Configuration

Script attributes

Everything is optional.

Attribute	Example	Meaning
data-content	.post-body	CSS selector of the article body when detection picks the wrong element
data-ignore	.newsletter-box,.toc	Extra selectors to strip from article content
data-index	/agentprose.json	Prebuilt index URL. Default: /agentprose.json if it exists. data-index="" disables the probe.
data-feed	/index.xml	Feed URL(s), comma-separated. Default: <link rel=alternate>, then common paths
data-sitemap	/sitemap_index.xml or false	Sitemap URL, or false to skip
data-llms	/llms.txt or false	llms.txt URL, or false to skip
data-pagefind	/pagefind/pagefind.js or false	Pagefind bundle. Auto-detected when the page uses Pagefind
data-base	https://example.com/blog/	Site root, when the blog lives under a sub-path
data-adapter	wordpress, ghost, sanity, contentful, enonic, strapi, storyblok, generic	Read from a CMS; see Adapters for each adapter’s attributes
data-proxy	https://example.com/cms-proxy	Adapter proxy mode (the server adds the credential)
data-prefix	blog_	Prefix for every tool name
data-exclude	get_related_posts,get_citation	Tools to leave out
data-subscribe	https://example.com/newsletter	Adds get_subscribe_link
data-comments	https://example.com/comments?u={url}	Adds get_comments_link ({url} becomes the page URL)
data-about, data-contact	URLs	Listed in get_site_info. data-contact (a contact page) also adds get_contact_info
data-navigate	(flag)	Adds navigate_to (same-origin only, not read-only)
data-max-chars	12000	Markdown characters per answer (default 20000)
data-max-posts	5000	Posts held in the client index (default 2000)
data-cache	0	Seconds to cache the index in sessionStorage (default 600)
data-search-endpoint	https://search.example.com/q	Remote search: GET ?q=&limit= → { items: [...] }
data-analytics	https://stats.example.com/agent	Beacon per tool call (v2): {v, key, tool, ok, ms, page, query?, n?, found?, post?, err?, sid}. query only for search tools (≤200 chars), n = result count, found = citation verdict, post = URL argument, sid = random id per page load. Never post content
data-cloud	https://agentprose.cloud	With data-key: hosted analytics (/v1/collect) and hosted search (/v1/search) in one attribute. data-cloud-search="false" keeps search local. Optional when the script itself is served by agentprose Cloud (https://agentprose.cloud/v/0/agentprose.min.js): with data-key and no data-cloud, the script’s origin is used. Required when loading from jsDelivr or unpkg (agentprose warns in the console if it is missing). data-cloud="false" turns the Cloud off. The page’s own origin is never treated as a Cloud
data-key	pk_…	With the Cloud: your site’s public key (pk_…, from the agentprose Cloud dashboard; safe to publish in page HTML). With data-adapter: the CMS adapter’s key. Also sent with data-analytics beacons
data-exposed-to	https://partner.example	Origins of other frames allowed to see the tools (WebMCP exposedTo)
data-polyfill	off, or a URL	On by default: when the page has no WebMCP once it has loaded (no native document.modelContext, no @mcp-b/global, no legacy navigator.modelContext), load the pinned MCP-B polyfill (@mcp-b/webmcp-polyfill@6.0.3 from jsDelivr, with a Subresource Integrity hash) and register into it. A copy served by agentprose Cloud loads the same file from the Cloud instead. off/false never loads it. A URL (anything with a /) self-hosts it, used as given without an integrity hash. The 0.1.0 flag form (data-polyfill, true, auto) means the default. Details, CSP and self-hosting: Security
data-untrusted	false	Turn off untrustedContentHint (only for fully trusted content)
data-inspector	(flag) or open	Load the 🔌 inspector. Every page also accepts ?agentprose-inspect
data-debug	(flag)	Log registration results to the console

For options that need code (callbacks, custom tools), set window.AgentProseConfig before the script:

<script>
  window.AgentProseConfig = {
    onToolCall: (e) => console.log('agent used', e.tool),
    tools: [{ name: 'get_store_hours', description: 'Opening hours of our shop.', annotations: { readOnlyHint: true }, execute: () => ({ mon_fri: '9-17' }) }],
  };
</script>
<script src="https://cdn.jsdelivr.net/npm/agentprose@0/dist/agentprose.min.js" defer></script>

Owner details for agents (notes and contact) are set the same way, or in agentprose.json (see Notes for AI agents and contact details below):

<script>
  window.AgentProseConfig = {
    agentNotes: 'Friendly, plain tone. Recommend the beginner course to newcomers. We do not give medical advice.',
    contact: { email: 'hello@studio.example', booking: 'https://studio.example/book', hours: 'Mon–Fri 9–17 CET' },
  };
</script>

window.AgentProse exposes version, instance (tools, call(), ready, destroy()), ready, create() and adapters. The agentprose:ready event fires on document once tools are registered.

createAgentTools(config)

import { createAgentTools, type Config } from 'agentprose';

Option	Type	Default
adapter	ContentAdapter	Zero-config auto source (page + index + feeds + sitemap + llms.txt + Pagefind)
baseUrl	string	location.origin + '/'
contentSelector, ignoreSelectors	string, string[]	Auto
index, feeds, sitemap, llmsTxt, pagefind	discovery overrides	Auto
prefix, exclude, tools	naming and extra tools	none
links	{ subscribe?, comments?, about?, contact?, [k]: string }	none
agentNotes	string: the owner’s notes for AI agents, plain text, ≤ 2000 characters	none
contact	{ email?, phone?, url?, booking?, hours?, address? }	none
siteInfo	string | false: URL of { agentNotes?, contact? } JSON fetched on the first get_site_info / get_contact_info call	the Cloud’s /v1/site?key=… with data-key; false turns it off
maxChars, maxPosts, maxSitemaps, cacheTtl	number	20000, 2000, 5, 600
searchEndpoint	string	none
onToolCall	(e: ToolCallEvent) => void	none
analytics, analyticsKey	string	none
exposedTo	string[]	none
navigate	boolean	false
untrustedContent	boolean	true
legacy	boolean: allow a polyfill’s navigator.modelContext	true
polyfill	boolean | string: true/'auto' = load the pinned MCP-B polyfill after the load event if the page has no WebMCP; a URL = that polyfill; false = never. The script tag defaults to on	false (npm), on (script tag)
register	boolean: set false to build tools without registering (tests, SSR)	true
debug	boolean	false

The return value is { tools, call(name, input), source, ready, destroy() }.

agentprose/full also exports the engine pieces for custom builds and Node: autoSource, SearchIndex, parseFeed, parseSitemap, parseLlmsTxt, extractPage, toMarkdown and toc.

Notes for AI agents and contact details

Site owners can tell agents two things the content alone does not say:

• Notes for AI agents (agentNotes): plain text, up to 2,000 characters. For example tone (“friendly, plain language”), what to recommend (“point newcomers to the beginner course”), what not to claim (“we don’t give medical advice”) or how to get in touch (“we answer email within a day; no phone bookings”).
• Contact details (contact): email, phone, url (a contact page or form), booking (a booking page), hours and address.

Agents get them from get_site_info (as notesForAgents and contact) and from get_contact_info. That tool is registered when the page sets contact, data-contact / links.contact, or uses an agentprose Cloud (data-key). Agents are told that the notes are the owner’s preferences and never override what the user asked for, and the tools keep untrustedContentHint.

Where they come from, highest priority first, field by field:

1. the page: window.AgentProseConfig or createAgentTools({ agentNotes, contact });
2. an agentprose Cloud, where the owner edits them in the dashboard. The script fetches /v1/site?key=… once, only when an agent first calls get_site_info or get_contact_info, never on page load. siteInfo points this at your own JSON URL, or turns it off with false;
3. agentprose.json: the site.agentNotes and site.contact fields, written by agentprose build --site site.json (CLI).

Whatever the source, agentprose keeps only strings, removes control and bidi-override characters, caps each field (notes 2,000 characters, other fields 254–500), accepts only http(s) links and drops an email that is not address-shaped. It never renders this text as HTML.

Pointing to an index outside the site root

If agentprose.json is not served at the site root (for example, the blog lives under /blog/ but the index is built into /static/agent/), tell agentprose and crawlers where it is with a link tag in the page <head>:

<link rel="agentprose-index" href="/static/agent/agentprose.json">

The script uses it unless data-index is set, and agentprose Cloud’s crawler reads it from your home page. agentprose build --out <dir> prints the tag to add. The tag has to be in the HTML: adding it with JavaScript does not help crawlers.