Architecture

            page load                                   first site-wide tool call
┌─────────────────────────────────┐         ┌──────────────────────────────────────────────┐
│ agentprose.min.js  (9.9 KB gz)     │ ──────▶ │ agentprose.index.min.js  (6.6 KB gz, lazy)      │
│ • reads data-* config           │  loads  │ • autoSource: agentprose.json → feeds → sitemap │
│ • builds tool definitions       │         │   → llms.txt (+ Pagefind for search)         │
│ • registers on document.        │         │ • BM25F SearchIndex, related posts           │
│   modelContext (AbortSignal)    │         │ • site-wide tool executors                   │
│ • page extraction + Markdown    │         └──────────────────────────────────────────────┘
│   (current-article tools need   │         ┌──────────────────────────────────────────────┐
│    no network at all)           │ ──────▶ │ adapters/<cms>.min.js (2–3 KB gz, on demand) │
└─────────────────────────────────┘ data-   │ WordPress auto-detected; others by attribute │
                                    adapter └──────────────────────────────────────────────┘

Source layout

Path	Role
src/auto.ts	Script-tag entry: config from data-*, adapter loading, inspector, the window.AgentProse global
src/index.ts	createAgentTools(), the public API
src/core/tools.ts	Tool definitions (names, descriptions, JSON Schemas) and page-level executors
src/core/site-tools.ts	Site-wide executors (search, list, taxonomies, related, citation, site info)
src/core/register.ts	WebMCP registration, analytics instrumentation, polyfill loading
src/core/extract.ts	JSON-LD/OpenGraph/meta reading, article-body detection, TOC and sections
src/core/markdown.ts	HTML → GitHub-flavoured Markdown (tables, nested lists, fenced code)
src/core/feeds.ts	RSS 2.0/1.0, Atom, JSON Feed, sitemap and llms.txt parsers
src/core/source.ts	The zero-config content source, merging, filtering, caching
src/core/search.ts	BM25F index with prefix matching and accent folding
src/core/lib.ts / lib-lazy.ts	Static engine (npm) vs. lazily loaded engine (script tag)
src/adapters/*	CMS adapters and rich-text renderers
src/cli/index.ts	agentprose build and agentprose inspect (Node, linkedom)
src/inspector.ts	The 🔌 panel

Design rules

1. The page is the API. Use what sites already publish (JSON-LD, OpenGraph, feeds, sitemaps, llms.txt) before asking for configuration.
2. Pay on use. Registration is free. Network and engine code load on the first site-wide call.
3. Agents get Markdown, paged. Never truncate silently. Return next_offset.
4. Errors are results. { error, message } lets agents recover; exceptions don’t.
5. Read-only by default. Anything else is opt-in and named for what it does.
6. Secrets never reach the page. Credentials shaped like management keys are refused, and proxy mode exists for private content.
7. One code path. The inspector, instance.call(), the CLI and the agent all run the same tool code.

Premium seams (open, documented, not crippling): onToolCall / agentprose:call / data-analytics (analytics), searchEndpoint (hosted search), index (hosted index), proxy (managed CMS proxy). See BUSINESS.md.