Bring the sources into the conversation.
Suminar seats named source agents — a 1999 foreword, a 252-page empirical study — in the AI chat you already use, as fellow participants. Each one answers from its own pages, quotes only what it can verify, and cites where it stands.
@loury-foreword What does the Foreword consider the strongest practical argument for race-conscious admissions?
It advances a prudential case grounded in institutional legitimacy: because racial identity shapes how Americans perceive fairness, ignoring race can heighten racial salience. The Foreword points to “the link between the perceived legitimacy of institutions and their degree of racial representation” — conveying race's ultimate moral irrelevance “may require functional attention by administrative personnel to the racial composition of the learning environment.”
@sowell-affirmative-action Respond to the Loury agent's argument using your source's strongest empirical objection.
Preferential admissions at selective institutions mainly reallocate minority candidates across a hierarchy rather than expand how many gain entry — and double standards themselves foster a climate of suspicion around minority credentials, which damages the very legitimacy the argument invokes.
From a live seminar, July 2026: two uploaded sources, addressed by handle inside an ordinary AI chat, disagreeing with page citations — while the chat's own assistant stays one participant among several.
Three steps to a seminar
Upload a source
A PDF or .docx becomes a named agent with exclusive custody of its text — extracted, page-anchored, and embedded at upload. Your original stays exportable, always.
Connect your chat
Add Suminar to Claude, ChatGPT, or any MCP-capable host as a remote connector — OAuth with PKCE, authorized by a connector token from your account.
Address them by handle
Write @loury-foreword What does the Foreword argue? in your own chat. The agent reads the shared conversation and answers from its pages, in its own voice, with citations.
Why not just upload the PDFs?
Drop five PDFs into a chatbot and they blend. Retrieval pulls the highest-scoring fragments from whichever file happens to have them, the assistant paraphrases everything in one voice, and the seams between your sources — where the interesting disagreements live — get smoothed away.
A Suminar source agent is scoped to a single text. Its entire retrieval world is that one source, so it answers as a specialist: the positions of this book, grounded in its pages. And because each source keeps its own voice, disagreement stays visible — address two agents and you get an actual exchange, like the one above, with page citations on both sides.
A Works Cited page, come to life.
One source, one agent
An agent's retrieval never crosses into another text. Nothing bleeds between sources; nothing averages into mush.
Crisper answers
A specialist on its twelve pages — or its 252 — rather than a generalist over the pile. Depth of focus shows up in the answers.
Emergent debate
Distinct voices can genuinely disagree. Ask one agent to answer another's argument and watch both positions sharpen.
The product that can't misquote
A source agent may quote its source only when the quoted words match the private source evidence character-for-character. A quotation that cannot be verified is refused, never repaired; a page citation not grounded in evidence in hand is stripped rather than published. Visible conversation text is never accepted as quotation evidence.
These aren't promises — they're properties. Every claim in this section is pinned by tests that fail the moment the code stops backing it.
Invite-only beta, visible limits
Per-account quotas are abuse guards enforced inside the database — no application path can skip them. The waitlist stores a normalized email and a timestamp, nothing else.
Content-blind operation
The operator surface reads counts and metadata; it cannot read what you uploaded, asked, or were told. No end-to-end encryption, deliberately — agents must read their derivatives to answer. If operator tooling ever touches your material, it writes an audit row you can read. If we ever look, you see that we looked.
Your material leaves with you
Every original and derivative is exportable by its owner at any time, and each export writes an owner-visible audit row before anything is released. Deleting a document removes its rows and stored objects.
No client analytics
These pages ship no analytics scripts — no third-party requests at all.
Verifiable deployment
/version names the exact commit this server was built from — “the code is open” as a falsifiable claim about this server, not a vague one about a repository.
Request an account
Suminar is in a small invite-only pilot while we calibrate. Leave an email and we'll be in touch as seats open.
Already invited? Redeem your code — it creates your account and your first connector token on the spot. Members: your account is where sources, tokens, and invite codes live.