Your shelf, staffed.

The one rule

Most retrieval systems fetch everything, then filter out what you should not see. Ours inverts that: what you own is the search scope, not a filter applied after the fact. Material outside your shelf is never ranked, never loaded — it was never in.

“A post-filter that fails leaks paid material. A scope that fails returns nothing.”

Before you buy, and after

This is the whole pitch, as a table. It is exactly what an agent gets back, per caller, per guide.

A held guide, a withheld guide, or an unknown slug all resolve the same way: the catalog view, never text. If anything about entitlement can’t be confirmed, the answer is the closed one.

What your AI can do with a shelf

These are the real tools, wired to one server. There is no separate “lite” version — the REST routes and the MCP tools call the same code underneath.

shelf_list

What’s on your shelf: the guides you own, readable, and the ones you don’t, shown as bookmarks with price and scale — never text.

shelf_read

Read one guide you own, unit by unit, verbatim, each with a citation anchor pointing at the exact place in the source. A guide you don’t own returns its catalog and not one word of prose.

shelf_ask

Ask a question of the books behind the guides you can read. Answers come back as verbatim passages with their book, chapter, and type — never a paraphrase. If your shelf doesn’t cover it, you get a refusal naming what would, and what it costs.

shelf_disagreements

Where the field disagrees with itself — misconceptions, named tensions, and places where a guide’s own sources contradict each other. One voice cannot disagree with itself, so this is a real signal, not a search trick.

shelf_verify

Check a quote against its source before your agent attributes it to a guide. A cited answer is only worth more than an opaque one if the citation can be checked.

Connect it

This runs over MCP. Claude Desktop and Claude Code support it today; other assistants that speak MCP will work the same way as they add support — we are not claiming it works everywhere yet.

1. Sign in and mint a shelf token from your account page. It’s a 30‑day token scoped to your shelf.
2. Paste the config below into your MCP client, with your token in place.
3. Ask it what’s on your shelf, or ask it a question — it will cite the guide and the book behind every answer.

{
  "mcpServers": {
    "bicycle-guide": {
      "command": "npx",
      "args": ["tsx", "scripts/mcp-bicycle-guide.ts"],
      "env": {
        "BICYCLE_SHELF_TOKEN": "<your shelf token, from /account>"
      }
    }
  }
}

No token, no shelf — the server answers with an empty shelf and connection instructions rather than an error. The registry tools (browsing what exists) work with no token at all; the shelf tools (reading what you own) always require one.

Limits, honestly

• This is MCP, not a universal plugin. It works with an MCP‑capable client — Claude Desktop and Claude Code today — not with every assistant.
• Semantic search over book passages depends on how much of a guide’s corpus has been indexed. Every answer states its coverage, so you can tell a thorough search from a partial one.
• Nothing is summarized on your behalf. If your shelf doesn’t contain the answer, you get a refusal and a price, not a confident guess dressed up as one.