ShelfLayer Connect
Folio № 005 — MCP reference for the public beta

Reading,
as a protocol.

Connect an MCP host to ShelfLayer, then let agents browse shelves, find books, search passages, read sections, and cite what they found.

§ 01 · AuthenticationGoogle sign-in.

ShelfLayer is free during public beta and protected by OAuth. When your AI client asks to connect, continue with Google and approve the ShelfLayer authorization flow.

MCP endpoint: https://mcp.shelflayer.com/mcp ⌘C · Copy

Use the HTTPS endpoint exactly as shown. OAuth clients and modern MCP hosts expect the server to be reachable over the public internet.

§ 02 · Connect clientsAdd ShelfLayer to your AI.

Use the same remote MCP endpoint in every client. If the client asks for a connector name, use ShelfLayer. If it asks for a URL, use https://mcp.shelflayer.com/mcp.

ChatGPT

  1. Open ChatGPT on web and go to Apps or Settings.
  2. Enable developer mode for custom MCP connectors.
  3. Create a new app or connector.
  4. Name it ShelfLayer.
  5. Set the MCP server URL to https://mcp.shelflayer.com/mcp.
  6. Scan tools, connect, and complete the Google sign-in.

Claude

  1. Open Claude settings and go to Connectors.
  2. Click Add connector or Add custom connector.
  3. Name it ShelfLayer.
  4. Set the connector URL to https://mcp.shelflayer.com/mcp.
  5. Click Connect and approve the Google sign-in flow.

Claude Code

Add the remote HTTP MCP server, then authenticate from the MCP panel.

claude mcp add --transport http shelflayer https://mcp.shelflayer.com/mcp
claude
/mcp

Codex

Add the streamable HTTP server to Codex, then run the OAuth login.

codex mcp add shelflayer --url https://mcp.shelflayer.com/mcp
codex mcp login shelflayer

Cursor

Use Cursor's MCP settings or add a project/global MCP config.

{
  "mcpServers": {
    "shelflayer": {
      "url": "https://mcp.shelflayer.com/mcp"
    }
  }
}

Other MCP hosts

Choose Streamable HTTP or remote HTTP transport when available.

Name: ShelfLayer
URL: https://mcp.shelflayer.com/mcp
Auth: OAuth / browser sign-in
If a client shows Missing Bearer token after opening the endpoint directly in a browser, that is expected. The MCP URL is a server endpoint, not a public web page; connect through an MCP client.

§ 03 · First callAsk like a librarian.

Use your MCP host's remote server configuration, then call the librarian like any other tool.

QUICKSTART · MCP HOST
{
  "name": "ShelfLayer",
  "transport": {
    "type": "streamable-http",
    "url": "https://mcp.shelflayer.com/mcp"
  }
}
QUICKSTART · TOOL CALL
{
  "name": "search_passages",
  "arguments": {
    "query": "\"compound interest\" OR thrift economy",
    "shelf": "economics-and-finance",
    "limit": 5
  }
}
QUICKSTART · SEARCH SYNTAX
# Keyword full-text, not semantic search.
"hard work" thrift economy
"compound interest" OR usury
+virtue -vice
"division of labour" shelf:economics-and-finance
QUICKSTART · CITATION FLOW
# 1. search_passages returns passage_id values.
# 2. read_section opens surrounding context when needed.
# 3. get_citation turns a passage_id into a user-facing citation.

§ 04 · Example promptsGood ways to ask.

ShelfLayer works best when the agent treats it like a library, not a general web search engine. Start broad, then ask it to inspect and cite.

Research

"Use ShelfLayer to find passages about thrift, industry, and compound interest in older economics books. Cite the strongest passages."

Book discovery

"List philosophy books about ethics or self-command, then inspect the most relevant tables of contents before quoting."

Context check

"Search for 'liberty' in John Stuart Mill, read the surrounding section for the best match, and summarize the argument with a citation."

Vocabulary retry

"If modern terms do not match, retry with older vocabulary: industry for hard work, economy for thrift, political economy for economics."

§ 05 · MCP toolsThe public beta surface.

The server exposes eight read-only tools. Six are the native ShelfLayer tools; search and fetch are compatibility tools for clients that expect that two-step pattern.

ToolUse it whenStatus
searchCompatibility search that returns result IDs for clients that expect search/fetch.Beta
fetchCompatibility fetch that opens a result ID returned by search.Beta
list_shelvesDiscover available categories before searching.Beta
list_booksFind books by title, author, year, language, or shelf.Beta
inspect_bookRead metadata and the full table of contents.Beta
search_passagesSearch passage text with optional shelf, book, or section scopes.Beta
read_sectionRead all passages in a section in order.Beta
get_citationTurn a passage id into a human-readable citation.Beta

§ 06 · search_passagesThe tool that opens the page.

MCP  tools/call search_passages ⌘C · Copy

Searches the indexed passage text. This is PostgreSQL full-text search with English stemming and websearch-style syntax, not a hidden vector index. Agents should try several vocabulary variants for serious questions.

Arguments

FieldTypeDescription
query string Required Keyword query. Supports quoted phrases, OR, required terms, and excluded terms.
shelf string Optional Limit to one shelf, e.g. economics-and-finance or philosophy.
book_id string Optional Limit to one book after using list_books or inspect_book.
section_id string Optional Limit to one section after inspecting a book's table of contents.
limit integer Optional Max passages to return. Keep small for agent loops. Default is intentionally conservative.
include_notes boolean Optional Include notes, indexes, bibliographies, and similar apparatus when deliberately needed.

Response

Returns ranked passages with IDs, book metadata, section locators, and citation info. Use read_section for surrounding context and get_citation for a clean final citation.

Passages found Try alternate vocabulary Shelf not held Query too broad

§ 07 · Books and sectionsNavigate before quoting.

MCP  list_books · inspect_book · read_section ⌘C · Copy

Use list_books to discover a title, inspect_book to get metadata and its table of contents, and read_section when the agent needs the surrounding argument instead of a single passage.

§ 08 · ShelvesThe beta taxonomy.

list_shelves returns the top-level categories currently used by the public catalog. The taxonomy is practical and evolving; expect it to improve as more corpora come online.

§ 09 · CitationsStructural, not page-perfect.

ShelfLayer citations are book and structure based: title, author, year where known, section, and paragraph locator. They are designed for transparent agent answers, not as a guarantee of a specific print page, ISBN, publisher, or modern edition.

CITATION
{
  "name": "get_citation",
  "arguments": {
    "passage_id": "psg_..."
  }
}

§ 10 · MCP ServerThe shelf, in your agent's hands.

ShelfLayer is the MCP server. Connect your host with Streamable HTTP and your agent gets the shelf as tools, without waiting for a separate SDK.

Remote server

https://mcp.shelflayer.com/mcp

The server advertises OAuth. Connect through your client, sign in with Google, then let the client store and refresh the access token.

§ 11 · TroubleshootingWhen the librarian cannot help.

Empty results usually mean the query vocabulary is too modern, too broad, or simply absent from the public shelf. Agents should retry with older terms, quoted phrases, and alternate shelf scopes before giving up.

SymptomWhat it meansWhat to try
Missing Bearer tokenYou opened the MCP URL directly in a browser.Connect from an MCP client and complete Google sign-in.
No matching passagesThe terms may be too modern or absent from the current corpus.Retry with older vocabulary, quoted phrases, or a shelf/book scope.
OAuth loopThe client may have cached an old connector or callback.Remove the connector, add it again, and use https://mcp.shelflayer.com/mcp.
Tool not foundThe host has not refreshed the tool list.Reconnect the MCP server or restart the host application.

§ 12 · LimitationsReading at a sustainable pace.

Public beta limits are operational and may change while the server is being tuned. Keep agent loops modest: search first, inspect only the books that matter, then read sections when context is needed.

  • The public beta corpus currently starts with Project Gutenberg and does not include every public-domain book.
  • Search is keyword full-text search, not semantic/vector search yet.
  • Citations are structural locators, not page numbers from a specific print edition.
  • All exposed tools are read-only. ShelfLayer does not modify files, accounts, or books.
  • Very broad agent loops may be rate limited during beta.

§ 13 · SupportContact the curators.

For connection problems, review questions, corpus suggestions, or enterprise shelves, email hello@shelflayer.com. Include the MCP host you used, the approximate time, and any error message.

You can also open the support page or read the privacy policy.