A reference your agent
can actually read.

No separate summary layer. Browse the same knowledge as a person, retrieve an exact document, or take the whole library with you.

Start with the index

Give your agent /llms.txt for a compact map, or /agent.json for structured discovery. The index currently contains 181 references across 7 subjects.

  1. Discover. Read the index and choose references relevant to the question.
  2. Retrieve. Follow each document’s markdownUrl or jsonUrl. Keep the stable url for citations.
  3. Ground the answer. Distinguish what the reference says from your own inference. Follow source links when the distinction matters.

Choose a format

Agent manifest/agent.json
Machine-readable entry point and available capabilities.
Compact text index/llms.txt
A lightweight directory for choosing the right references.
Full text for agents/llms-full.txt
The complete reference text in one file.
JSON index/api/v1/index.json
Document metadata, stable URLs, subjects and revision.
Full-text search data/api/v1/search.json
Download once, then filter locally. This is not a search API.
Source coverage/api/v1/coverage.json
Inspect the relationship between references and source material.
Complete Markdown/downloads/knowledge-base.md?raw=1
A portable copy of the full knowledge base.
Complete JSON/downloads/knowledge-base.json
Structured knowledge-base export.
Agent CLI/agent-cli.mjs
A downloadable command-line entry point. Inspect the script before running it.
Read-only API/developers/
Typed contracts, working examples, errors and compatibility policy.
Agent instructions/AGENTS.md?raw=1
When to use this library, how to retrieve references and where to stop.

Work through the browser

Every page exposes window.WorkshopAgent, interface version 1. The visible search form uses the same query method. Calls return structured outcomes with an ok flag; clipboard failures are reported, never treated as success.

const library = await window.WorkshopAgent.discover();
const matches = await window.WorkshopAgent.query({
  q: "verification",
  category: "",
  expectedRevision: library.revision
});
const state = window.WorkshopAgent.inspect();
await window.WorkshopAgent.reset({ expectedRevision: library.revision });

// Use a real id returned in matches.results.
if (matches.ok && matches.results.length) {
  await window.WorkshopAgent.copy({
    id: matches.results[0].id,
    format: "markdown",
    expectedRevision: library.revision
  });
}

query({ q, category, expectedRevision }) accepts a text query and an exact subject ID. An empty category searches every subject. Matching is case-insensitive and requires every search term to occur in a reference. Text search downloads the full-text JSON file and filters it in this browser; there is no server-side search endpoint. Category-only filtering does not download the full text.

copy({ id, format, expectedRevision }) accepts markdown or url. Markdown is fetched from the actual document. If clipboard access is unavailable, a read-only selection appears for manual copying. A failed copy returns copied: false.

All operations return a requestId, status, revision and current state. Pass the optional expectedRevision to query, reset or copy to reject a changed version with STALE_REVISION. Unknown document and category IDs return typed errors. A downloaded index must match this page’s revision; reload the page if it does not.

Use the command line

Download and inspect the CLI script, then run it with Node.js. Replace the example origin with this site’s origin. Add --revision with a known revision to pin your request.

node agent-cli.mjs query --source https://SITE.here.now --q 'verification' --limit 5

Cite a reference, keep its revision

Use the document’s permanent topic URL for a citation, its Markdown URL for extraction, and the index revision to identify the version you consulted. Downloads and readable pages are generated from the same reference content.

Current revision 22ce5d34dbcb5a78

Public files, clear boundaries

This site serves public reference files. It does not run inference, authenticate agents, or expose an MCP server. The browser interface only discovers, filters and copies published material; your agent’s own tools remain under your control.

Search runs on your device, without analytics code. Queries appear in the address bar so results can be revisited or shared; navigating a query URL may send that URL to the hosting service. Don’t put private information in a search you intend to share.

Original source material remains attributed to its authors. Public access to these references does not grant new ownership or licensing rights over linked material. Some original source links may require separate access.