Ask the Archives
Open, machine-transcribed data archives,
searchable by any MCP client.
Public-domain collections · released CC0
What this is
The great public-domain document collections, a century of government research reports, archival records, and technical series, survive mostly as scans: photographs of pages, invisible to search and to the language models researchers now work with.
Ask the Archives transcribes these collections page by page with vision models, tags every page with its source engine and measured quality, and serves them through one public MCP endpoint: hybrid keyword-and-semantic search over section-aware chunks, full page text with provenance, figure records, page images, and ready-made citations. Every result carries its collection, so one connection searches everything.
The collections
| Collection | Span | Size | Status |
|---|---|---|---|
| NACA Technical Reports (NASA NTRS) | 1915–1958 | N_REPORTS reports · N_PAGES pages | Live |
| U.S. National Archives Catalog (NARA) | In preparation |
Final counts land with each collection's first data release; the endpoint reports live counts when you connect.
How to connect
Claude Code or Claude Desktop:
claude mcp add --transport http archives https://mcp.askthearchives.org/mcp
Any other MCP client:
{
"mcpServers": {
"archives": {
"type": "http",
"url": "https://mcp.askthearchives.org/mcp"
}
}
}
No account, no key. The endpoint is rate-limited per connection and reports its own live corpus counts in the server instructions.
Tools
- search
- Hybrid, keyword or semantic search over the corpus; hits carry report, page span, section heading, snippet, score, source engine and quality tag.
- get_page
- Full text of one page with provenance (engine, quality, config hash) and a ready-made citation.
- get_report
- Report metadata, NTRS link, and page availability by quality.
- get_figures
- Structured figure records (charts, drawings, photographs) for a report or a single page;
verified: falsemeans machine-drafted. - get_page_image
- The scanned page itself (PNG), optionally cropped to a bounding box.
- cite
- Formatted citation with stable identifiers and the NTRS URL.
Measured accuracy
Transcription quality is measured, not asserted: every collection ships with its own benchmarks against held-out human ground truth. For the NACA collection, character error rate (lower is better):
| Engine | Median CER |
|---|---|
| Campaign engine (8B vision model, fine-tuned) | 0.0940 |
| 27B escalation lane | 0.0975 |
Retrieval is benchmarked too: hybrid search finds the right document in its top five results for 76% of held-out research questions, against 54% for whole-page embedding search. The full engine comparison lives in BENCHMARKS.md.
Caveat on tables. Agreement between engines can miss wrong digits inside numeric tables; flagged pages go through an escalation and verification lane. Treat table digits as less certain than prose, and check the page image for anything load-bearing.
Citation policy
Use the cite tool rather than composing citations by hand. Spans a
transcriber could not read are marked <?> in the text; they are
illegible in the original scan and must not be guessed or filled in. The page images
served by this archive are the citable original.
Bulk download
Each collection's data releases ship as torrents: page images, transcriptions in every quality layer, figure records, errata, and the search index, under CC0. TORRENT_LINKS_PLACEHOLDER