CONTENTdm State Archives

contentdm-mcp

An MCP server for the record images that US state archives and state libraries publish on CONTENTdm: death certificates and their annual indexes, Confederate pension files, voter registers, prison registers, county court and estate papers, letters and diaries, with whatever transcript or OCR text the institution added.

CONTENTdm is OCLC's hosted digital-collections platform. The Alabama Department of Archives and History, the Georgia Archives' Virtual Vault, the Tennessee State Library and Archives, Ohio Memory, the Illinois Digital Archives, the Indiana State Library, Missouri Digital Heritage and many more run on it, and every one of them answers the same keyless JSON API and IIIF image service. The server ships a curated list of those sites, searches one or several of them at once, reads an item and its pages, and downloads a page image to read.

It works the way a careful genealogist does. The page image is the evidence. A transcript or OCR text is a lead that says which page to read, and an index card is a finding aid to the record behind it. Every result carries the holding institution's item address and a citation core (institution, collection, title, identifier, address), because the item page is what gets cited, not this server.

Nothing here writes anywhere, and nothing here keeps a family tree. It sits well beside dpla-catalog-mcp, which finds an item across hundreds of collections and hands you the institution's address (get_item takes that address directly), and nara-catalog-mcp for federal records.

This is an independent project. It is not affiliated with, endorsed by, or supported by OCLC or any of the institutions whose sites it reads.

Tools

The server publishes eight tools. All but get_image are read-only; get_image writes one new file and never overwrites one.

Finding

Reading

The curated sites

list_instances gives the whole list. On 2026-10-06, 18 sites answered, and 7 entries record institutions that have left CONTENTdm, so a search can say where they went:

Any other CONTENTdm site works too, at its cdmNNNNN.contentdm.oclc.org address: pass https://cdmNNNNN.contentdm.oclc.org as instance. Every classic site answers there, whatever its own domain, and the source of its pages names the number ("cdmServerUrl": "serverNNNNN.contentdm.oclc.org"). There are more than 675 of them, and no registry. To use a site by its own domain, add it to CONTENTDM_EXTRA_INSTANCES; a tool argument cannot, for the reason under Security.

Setup

You need Python 3.11 or later and uv. There is no key to request.

Without cloning. uvx fetches it from PyPI and runs it in one step:

uvx contentdm-mcp

From a clone, which is what you want if you will change it:

git clone https://github.com/ianderso/contentdm-mcp

cd contentdm-mcp uv sync uv run contentdm-mcp # stdio server, usually launched by the client

Either way the server speaks MCP over stdio, so you will normally let an MCP client start it rather than run it by hand.

Claude Desktop

{

"mcpServers": { "contentdm": { "command": "uvx", "args": ["contentdm-mcp"] } } }

A desktop app does not always inherit your shell's PATH. If the server fails to start because uvx cannot be found, give the full path that which uvx prints as the command.

Claude Code

claude mcp add contentdm -- uvx contentdm-mcp

Configuration

Nothing is required. A .env file in the directory the server starts in supplies anything the environment does not; only that directory is read.

An unusable value is reported on the first tool call as a not_configured result naming the variable.

Being a good guest

Each site is one institution's server. The client sends one request at a time to each site, at least a second apart; a search across several sites runs them side by side, six at a time, so no site sees more than one caller. Two identical calls in flight share one request. Collection lists and field definitions are cached for 30 days, items for 7, searches for a day. A 429, a 5xx or a dropped connection gets one retry, honouring Retry-After. The User-Agent names the package, its version and this repository.

robots.txt on these sites disallows the website's search pages, not the API this server uses. Each institution's own terms govern what you do with its images: list_instances and get_item pass on what the sites say, and several ask for permission before publication.

How to read what comes back

• A search is only as wide as `answered`. failed, timed_out and not_searched list the sites a negative does not cover. A site that redirects, asks for a bot check or lists no collections has usually moved.
• Search covers metadata and text fields, never the image. Many record images have no text at all, so a name can be on a page that no search finds. Browse by collection, or by the index volume, instead.
• A compound object's text is on its pages. Volumes, files and newspaper issues are compound objects; their own transcript field is usually empty. Use page_hits=true, or get_pages with a query, to find the page.
• Field names are per collection. "Name of Father" is namea in one collection and father in another. get_collection says which nick holds what, and which field is the transcript: that is marked by type, because the name differs from site to site.
• "Transcript" can be machine OCR. The Tennessee vital-records indexes carry OCR under that label. Read the image.
• `exact` is a phrase search (the words together, in order), not whole-field equality. all needs every word, any one of them.
• A PDF item can hold many pages. Its IIIF image shows the first; get_image with pdf_page reaches the others.
• `page` is a position in the object, counted from 1, not the number printed on the scan: the entry for Sgt. Alvin C. York is on page 461 of volume 8 of Tennessee's 1960-1964 death index, a scan named 54784_460 and stamped 457. Cite the printed number too when it differs.
• Cite the item page. Use the citation each result carries, and the institution's own cite_as line where get_item finds one (Georgia, Oklahoma and Alaska items have them). Record the identifier (alias:pointer) so the item can be found again.

Deliberately not here

• Writing to any site. CONTENTdm's write functions need a staff login.
• Sites on other platforms. The new CONTENTdm (launched 2026-09-22) has no published public API, and Quartex's is undocumented. The tools speak to an adapter layer, so either can be added without changing them; see docs/DESIGN.md.
• Working around bot checks. A site that answers with a challenge is reported as blocked and left alone.

Security

Tool arguments are written by a model, and the model reads text this server does not control: titles, transcripts, DPLA records, web pages. The server assumes that text can steer the model, and limits what a steered model can make it do.

• Which hosts. A tool reaches the curated sites, any address under contentdm.oclc.org (OCLC's own servers), and the sites you list in CONTENTDM_EXTRA_INSTANCES. Any other host is refused before it is even looked up, because a DNS query for secret.attacker.example would already deliver the name; the refusal says how to reach the site instead.
• Which addresses. Each connection is checked where it is made: a name that leads to a private, loopback, link-local, CGNAT, multicast, reserved or unspecified address, IPv4 or IPv6, is