The platform manual as agent tools: search, read, validate and write

An agent working on an application is only as good as what it knows about the platform. This release exposes the manual itself as MCP tools, so an agent looks up how a feature works instead of guessing, and a documentation author can have an agent draft and maintain pages under the same controls as a person.

Reading the manual

• Search and read. docs_search finds pages and docs_read returns one, in windows for long pages, each naming the full section path where the page lives.
• The table of contents. docs_section_list returns the manual's structure, so an agent can navigate it rather than search blindly.
• Live examples. A manual page can embed an application object from another dictionary and render it live, so an example shows the real component rather than a picture of it.

Writing it, under a grant

• Gated by documentation level. The tools are listed only to a caller holding a documentation grant on the dictionary; reading needs any level, writing needs the editor level, and a refusal says which level is missing and where it is set.
• Validated before it lands. docs_validate checks a page against the manual's schema, and docs_write validates again before writing. A write is a dry run by default, and an existing page is replaced only when asked to.
• Versioned like any edit. A page written by an agent goes through the same path as one edited in the viewer, so it carries the same revision history and the same change journal entry.
• Structure tools. Sections can be created and removed and pages moved or deleted, each under the same editor grant.

The documentation tools are served on scoped endpoints only, where the dictionary they act on is declared at connect, and every call is recorded in the MCP audit like any other tool.

See the feature →

← All posts