Overview

@eigenpal/docx-editor-agents lets a language model read and edit a Word document: pull out paragraphs, leave comments, suggest tracked changes, apply formatting. The same tools run in three places: a live editor in the browser, a headless reviewer on the server, or an MCP server. You point them at whatever model you already run; the tool definitions are plain JSON Schema, with adapters for the Vercel AI SDK, OpenAI-style function calling, React, and Vue.

npm install @eigenpal/docx-editor-agents

Integration shapes

1. Live editor panel. Tool calls execute against a <DocxEditor> running in the user's browser. Comments and tracked changes appear in the open document. In this setup your API route sees chat messages, tool names, arguments, and tool results; the DOCX file stays client-side.

const { tools, executeToolCall, getContext } = useDocxAgentTools({
  editorRef,
  author: 'Assistant',
});

Full walkthrough: AI editing with the Vercel AI SDK.

2. Headless DocxReviewer on the server. Parse a buffer, run a review, serialize the modified DOCX back. No DOM, no editor instance. Fits a queue worker, Lambda, or CI bot.

import { DocxReviewer } from '@eigenpal/docx-editor-agents';

const reviewer = await DocxReviewer.fromBuffer(buffer, 'Reviewer');
reviewer.addComment({ paragraphIndex: 5, text: 'This cap seems too low.' });
reviewer.replace({ paragraphIndex: 5, search: '$50k', replaceWith: '$500k' });
const out = await reviewer.toBuffer();

Reference: DocxReviewer.

3. MCP server. Expose the tool catalog over JSON-RPC for any MCP-compatible client. Transport-agnostic core; you own auth, storage, and the transport.

import { McpServer } from '@eigenpal/docx-editor-agents/mcp';
import { createReviewerBridge } from '@eigenpal/docx-editor-agents';

const server = new McpServer(createReviewerBridge(reviewer), {
  name: 'acme-review',
  version: '1.0.0',
});

Reference: MCP server.

Which shape to pick

If you are building a product UI, start with the live editor. If you have documents and no UI, start with DocxReviewer. Pick MCP only when the client already supports MCP.

The tools

• Locate: read_document, read_selection, read_page, read_pages, find_text, read_comments, read_changes
• Mutate: add_comment, reply_comment, resolve_comment, suggest_change, apply_formatting, set_paragraph_style
• Navigate: scroll

The pattern is locate-then-mutate, mirroring Word's JS API: locate tools return paragraphs tagged with a stable paraId, and mutate tools address by it (how paraId anchoring works). Full parameter and return shapes: tools reference.

Subpaths

Each subpath tree-shakes independently. The /ai-sdk/* entries import ai, which is an optional peer dependency; skip them if you are not on the AI SDK.

Next steps

• AI editing with the Vercel AI SDK
• AI redlining and tracked-change review
• Live editor bridge reference
• API reference