Default tools

Content tools (CDA-backed, read-only)

All set annotations: { readOnlyHint: true }.

contentstack_search_entries

Full-text search across configured content types.

Input	Type	Description
query	string	Free-text search
content_types	string[]	Filter by UID
limit	integer	1–50, default 10
offset	integer	Pagination offset
locale	string	e.g. 'en-us'

contentstack_get_entry_by_url

Fetch entry by public URL path.

Input	Type	Description
url	string	Path starting with /
locale	string	Optional locale

contentstack_get_entry_by_uid

Fetch entry by content type UID and entry UID.

Input	Type	Description
content_type	string	Content type UID
uid	string	Entry UID
locale	string	Optional locale

contentstack_list_content_types

List configured/exposed content types. No input.

Browser context tools

get_current_page

Returns window.location and document.title. No SDK call.

get_site_info

Returns discovery pointers (llms.txt, OpenAPI, server MCP, etc.). Configurable via siteInfo.

navigate_to

Client-side navigation. Requires navigate callback in register(). Not read-only.

Input	Type	Description
path	string	Internal path starting with /

How search works

contentstack_search_entries matches CDA-side, not in the browser:

• The query is capped at 200 characters and 8 tokens, and each token is regex-escaped
• Per content type: AND across tokens, OR across searchableFields (case-insensitive)
• searchableFields defaults to [titleField, bodyField] when omitted
• meta.total is the summed CDA match count, so it reflects the real result set
• Each type fetches only enough rows to cover the requested page — no full content download
• References are not fetched for search results, only for get_entry_by_*

Two limits worth knowing:

Limit	Detail
Deep offsets	The CDA returns at most 100 entries per request, so offset + limit above 100 cannot be paged into
Ordering	Results are grouped by content type in config order, not interleaved by relevance

Tool execute response format

Per the WebMCP spec, execute resolves to any serializable value and the agent receives it directly — there is no MCP-style content envelope. Tools return their response shape as-is:

{
  "data": [{ "uid": "blt123", "title": "Getting Started" }],
  "meta": { "total": 42, "limit": 10, "offset": 0, "has_more": true }
}

On failure they resolve to an error object instead:

{ "error": { "code": "tool_error", "message": "No entry found for URL: /missing" } }

Wrap custom handlers in safeExecute() from @timbenniks/contentstack-webmcp to get that error shape automatically, or build one with toolError().

Annotations

Annotation	Set on	Meaning
readOnlyHint	all tools except navigate_to	Tool does not modify state
untrustedContentHint	the three CDA entry tools	Output is author-written content the package has not validated, so agents should treat it as untrusted input