1865) as one TEI download, and its website has no public search service. So this server downloads that file once, indexes it into a local SQLite database, and searches that. After setup a search makes no request to the university.
heritage-research-mcp dispatch-index # or ask your assistant to run the dispatch_prepare toolThat fetches about 92 MB (with a provenance sidecar), takes about a minute, and leaves a 260 MB index in the cache folder: 1,384 issues, about 107,000 articles, notices and advertisements. Until it is built the Dispatch is left out of search and list_sources shows it as not ready.
dispatch_searchranks by relevance and shows the matching words in bold. Words are ANDed,"quoted phrases"stay together, and an uppercaseORjoins spellings (Mosby OR Moseby). Filter bydate_from,date_to(year, month or day) andsection(article, advert, orders, death, married, ...).dispatch_get_articlereturns the whole text, the page, rights, a link to the page on the university's site, and a link to the issue's TEI file (whichdownload_mediacan fetch).dispatch_issuelists a day's headlines by page, to browse an issue.Ids look like
D_025_053/1/15(issue file, page, article number) and match the numbering on dispatch.richmond.edu.
The paper itself is out of copyright, and records say so with basis: date-heuristic. The digital edition carries "© 2004 by the University of Richmond", which every record repeats along with a ready-made citation. The text is the university's keyed transcription from microfilm, so spellings and names vary; search variants. The project keyed all the news but only one day of advertisements every two weeks. Three issue files have a date field that disagrees with their printed title; each record's issue_title is the title as printed.
Not included
Virginia Chronicle (Library of Virginia). It runs on Veridian. Its OAI-PMH endpoints (
/cgi-bin/oaiand/oai) answer scripts with the same Cloudflare challenge, and itsrobots.txtdisallows every crawler except named search engines. Asking the Library of Virginia for its data is the remaining route.
If it opens a programmatic route, the adapter shape is the same as the others.
Tools
Tool | What it does |
| What is ready to use, and how many requests have been made |
| Query every ready archive at once, optionally by |
| One record, normalised, with its media files and rights |
| Cache one original locally with a provenance sidecar |
| Internet Archive: metadata search, search inside books, files, and reading or grepping OCR text |
| Commons files and categories |
| DPLA |
| National Archives Catalog |
| Smithsonian Open Access |
| Richmond Daily Dispatch (local index) |
| Chronicling America (live API) |
| Chronicling America (local index of bulk OCR) |
A source that needs a key is skipped, with a message saying where to get one, until the key is set. The Internet Archive and Commons work immediately.
Search results are compact (a short description, the first two files, no thumbnails) so that many can be read cheaply; get_record and the *_get_* and *_file_info tools return the full record.
Install
You need uv (or any Python 3.11+ environment).
Claude Code
claude mcp add heritage-research -- uvx --from git+https://github.com/magicianmarty/heritage-research-mcp heritage-research-mcpClaude Desktop and other MCP clients
{
"mcpServers": {
"heritage-research": {
"command": "uvx",
"args": ["--from", "git+https://github.com/magicianmarty/heritage-research-mcp", "heritage-research-mcp"]
}
}
}Pin a release for reproducible installs: git+https://github.com/magicianmarty/heritage-research-mcp@v0.2.0 (see the releases).
Hosted agent platforms. Any platform that runs stdio MCP servers can run it. Use the same uvx command, pass the keys as environment variables from the platform's secret store, and set HERITAGE_MCP_DISABLE_DOWNLOADS=1 (a hosted agent cannot read files the server downloads, so download_media is switched off and the text-returning tools do the work). To get files onto a machine you control, run the command-line downloader there (below) and let the agent call it through whatever shell access the platform gives it.
Download from a shell. heritage-research-mcp download <source> <record_id> [--kind image|pdf|text|audio|video|archive] [--media-index N] [--max-mb N] [--overwrite] runs the same checked downloader as download_media (https only, public addresses, size limit, provenance sidecar) and prints the result as JSON, including the file path. It exits 1 with {"error": ...} when it fails. HERITAGE_MCP_DISABLE_DOWNLOADS only removes the MCP tool; the command always works.
From source
git clone https://github.com/magicianmarty/heritage-research-mcp && cd heritage-research-mcp
uv venv && uv pip install -e ".[dev]"
.venv/bin/heritage-research-mcp doctorWorks with both the 1.x and 2.x generations of the MCP Python SDK.
Keys
Three sources need a free key: DPLA, NARA and Smithsonian. docs/KEYS.md walks through each, with what the key gives you and its limits. Put a key in an environment variable or on one line in ~/.config/heritage-research-mcp/keys/<dpla|nara|smithsonian> (mode 600). Files are the better choice: they stay out of client configs and shell history.
heritage-research-mcp doctor --live # shows what is configured, then makes one small real request to each sourceConfiguration
Variable | Default | Purpose |
| none | API keys (or key files, above) |
| none | Added to the User-Agent, so providers can reach you |
|
| Where |
|
| Refuse larger downloads |
| off | Remove |
|
| Stop before NARA's monthly quota |
|
|
|
Rights
Archives state rights in many vocabularies: Creative Commons URLs, rightsstatements.org URIs, free text, or nothing. Each record's rights reduces that to a reuse value:
| Meaning |
| Public domain, CC0, or a no-copyright statement |
| Free if credited ( |
| Reuse must carry the same licence |
| Conditions that limit reuse |
| The holder stated nothing. This is not permission. |
rights.basis says where the value came from: holder, or date-heuristic for one narrow case (a US text first published 96 or more years ago). This is a filtering aid, not legal advice. download_media adds a rights_warning when an item is not marked free.
Being a good citizen
Requests are paced per source and retried with backoff;
Retry-Afteris honoured. Every request carries a User-Agent naming this project.NARA allows 10,000 queries per key per month and forbids scraping or bulk download through the API. This server counts requests, stops at the limit, returns one page of at most 100 records per call, and includes the attribution NARA requires.
The Internet Archive full-text endpoint is experimental and may change.
Chronicling America (loc.gov) is paced to its published 20 requests a minute and left alone for 10 minutes after any block or bot challenge.
The Dispatch costs the university one 92 MB download, made only when you ask for it (
dispatch_prepare) and cached withrefreshto repeat it.
Security
Keys are never returned in results and are redacted from errors and logs (HTTP request logging is filtered, because two providers take the key as a query parameter).
download_mediafetches onlyhttpsURLs that resolve to public addresses, refuses credentials and non-standard ports, re-checks every redirect, and enforces a size limit. The URL always comes from a record a source returned, never directly from the model.See SECURITY.md to report a problem.
Status
Verified against the live services on 6 and 7 October 2026: the Internet Archive, Wikimedia Commons, DPLA, Smithsonian Open Access and the National Archives Catalog. The Dispatch was verified on 17 October 2026 against the live archive (all 1,384 issues indexed, and article numbering checked against the site's own issue data for five issues sampled across 1860 to 1865); its test fixture is a trimmed real issue. The NARA adapter was first written from the published specification and then rebuilt against real responses, which showed that its creators, dates, restrictions and file sizes all needed different handling; its test fixtures are trimmed live captures. If you find a response shape that differs, please open an issue. The Live smoke workflow re-checks every configured source weekly.
Development
uv venv && uv pip install -e ".[dev]"
.venv/bin/pytest # offline: every request is mocked
.venv/bin/ruff check src tests && .venv/bin/pyright srcCI runs the suite on Python 3.11 and 3.13 against both SDK generations. See docs/DESIGN.md for how it is put together.
Licence
MIT. Records and files returned by this server remain subject to their holders' terms.
Available Tools
23 toolscommons_category_membersA
List the files in a Commons category (for example "Mosby's Rangers").
Args: category: The category name, with or without the "Category:" prefix. limit: Results to return (1 to 50). filetype: Optionally keep only bitmap, drawing, audio or video files.
Inputs: category (required), limit, filetype. Returns: object
commons_file_infoA
Get one Commons file's licence, author, description, categories and original size.
Args: title: The file name, with or without the "File:" prefix.
Inputs: title (required). Returns: object
commons_searchB
Search Wikimedia Commons files. Each record includes the licence read from the file's metadata.
Args: query: Search words. limit: Results to return (1 to 50). filetype: bitmap, drawing, audio, video, office or multimedia. kind: text, image, map, audio or video. Maps are found by the word "map" in the file title.
Inputs: query (required), kind, limit, filetype. Returns: object
download_mediaA
Download one original file from a record into the local cache and write a provenance sidecar.
The sidecar (<file>.provenance.json) records the source, URL, retrieval time, SHA-256 and the rights
the holder stated, so the file can be traced later. Only https URLs on public addresses are fetched.
If rights.reuse is not free the result carries a rights_warning: treat the file as reference only.
Args:
source: internet_archive, commons, dpla, nara or smithsonian.
record_id: The record's id, as returned by search.
media_index: Which entry of the record's media list to fetch (default the first).
media_kind: Pick the first media of this kind instead (pdf, text, image, audio, video, archive).
overwrite: Fetch again even if the file is already cached.
max_mb: Refuse files larger than this (default 250, or HERITAGE_MCP_MAX_DOWNLOAD_MB).
Inputs: source (required), record_id (required), max_mb, overwrite, +2 more. Returns: object
dpla_get_itemB
Get one DPLA item by its id. Needs DPLA_API_KEY.
Inputs: item_id (required). Returns: object
dpla_searchA
Search the Digital Public Library of America for items held by US libraries, archives and museums.
Needs DPLA_API_KEY. Give a search term, any filter, or both. DPLA indexes descriptions, not the item
itself: follow landing_url to the holding institution, and read rights before reusing anything.
Args: q: Free-text search across the record. title: Words in the title. creator: Creator or author. subject: Subject heading. place_state: US state name, e.g. Virginia. date_after: Earliest date (YYYY or YYYY-MM-DD). date_before: Latest date. type: image, text, sound, moving image, physical object, etc. provider: DPLA hub name. data_provider: Contributing institution name. page: Page number from 1. page_size: Results per page (1 to 100). sort_by: A DPLA field such as sourceResource.date.begin. kind: text, image, map, audio or video. Maps are images with the subject "Maps".
Inputs: q, kind, page, type, +10 more. Returns: object
get_recordA
Fetch one record in the normalised shape, including its media files and rights.
Args: source: internet_archive, commons, dpla, nara or smithsonian. record_id: The id from a search result (an Internet Archive identifier, a Commons "File:" title, a DPLA id, a NARA naId, or a Smithsonian id).
Inputs: source (required), record_id (required). Returns: object
ia_fulltext_searchA
Search the OCR text inside Internet Archive books and reports, with snippets and page numbers.
This is the way to find a name, place or phrase inside a memoir, regimental history or official report. Matches are wrapped in ** in the snippets. It uses an experimental endpoint that may change.
Args: query: Words or a "quoted phrase". hits: Results to return (1 to 50). page: Page number from 1.
Inputs: query (required), hits, page. Returns: object
ia_get_itemB
Get one item's metadata, rights and downloadable files (PDF, OCR text, page images).
Inputs: identifier (required). Returns: object
ia_grep_textA
Find every occurrence of a word or phrase in an item's OCR text, with surrounding context and offsets.
The text is fetched once and held in memory, so repeated searches of the same book are fast.
Args:
identifier: The Internet Archive identifier.
pattern: Text to find. A plain phrase matches across the line breaks, repeated spaces and split
hyphens that OCR leaves in scanned books ("twenty-five prisoners" finds "twenty-five prisoners" and
"twenty- five prisoners"). A regular expression only if regex is true.
regex: Treat the pattern as a regular expression.
ignore_case: Ignore capitalisation.
context: Characters of context on each side (0 to 1000).
max_matches: Most matches to return (1 to 100); the total is always reported.
Inputs: pattern (required), identifier (required), regex, context, +2 more. Returns: object
ia_read_textA
Read a slice of an item's OCR text by character offset (up to 20,000 characters at a time).
Use ia_grep_text first to find the offset you want.
Inputs: identifier (required), start, length. Returns: object
ia_searchA
Search Internet Archive item metadata (title, author, subject, description). Not the text inside books.
Use ia_fulltext_search to find a phrase inside scanned books. The query accepts Lucene syntax, for
example creator:(mosby) AND subject:"Virginia".
Args: query: The search. mediatype: texts, image, audio, movies, software, data, etc. year_from: Earliest publication year. year_to: Latest publication year. collection: Restrict to a collection identifier, e.g. americana. rows: Results per page (1 to 100). page: Page number from 1. sort: For example "downloads desc" or "date asc". kind: text, image, map, audio or video, translated into Internet Archive terms (maps are matched on the "maps" subject and the map collections, so results are noisy).
Inputs: query (required), kind, page, rows, +5 more. Returns: object
list_sourcesA
List the archives this server can search, whether each is ready to use, and what it is good for.
Sources that need a free API key show configured: false until a key is set; the Internet Archive and
Wikimedia Commons need none.
Inputs: no named parameters. Returns: object
nara_childrenB
List the immediate children of a series or file unit (one level down). Needs NARA_API_KEY.
Inputs: parent_na_id (required), page, limit. Returns: object
nara_extracted_textB
Get OCR text extracted from a record's digital objects. Needs NARA_API_KEY.
The response is passed through from NARA with long strings shortened to 4,000 characters.
Inputs: na_id (required), page, limit, object_id. Returns: object
nara_get_recordA
Get one Catalog record by its numeric naId, with digital objects and ancestry. Needs NARA_API_KEY.
Inputs: na_id (required). Returns: object
nara_searchA
Search the US National Archives Catalog. Needs NARA_API_KEY (10,000 queries a month by default).
One page per call, at most 100 records: the API's terms forbid scraping or bulk download. The remaining monthly budget is visible in usage_report. Results include the attribution NARA requires.
Args: q: Search words; supports AND, OR, NOT, wildcards (*) and "exact phrases". title: Words in the title. start_date: Earliest date (YYYY, YYYY-MM or YYYY-MM-DD). end_date: Latest date. available_online: Only records with digitised objects. type_of_materials: e.g. Photographs and other Graphic Materials, Textual Records, Maps. level: series, fileUnit, item, recordGroup, etc. record_group: Record group number, e.g. 109 (Confederate records). ancestor_na_id: Only records beneath this naId. geographic: Geographic subject heading. creators: Creator heading. include_extracted_text: Include OCR text in the results where NARA has it. limit: Results per page (1 to 100). page: Page number from 1. kind: text, image, map, audio or video, mapped to NARA's type of materials (unverified until a key has been used against the live service).
Inputs: q, kind, page, level, +11 more. Returns: object
searchA
Search several archives at once and return normalised records grouped by source.
Every record carries rights.reuse (free, attribution, share_alike, non_commercial, no_derivatives,
restricted or unknown). This is a filtering aid, not legal advice: unknown means the holder stated
nothing.
Args:
query: What to look for. Plain words work everywhere.
sources: Limit to some of internet_archive, commons, dpla, nara, smithsonian. Default: every source
that is ready to use (see list_sources).
limit: Records per source (1 to 25).
date_from: Earliest year. Applied by the Internet Archive, DPLA and NARA; ignored by the others.
date_to: Latest year, applied the same way.
fulltext: Also search the text inside Internet Archive books (an experimental endpoint). Returned
under fulltext; this is the way to find a name or place inside a memoir or official report.
kind: Only this kind of material: text (books, reports, manuscripts), image (photographs, prints),
map, audio or video. Each archive is asked in its own vocabulary, and kind_applied in each
result says how. Maps are the least uniform: Commons matches on the word "map" in the title and
the Internet Archive on its map subjects and collections, so expect some misses and some noise.
Every record also carries a best-effort kind.
Inputs: query (required), kind, limit, date_to, +3 more. Returns: object
si_get_contentA
Get one Smithsonian record by id (for example edanmdm-nmaahc_2012.36.4ab). Needs SMITHSONIAN_API_KEY.
Inputs: content_id (required). Returns: object
si_searchA
Search Smithsonian Open Access records. Needs SMITHSONIAN_API_KEY.
Records whose media are all marked CC0 come back with rights.reuse "free"; otherwise "Usage conditions apply" is surfaced as restricted.
Args: q: Search words; supports AND/OR and fielded terms such as topic:Cavalry (see si_terms). rows: Results to return (1 to 100). start: Offset of the first row. sort: relevancy (default), newest or updated. type: EDAN record type, e.g. edanmdm. row_group: objects or archives. category: Search within art_design, history_culture or science_technology instead. kind: text, image, map, audio or video, added to the query as object_type or online_media_type.
Inputs: q (required), kind, rows, sort, +4 more. Returns: object
si_statsA
Counts of CC0 objects and media in the collection. Needs SMITHSONIAN_API_KEY.
Inputs: no named parameters. Returns: object
si_termsB
List the vocabulary for a search field: culture, data_source, date, online_media_type, place, topic or unit_code. Useful for building fielded queries. Needs SMITHSONIAN_API_KEY.
Inputs: category (required), limit, starts_with. Returns: object
usage_reportA
Requests made this month and this session per source, and the provider rate limit last seen.
Inputs: no named parameters. Returns: object
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 23 tools
The general `search` and `get_record` tools overlap with the five per-source search and get tools (si_search, nara_search, ia_search, commons_search, dpla_search, si_get_content, nara_get_record, ia_get_item, commons_file_info, dpla_get_item). Descriptions help by distinguishing normalized output from source-specific details, but an agent could still misselect for simple queries.
Most tools follow a predictable snake_case pattern with source prefixes (nara_, si_, ia_, commons_, dpla_) for source-specific operations and unprefixed names for general operations. Minor deviations are noun-only names like `si_terms`, `si_stats`, and `usage_report`, but overall the scheme is clear and consistent.
23 tools is on the heavy side for a single server, even one covering five archives. While many tools serve distinct purposes (e.g., `ia_grep_text`, `nara_children`, `si_terms`), the overlapping general and source-specific search/get tools add redundancy.
The set covers discovery, retrieval, fulltext search, media download, rights, provenance, and usage reporting across five archives. Minor gaps remain, such as cross-source fulltext search (only Internet Archive supports it) and OCR retrieval for some sources, but core research workflows are well supported.
Maintenance
Related MCP Connectors
SEC filings, Wikipedia, arXiv, US Census and more for AI agents, with every answer source-linked.
License-checked research archive and technical-spec search for AI agents. Pay per call via x402.
Sourced public knowledge for AI agents: search and read records with evidence; moderated writes.
Academic papers, reference verification and research data for AI agents.
Related MCP Servers
- AlicenseAqualityBmaintenanceEnables users to search and access digital collections from the Swedish National Archives (Riksarkivet) through multiple APIs. Supports searching records by keywords, exploring collections, and downloading historical images and documents.227Apache 2.0
- FlicenseNot gradedqualityNot gradedmaintenanceEnables semantic search across the Arke Institute's extensive archive of NARA records and presidential libraries using natural language queries. Provides access to millions of historical documents, photographs, and records with OCR'd content and complete metadata.-
- AlicenseNot gradedqualityDmaintenanceEnables natural language search and exploration of the Dutch WWII Oorlogsbronnen archives, allowing users to query historical documents, photographs, and personal accounts through AI assistants.15MIT
- FlicenseNot gradedqualityCmaintenanceEnables querying genealogical records, archive statistics, historical weather, and full-text transcriptions from Open Archives via natural language.4-