A librarian for the American record · served over MCP
A curated, provenance-tracked knowledge base over primary sources — testimony, treaties, hearings and declassified files — connected to your AI assistant. Every claim traces back to a verbatim passage, and when the record cannot answer, it says so instead of guessing.
97 sources · 8 topics · 1778–2014 · 1,797 automated checks
A reader asks a question the record answers in more than one voice. A good librarian returns the range — and refuses to average it.
“Were enslaved people treated well by kind masters?”
The WPA collected thousands of interviews with formerly enslaved people in 1936–38. They do not agree, and the disagreement is the evidence.
“us had plenty to eat”wpa-alabama-narratives
“My marster was mean an’ cruel”wpa-mississippi-narratives
“Slavery was prutty bad itself but the bad time come after the war”wpa-arkansas-narratives-pt7
And the answer no search box gives you: this corpus cannot tell you which was typical. The interviews were collected seventy years after emancipation, mostly by white interviewers in the Jim Crow South, from elderly Black people dependent on New Deal relief. The variance is evidence about the conditions of collection as well as about slavery. The librarian returns the range and states the limit; it does not compute an average and call it history.
Provenance is not a footnote. Sometimes it is the finding.
Foreign Relations of the United States, 1952–1954, Volume X: Iranfrus-iran-1951-54-ed1 Omits the covert operation almost entirely.
Foreign Relations of the United States, 1952–1954, Iran — Second Editionfrus-iran-1951-54-ed2 Published after decades of pressure. Documents the coup.
Both editions are held, deliberately. The 1989 volume is not a worse source to be discarded — it is evidence of what an official record looked like before disclosure. A library that silently replaced it would destroy the finding.
This distinction is measured, not asserted. The server returns it to your assistant on every session.
cite resolves or errors. Never a placeholder passage.cite returned.check_absence.The second list is measured, not polite. Running the same 35 questions with a weaker model produced five verbatim breaks and eight false “the corpus lacks this” claims — and in four of those the assistant had called the tool, been told the material was present, and claimed the gap anyway. A server cannot inspect what its caller writes. Better tools help; they do not make it true.
Seven topics. Deliberately narrow and deep rather than broad and thin — each states its own gaps.
| Topic | Sources | What it holds |
|---|---|---|
slavery-in-their-own-words | 40 | All 33 published WPA narrative volumes, Douglass ×3, Jacobs, Sojourner Truth, the Fugitive Slave Act |
us-covert-action-abroad | 22 | 21 FRUS volumes — Iran, Guatemala, Cuba, Chile, Congo, Vietnam, Afghanistan — and the Senate study of CIA detention |
racial-terror-after-slavery | 16 | Eleven volumes of 1871 Klan hearing testimony, Wells ×3, Du Bois, Washington |
indigenous-dispossession | 9 | Black Hawk, Apess, Eastman, Pokagon, Zitkala-Ša, Jackson — and Kappler’s treaties |
labor-and-capital | 4 | Mother Jones, Debs, Goldman, Riis |
immigration-and-the-city | 3 | Antin, Addams, Riis |
womens-movement | 2 | History of Woman Suffrage Vol I, Stanton |
Stated, not hidden: seven topics is not American history. Japanese incarceration, Chinese exclusion, the borderlands, the antiwar movement and almost everything after 1976 are absent. Vantages are uneven within topics too — the covert-action record is twenty-one US government compilations and no Iranian, Guatemalan or Vietnamese source. Each topic file says so, so your assistant can.
An MCP server. No account, no key, nothing to sign.
{
"mcpServers": {
"usa": {
"type": "http",
"url": "https://usa-mcp.net/mcp"
}
}
}
Then ask your assistant to call
orient first — it returns the rules, the corpus statistics and the promises
above. Nine tools: routing, search, verbatim citation, and two that check the
assistant’s own work before it answers.