AI Tools

The office-open package provides AI SDK tools that allow AI agents (Claude, GPT, etc.) to generate .docx, .pptx, .xlsx, .odt, .rtf, .ods, and .odp files using the Vercel AI SDK.

#Installation

pnpm add office-open ai

npm install office-open ai

yarn add office-open ai

bun add office-open ai

Requires the Vercel AI SDK (ai package) v6+.

#Tools

Eight tools are provided — one per supported root format plus an on-demand schema lookup:

Each generate tool validates the input with JSON Schema (ajv) before generating. Validation errors include the instance path (e.g. at "sections.0.children.1.paragraph.children") so the model can fix and retry.

#Progressive Disclosure

The full format schema is far too large for a tool definition (~675 KB for docx), so the generate tools carry a skeleton input schema: the top-level shape plus the child wrapper keys, with stubs naming the definitions they stand for. The workflow is:

1. Call office-open-schema-lookup with { type, definitions: [...] } for the option types you are about to fill (e.g. ParagraphOptions, RunOptions, TableOptions). Names come from the skeleton stubs, or from npx office-open schema index <type>.
2. Fill the options and call the generate tool.
3. If validation rejects the input, fix the reported paths and retry — up to five issues are reported per call.

A lookup returns the requested definitions plus their dependency closure; cataloged domains you did not request stay as expandable stubs, so request each domain root you need. At most 8 definitions per call; unknown names return closest-match suggestions instead of an error.

Provider strict structured-output modes (which require additionalProperties: false everywhere) are not supported by the skeleton; validation is owned by the tool's execute via ajv.

#Usage

import { generateText } from "ai";
import { officeOpenTools } from "office-open/ai";

const result = await generateText({
  model,
  prompt: "Create a quarterly report document",
  tools: officeOpenTools,
});

#Tool Input

Each tool accepts the document options directly (no wrapping — pass the document structure as-is):

#generate-docx

{
  "sections": [
    {
      "children": [{ "paragraph": { "children": [{ "text": "Hello, World!", "bold": true }] } }]
    }
  ]
}

#generate-pptx

{
  "title": "My Presentation",
  "slides": [
    {
      "children": [
        {
          "shape": {
            "x": 100,
            "y": 100,
            "width": 760,
            "height": 340,
            "textBody": {
              "paragraphs": [{ "children": [{ "text": "Hello, World!", "size": 32 }] }]
            }
          }
        }
      ]
    }
  ]
}

#generate-xlsx

{
  "worksheets": [
    {
      "name": "Sheet1",
      "rows": [
        { "cells": [{ "value": "Name" }, { "value": "Score" }] },
        { "cells": [{ "value": "Alice" }, { "value": 95 }] }
      ]
    }
  ]
}

#Tool Output

Each tool returns a base64-encoded file:

{
  "base64": "...",
  "mimeType": "application/vnd.openxmlformats-officedocument.wordprocessingml.document"
}

#JSON Schemas

The full schemas and the slicing API are exported for custom use:

import {
  docxSchema,
  pptxSchema,
  xlsxSchema,
  validateDocumentInput,
  sliceDocumentSchema,
  SCHEMA_ENTRIES,
} from "office-open/schemas";

// Extract the dependency closure of specific option types:
const slice = sliceDocumentSchema("docx", ["ParagraphOptions", "RunOptions"]);

The same lookup is available on the CLI (npx office-open schema index docx, npx office-open schema slice docx ParagraphOptions).

For full schema details, refer to the DOCX, PPTX, and XLSX documentation.