Deckary Word AI agent tool reference
Reference the reading, search, inspection, authoring, and run-control operations used by Deckary's Word AI agent.
Last verified: 2026-10-06
Deckary's Word AI agent uses four document tools plus shared run-control tools. The model chooses an operation, but the Word add-in resolves handles against the live document, enforces the destructive-edit policy, performs the Word JavaScript API calls, and returns host-observed evidence. The same document tools serve an external MCP agent when you share the document.
Read tools
| Tool | Purpose |
|---|---|
read_document | Read the outline (headings, block count, styles, selection), or exact blocks for up to ten heading sections, blocks or spans |
search_doc_text | Find the blocks that contain a phrase and return their handles |
inspect_document | Read exact runs, paragraph formatting and tracked revisions of a block, table cells, or a comment |
Handles name blocks (bk_…), heading sections (hd_…), tables (tb_…) and comments (cm_…). A handle stays valid while its block exists in the document.
Write tool
author_document applies one bounded program: it can insert headings, paragraphs, lists, tables and page breaks; replace blocks or a heading section; patch blocks with literal find-and-replace edits, new text, styles or paragraph formatting; delete blocks; and add comments. One program touches at most 40 existing blocks. review tracked records the program's changes as tracked revisions and then restores your tracking setting.
Deckary checks the complete program against the current document before anything changes. A program that names a passage that does not occur, an unknown style or a block that no longer exists is not applied at all. If the document changed after the agent's latest read, the edit is refused and the agent reads again. Deleting, replacing or rewriting existing blocks asks for your approval first.
A successful edit returns the affected blocks with fresh handles and exact text.
Shared run tools
| Tool | Purpose |
|---|---|
ask_user | Ask for a decision required to continue safely |
read_run_context | Retrieve an earlier request, tool arguments or original observation from the current workspace |
finish | Finish after observing the results; the host checks receipts before releasing the final response |
The shared Office run treats conversation text as context, not proof that a document changed. A mutation must have a host-observed receipt before finish can support a completion claim.
See Editing and review for workflow examples.