NARA Catalog

nara-catalog-mcp

An MCP server for genealogical research in the US National Archives Catalog. Find an ancestor's pension file, service record or census page, read what machines and volunteers have transcribed from it, and fetch the page images you will cite. It also searches NARA's Access to Archival Databases (AAD), which the Catalog cannot reach: the NUMIDENT's Social Security applications and deaths, the WWII Army enlistment cards, and the 1820–1912 ship passenger lists among them.

It works the way a careful genealogist does. A catalogue description is a finding aid, and OCR text, transcriptions and tags are somebody else's reading of the page: all of them are leads. The page image is the source. The tools say which is which, and the server tells the model to read the image before citing it.

Nothing here writes to the Catalog, and nothing here keeps a family tree. The server finds and reads records; what you conclude from them belongs in your genealogy software, or in a family-tree MCP server running alongside this one.

The tools work on any record the Catalog describes, so a historian or a journalist can use them too. The examples throughout are about tracing people.

This is an independent project. It is not affiliated with, endorsed by, or supported by the National Archives and Records Administration.

Tools

The server publishes 21 tools. Twenty only read, and are annotated read-only for the client. One, download_page_image, writes a new file to local disk; it never overwrites one.

Finding a record

Reading a record

What machines and other people read in it

Everything in this group is somebody else's reading of the document. It tells you which page to open. It does not tell you what the page says.

Searching the documents rather than the catalogue

NARA's Access to Archival Databases (AAD)

AAD serves a selection of NARA's electronic records as searchable tables. Each row is a database entry an agency clerk typed from a record: a transcription or index entry, not the record, and cited as an electronic record. These tools need no key and spend none of the Catalog allowance.

Housekeeping

Setup

You need Python 3.11 or later, uv, and a Catalog API key. Request a key from the Catalog API team via the API documentation; the default allowance is 10,000 calls per month.

There are two ways to run the server.

Without cloning. uvx fetches it from PyPI and runs it in one step, and caches the result:

NARA_API_KEY=your-key uvx nara-catalog-mcp

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

git clone https://github.com/ianderso/nara-catalog-mcp

cd nara-catalog-mcp uv sync cp .env.example .env # then put your key in it uv run nara-catalog-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

Without cloning:

{

"mcpServers": { "nara": { "command": "uvx", "args": ["nara-catalog-mcp"], "env": { "NARA_API_KEY": "your-key" } } } }

From a clone (the env block can be dropped if the key is in the clone's .env):

{

"mcpServers": { "nara": { "command": "uv", "args": ["--directory", "/path/to/nara-catalog-mcp", "run", "nara-catalog-mcp"], "env": { "NARA_API_KEY": "your-key" } } } }

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

Claude Code

claude mcp add nara --env NARA_API_KEY=your-key -- uvx nara-catalog-mcp

or, from a clone whose .env holds the key:

claude mcp add nara -- uv --directory /path/to/nara-catalog-mcp run nara-catalog-mcp

Configuration

All settings come from the environment. A .env file in the directory the server starts in supplies any that the environment does not; with uv --directory that is the clone. Only that directory is read — not its parents, and not the directory the package is installed in.

A missing key, or an unusable value, is reported on the first tool call as a not_configured result naming the variable.

The call budget

A Catalog key is capped per month, and a sweep across many names will exhaust it faster than expected. Every successful response is cached on disk, keyed by endpoint and query, so repeating a call costs nothing. Live calls are also written to a ledger, budget.json beside the cache, so api_budget reports the month's spend across sessions as well as this session's live calls and cache hits. A rejected call counts as live, since it reached the API. The ledger sees one machine and one cache directory; the Catalog's own count is the authority.

The cache never expires. Pass refresh=true to get_record, get_record_images, get_extracted_text, get_transcriptions, get_tags or get_comments to re-read one answer from the Catalog: that spends one call and replaces the cached copy. New transcriptions and tags arrive over time, and a file that was paper-only can be digitised, so a cached "nothing here" is the answer most worth refreshing. Deleting NARA_CACHE_DIR still works, and also resets the month's ledger.

Searching past the first 10,000 hits

page stops working beyond 10,000 results, which a common surname passes. search_records_advanced reports a paging_note when you hit that boundary; either narrow the search or re-run it with search_after="*" and follow the next_search_after value from each response.

How a search behaves

title matches words in the record title and is the precise option — case files are usually titled with the person's name. query searches the full description: broader, noisier, and worth reaching for only when a title search finds nothing.

Results come back as summaries rather than whole records. A raw Catalog record is large and mostly irrelevant to the decision you are making, which is whether this record is worth opening.

Searching AAD

AAD has no API, so aad_* read its web pages the way a browser does, and parse them. That shapes how they behave:

• One file at a time. A big series is split into files, and a search runs on one. The NUMIDENT is split by entry type (application, claim, death) and by surname range, so aad_list_series comes first.
• Fields by name. fields={"LAST NAME": "PRESLEY", "DATE OF BIRTH (YEAR)": "1930-1935"}. A field AAD stores as a code, such as a state or a country, cannot be searched by name: put its meaning in query, which matches the whole row, code meanings included.
• Slow on purpose. One request at a time, at least two seconds apart, and a search usually takes two requests because AAD answers first with a "Please Wait" page. Answers are cached for seven days, under NARA_CACHE_DIR/aad; refresh=true on aad_get_record fetches a row again.
• Record ids are not stable. NARA reloads files, and a reload can renumber the rows. Search again rather than reuse an old id, and cite a row by its fields, its file and its series.
• A row is not the record. It was keyed from an SS-5, a punch card or a manifest, with the keying errors that brings, and no file holds everyone. The NUMIDENT has entries only for people with a verified death or born before 1908, and its death file omits state-reported deaths. Each answer carries AAD's own notice for the file, and a NUMIDENT answer says what that file leaves out.

Notes

• A description is not evidence. It summarises a file; it says nothing about what any individual page contains. Read the images before citing.