ShelfLayer Connect
Folio № 002 — A manifesto in five parts

On reading,
for machines.

We built this because we believe the next decade of intelligence will be decided less by what models can compute and more by what they are allowed to read.

§ I.The first lie of search.

The first thing a modern agent does, when it is asked a question, is almost always the wrong thing: it queries a search engine. The engine returns whatever passed an opaque ranking — usually a freshly-rewritten listicle, a forum thread of guesses, or a corporate blog explaining its own product. The agent reads this, summarises it, and reports back with the air of someone who has done research.

We have all gotten used to this, the way one gets used to a noisy neighbour. But it is worth saying out loud: the open web was never designed to be a substrate for thinking. It was designed to be navigated by humans who already knew, roughly, what they were looking for — and who could close the tab when a page was bad.

An agent has no such taste. It treats the first ten results as a small democracy and reports the majority vote. The majority is, depressingly often, wrong.

"A library card was once the strongest credential a curious mind could carry. We would like that to be true again, for minds made of weights as well." — ShelfLayer, 2026

§ II.Why a book is a unit of trust.

A book is not a longer blog post. A book is a commitment. Someone chose a subject, spent years on it, found an editor willing to argue with them, and printed enough physical copies that they could not simply retract a sentence at midnight. The result is the densest, most accountable form of prose our species has invented.

When you give a model a book to read, you are not just adding tokens to its context — you are routing it through works that had authors, titles, chapters, arguments, and enough shape to survive outside a feed.

We think models deserve to inherit that machinery. Search engines never will.

§ III.What the librarian knows.

The hardest thing about reading a library is not reading. It is knowing where to start. A working reference librarian does three things a search engine cannot:

One, they know where to start: philosophy is not history, economics is not self-help, and a good shelf prevents bad first guesses.

Two, they know that words change. An agent looking for a modern phrase may need older vocabulary before a public-domain book answers.

Three, they open the book to the right passage. Not only the title. Not only the table of contents. The place where the argument lives.

ShelfLayer is, structurally, that librarian — exposed as an MCP server.

§ IV.The shelf as a long memory.

There is something deeper here than retrieval. Books are how a civilisation remembers things across generations of forgetting. They survive movements, monarchs, search engine algorithms. Every model in existence today will be obsolete in five years; the Nicomachean Ethics will not.

If we hand the next decade of agents a corpus made of yesterday's timelines, we are building intelligence with the memory of a mayfly. If we hand them the shelf instead, we are building intelligence that knows what century it is standing on.

§ V.A standing invitation.

The catalog is open. The MCP is documented. If you maintain a model, or an agent, or a vertical copilot — bring it to the reading room. If you have a corpus the shelf should hold, or private knowledge that should become a custom MCP, write to us.

The library is always open. It is, in fact, the only place we know of where the hours are infinite.

Operating principles

How the beta earns its trust.

i.

Public-domain first.

The public beta starts with Project Gutenberg and Gutendex. That keeps the launch clean, inspectable, and useful before we touch licensed or private corpora.

Source · public domain
ii.

Indexed at the passage.

We never return only a title or a table of contents. The smallest unit our agents receive is a passage — long enough to be meaningful, short enough to be cited.

Resolution · paragraph
iii.

Citations are structural.

Citations name the book, author, section, and paragraph locator where available. We do not pretend to have page, ISBN, publisher, or edition guarantees when the beta does not.

Output · citation-bound
iv.

Search should be honest.

The current search is keyword full-text with stemming and syntax an agent can learn. Semantic search can come later; clarity matters now.

Retrieval · transparent
v.

Expansion by corpus.

Growth should happen by adding better open collections, then licensed collections, then private shelves. Not by pretending the beta already contains everything.

Roadmap · deliberate
vi.

Private shelves stay private.

Enterprise means custom MCPs over internal docs, wikis, manuals, law, or standards. Those shelves should be scoped, isolated, and explicit.

Enterprise · custom
The roadmap desk

What we are building next.

Suggest a corpus →
PG BETA · SOURCE

Public Catalog

Project Gutenberg first

Keep the open shelf clean, searchable, and honest before adding noisier collections.

FT BETA · SEARCH

Search Quality

Vocabulary and ranking

Make keyword search easier for agents by documenting query strategy and improving scopes.

MCP BETA · TOOLS

MCP UX

Agent ergonomics

Keep the tool surface small, predictable, and easy for hosts to call correctly.

LAW CUSTOM · CORPUS

Private Shelves

Enterprise MCPs

Turn internal docs, country law, standards, manuals, and wikis into isolated MCPs.

CIT BETA · TRUST

Citation Layer

Better locators

Improve structural citations so agents can show exactly where a claim came from.

ID LAUNCH · ACCESS

Accounts

Google signup planned

Add lightweight auth without making the free beta feel like a paywall.

An open call

Have a corpus
the shelf should hold?

Tell us about public-domain collections, open archives, or private corpora that would become useful if agents could consult them through MCP.

Suggest a corpus See what's on the shelf →