Every published documentation page is now served two ways: the page you read,
and a raw Markdown route beside it. A manifest at /docs/corpus.json lists
each document with a stable id, its canonical URL and its Markdown URL, with
cache headers, so a service or an agent can treat the documentation as a
corpus rather than scraping pages. The console assistant reads from this
surface instead of a hand-copied fork, so what it says and what you read cannot
drift apart.
For agents running in the browser, two read-only WebMCP tools join
get_pricing. search_docs ranks documents by title, slug, headings,
description and category and returns ids and both URLs. get_doc returns one
document as Markdown with citation metadata, accepting a full id or an
unambiguous slug, and says so when a long document is truncated.
The MCP reference page now carries the four scope names with their exact descriptions, and documents each credential kind: person OAuth, MCP-only org agent keys, the SSO and MFA exemption that applies to them, scoring keys, and manual session tokens.