Tool catalog

All 14 built-in tools, with exact parameters and return shapes. The schemas come from getToolSchemas() (OpenAI function-calling format) or getAiSdkTools() (Vercel AI SDK shape); execution goes through executeToolCall(name, args, bridge) or the framework hooks.

Every tool returns the same envelope:

interface AgentToolResult {
  success: boolean;
  data?: unknown; // present on success; string for most tools
  error?: string; // present on failure; written so the model can self-correct
}

The pattern is locate-then-mutate. Locate tools return paragraphs tagged with a stable paraId (Word's w14:paraId); mutate tools take that paraId as input. Paragraphs without a paraId in the source file are addressed by their ordinal index as a string, so the ids the agent reads always resolve.

Locate

read_document

Read the document as [paraId] text lines. Returns the vanilla view: pending tracked insertions are hidden, pending deletions still show as plain text (they are part of the document until accepted), and comment markers are stripped. Use read_changes / read_comments to inspect what is pending.

// call
{ "name": "read_document", "arguments": {} }
// data
"[2A1F3B] Quarterly report\n[2A1F3C] Revenue grew 14% over the prior period..."

read_selection

Read the user's current cursor or selection. No parameters. Returns a SelectionInfo object: { paraId, selectedText, paragraphText, before, after }. Fails with No selection (editor not focused). when there is no live cursor, which is always the case in headless mode.

// data
{
  "paraId": "2A1F3C",
  "selectedText": "grew 14%",
  "paragraphText": "Revenue grew 14% over the prior period.",
  "before": "Revenue ",
  "after": " over the prior period.",
}

read_page

Read one rendered page. Live editor only: the headless reviewer has no layout, so it reports zero pages.

Returns the page text as [paraId] text lines. Out-of-range pages fail with Page 9 does not exist (document has 4 pages).; headless mode fails with No pages rendered (headless mode or empty document).

read_pages

Read a contiguous range of rendered pages in a single round-trip. Live editor only, same headless caveat as read_page.

// data
"--- Page 2 ---\n[3B1C44] ...\n\n--- Page 3 ---\n[3B1C59] ..."

find_text

Locate paragraphs containing a phrase. Returns handles the agent passes straight to the mutate tools: paraId plus the matched substring (use it as search).

Returns FoundMatch[], or the string No matches.:

// data
[
  {
    "paraId": "5D0A21",
    "match": "best efforts",
    "before": "The Supplier shall use ",
    "after": " to deliver the goods by the agreed date.",
  },
]

read_comments

List all comments with authors, anchored text, and threaded replies. No parameters.

// data
"[Comment #3] Legal AI: \"Define 'material breach'.\" (anchored to: \"material breach\")\n  Reply by Dana: \"Agreed, drafting a definition.\""

read_changes

List tracked changes (insertions and deletions) currently pending in the document. No parameters.

// data
"[Change #12] insertion by Legal AI: \"commercially reasonable efforts\"\n[Change #13] deletion by Legal AI: \"best efforts\""

Mutate

add_comment

Attach a comment to a paragraph, optionally anchored to a phrase within it. In the live editor it appears in the comments sidebar immediately.

// call
{ "paraId": "5D0A21", "text": "Replace with a defined-efforts standard.", "search": "best efforts" }
// data
"Comment 7 added on 5D0A21."

Fails when the paraId does not exist or search is missing from / ambiguous within the paragraph.

reply_comment

Reply to an existing comment; the reply threads under the original.

Returns Reply 9 added to comment 7. Fails with Comment #7 not found. for unknown ids.

resolve_comment

Mark a comment as resolved (done).

Returns Comment 7 resolved.

suggest_change

Suggest a tracked change the user accepts or rejects. Never edits text directly. Three modes selected by empty strings:

// call
{ "paraId": "5D0A21", "search": "best efforts", "replaceWith": "commercially reasonable efforts" }
// data
"Replacement proposed: \"best efforts\" → \"commercially reasonable efforts\" on 5D0A21."

Fails when the paraId is unknown, search is missing or ambiguous, or the target overlaps an existing tracked change.

apply_formatting

Apply character formatting to a whole paragraph or a unique phrase within it. This is a direct edit, not a tracked change; exclude it in strict redlining workflows.

marks keys:

// call
{ "paraId": "2A1F3B", "marks": { "bold": true, "color": { "rgb": "1A73E8" } } }
// data
"Formatting applied to 2A1F3B."

Out-of-spec values fail early with the allowed list in the error message, so the model can self-correct instead of writing OOXML Word would discard.

set_paragraph_style

Apply a paragraph style by id. The styleId must exist in the document's styles.xml; unknown ids are rejected. Direct edit, not a tracked change.

// call
{ "paraId": "2A1F3B", "styleId": "Heading1" }
// data
"Style \"Heading1\" applied to 2A1F3B."

Navigate

scroll

Scroll the editor viewport to a paragraph. Does not move the user's cursor. In headless mode it validates the paraId and reports success without doing anything (there is no viewport).

Returns Scrolled to 2A1F3B. or fails with paraId 2A1F3B not found.

Next steps

• AI editing tutorial, wire these tools to a live editor
• AI redlining, tool subsets for tracked-change-only review
• Word JS API parity, how each tool maps to Office.js
• API reference