german-archives-mcp

An MCP server for German archives' finding aids and Baden's civil registers: search what German archives, libraries and museums have described in the Deutsche Digitale Bibliothek and its archive portal Archivportal-D, read one entry with its place in the archive's fonds, find the Standesbuch of a Baden parish for a year at the Landesarchiv Baden-Württemberg, and download a page of it to read.

The DDB is Germany's counterpart of DPLA, and for archives its counterpart of ArchiveGrid as well: archives deliver their finding aids to it, item by item, with signatures (call numbers), dates and the fonds each item belongs to. Some deliver images. Its version 2 API answers without a key.

The Landesarchiv has put the Baden Standesbücher online in full: the civil duplicates of the parish registers of baptisms, marriages and burials that Baden's clergy kept from 1810 to 1870, one book per parish and confession for a span of years. South Baden's are fonds L 10 at the Staatsarchiv Freiburg; north Baden's are fonds 390 at the Generallandesarchiv Karlsruhe. The server covers both.

It works the way a careful genealogist does. A finding-aid entry says where a record is, not what it says: cite the archive and its signature, not the DDB. The page image is the evidence: read it before citing it, and cite the archive, the signature and the image number with the archive's permalink. A zero is not a negative: most German registers are not online, and many archives deliver nothing to the DDB.

Nothing here writes anywhere, and nothing here keeps a family tree. It sits beside dpla-catalog-mcp for American collections and snac-archives-mcp for finding which archive holds a family's papers.

This is an independent project. It is not affiliated with, endorsed by, or supported by the Deutsche Digitale Bibliothek, the Landesarchiv Baden-Württemberg, or any archive whose finding aids it reads.

Tools

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

The Deutsche Digitale Bibliothek and Archivportal-D

The Baden Standesbücher at the Landesarchiv Baden-Württemberg

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 german-archives-mcp

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

git clone https://github.com/ianderso/german-archives-mcp
cd german-archives-mcp
uv sync
uv run german-archives-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": {
    "german-archives": {
      "command": "uvx",
      "args": ["german-archives-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 german-archives -- uvx german-archives-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.

How to read what comes back

• A DDB entry is a finding aid. The archive wrote it to say where a record is and roughly what it holds. It supports a research task (order the file, open the images), not a fact. Cite the archive and its signature (citation in ddb_item), and the archive's own page where the entry links one; keep the DDB id only as a finder.
• A DDB search covers descriptions, never page text, and only what archives have delivered. Many German archives deliver nothing, and most that do describe registers by volume, not by name. No hits is not a negative.
• archive matches the start of the holder's name, exactly as the DDB writes it, capitals and umlauts included: Stadtarchiv Düsseldorf, not stadtarchiv duesseldorf.
• Dates match by overlap. year_to: 1812 keeps a register for 1811-1816.
• A series heading has no record of its own. ddb_item then answers with its place in the hierarchy and its children (grouping_node).
• The DDB lists only the first ten images of a Landesarchiv item. labw_image reads every page; ddb_search and ddb_item give the permalink it takes (labw_permalink).
• A Standesbuch is a civil copy of the parish register, 1810 to 1870. Before 1810, look for the parish register itself; after 1870, the civil registry office. A book often covers several years, and sometimes two confessions (confessions).
• labw_standesbuch finds books whose title names the place. A village incorporated into a town is filed under both names (Niederrimsingen, Breisach am Rhein FR), so a search for the town lists its villages too; place_named_first says which books are the place's own. Spell the place as the archive does, umlauts included.
• page is the archive's image number ("Bild"), counted from 1, not the number written on the page. A book's image numbers and its permalinks can fall out of step where an image was added later: in L 10 Nr. 510, Bild 119 is permalink …-380, and Bild 378 is …-377. labw_image reads the archive's own list, so either works.
• Read the page. Standesbücher are handwritten in German script. The image is the evidence; nothing in this server transcribes it.

Terms of use, and being a good guest

Deutsche Digitale Bibliothek. The DDB's developer documentation says its data is "publically accessible without any restrictions" (DDB-Backend / API) and that "using version 2 all endpoints are usable without an API key" (Differences between API versions 1 and 2). Its OpenAPI description says metadata delivered through the API is licensed CC0, while images carry each institution's own licence. Two older pages still describe a key: the DDBpro page on its interfaces and the general text of the OpenAPI description; the version 2 routes this server uses declare no security and answered without one on 2026-10-11. The former API terms page (/content/terms/api) now answers 404, and the DDB's portal sits behind an Anubis proof-of-work check, which this server never touches: it reaches only the API host. No rate limit is published. The API host has no robots.txt (it answers 404). If the DDB starts refusing keyless requests, the tools say key_required.

Landesarchiv Baden-Württemberg. Its terms (Nutzungsbedingungen auf einen Blick) put the online finding aids' metadata under CC0, mark each digitised item with its rights (the Standesbücher viewer shows the Public Domain Mark 1.0), and ask every user to "Archivsignatur oder den Permalink zitieren". They also ask for a deposit copy of a book made with substantial use of the archive's material (§ 8 Abs. 9 Landesarchivgesetz). They say nothing about automated access. As a fact: robots.txt on www2.landesarchiv-bw.de allows Googlebot and Yahoo and disallows everyone else. The DDB's copy of the Landesarchiv's entries lists the images as CC BY 3.0 DE; the archive's own viewer marks them Public Domain, and labw_image reports the archive's mark.

What the client does. It sends one request at a time to each host: the DDB's at least a second apart, the Landesarchiv's at least two. Two identical calls in flight share one request. Searches are cached for a day, DDB records for a week, and the Landesarchiv's permalink answers and image lists for 30 days, so reading a second page of a book costs one request. 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. Images are rendered at the size the archive's viewer shows at 100 per cent, and written to your file, never cached.

Deliberately not here

• The DDB's portal, newspapers and user features. The portal is behind a bot check; favourites and saved searches need an account.
• Images on the holders' own sites. ddb_item returns their links; this server fetches nothing outside its two hosts.
• Transcription. The page is read by a person, or by a tool built for German script.
• Working around bot checks. A site that answers with one 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: finding-aid entries, image links, web pages. The server assumes that text can steer the model, and limits what a steered model can make it do.

• Which hosts. Two, fixed: api.deutsche-digitale-bibliothek.de and www2.landesarchiv-bw.de. Any other host is refused before it is looked up, including addresses that come back in a response, such as an image link in a DDB record or a redirect. A permalink redirect is read, never followed.
• 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 refused, and the connection goes to the address that was checked. Proxy settings in the environment are not used.
• How much. A JSON or HTML answer over 10 MB, or an image over 60 MB, is refused as it streams in. A search returns at most 50 hits a page.
• What is sent. Search words have Solr's field, range and parameter syntax escaped; a place must be letters, spaces and a few punctuation marks; ids and signatures are matched against narrow patterns; an image is asked for only by a file name the archive's own list gave, and only if it has the archive's file-name shape.
• Which files. labw_image creates one new file and never overwrites one. The bytes must be an image, judged by their first bytes, and a JPEG, since the file must end in .jpg or .jpeg. Never a hidden file or folder, never under ~/Library, and with GERMAN_ARCHIVES_DOWNLOAD_DIR set, never outside it, all judged after links are resolved. A refused download leaves nothing on disk.
• Site text is untrusted. Titles and descriptions reach the model verbatim. The server's instructions tell the model to treat that text as material to weigh, never as instructions; the model still decides, so review what it proposes to do.

To report a vulnerability, see SECURITY.md.

Development

uv sync --extra dev
uv run pytest                      # mocked with respx; never touches a site
uv run ruff check .
uv run ruff format --check .
uv run python -m tests.live_check  # paced calls to the live sites

The live check asks the sites what the recorded fixtures cannot: whether their answers still have the shape the server reads. See CONTRIBUTING.md for how the suite is organised, docs/API-NOTES.md for what was observed of each site and when, and docs/DESIGN.md for why the server is shaped this way.

Credits

The finding aids belong to the archives that wrote them, and the images to the archives that hold the records. The Deutsche Digitale Bibliothek is run by a network of German cultural institutions; the Landesarchiv Baden-Württemberg is the state archive of Baden-Württemberg.

License

MIT.