How to Give an AI Agent the Ability to Create Documents

LLM agents are great at reasoning, but when they need to hand back a spreadsheet or a report they usually dump a Markdown table into the chat. Here is the pattern that turns that into a real, shareable document — and the runnable code to add it to your agent today.

The fastest way to give an agent document output is a single tool that calls a document API and returns a URL. The agent decides when to use it, supplies the data, and your code runs the request. The result is a link the user can open and edit — not 2,000 tokens of pipe-delimited text the model had to generate character by character.

Why a table-in-chat is the wrong output

Generating a large table token-by-token is slow, expensive, error-prone, and produces something the user still has to copy, clean, and paste into a real spreadsheet. Worse, the model often truncates or mis-aligns columns. A document API removes all of that: the agent sends structured rows, and a hosted file comes back with a shareable link. For the full reasoning, see why AI agents need a document API.

The pattern: one tool, one POST, one link

Every framework below uses the same three-step shape: (1) define a tool with a title androws input, (2) inside it, POST to a document API, (3) return the URL. Here is the bare API call the tool wraps:

The underlying call (any language)

POST https://openofficeai.com/api/v1/sheets
Authorization: Bearer YOUR_API_KEY
Content-Type: application/json

{ "title": "Q3 Revenue", "sheets": [{ "rows": [
  ["Month", "Revenue"], ["July", 48200], ["August", 61900]
] }] }

# → { "url": "https://openofficeai.com/s/..." }

OpenAI function calling

Declare a function tool; when the model calls it, run the request and feed the URL back.

Python

tools = [{
  "type": "function",
  "function": {
    "name": "create_spreadsheet",
    "description": "Create a shareable spreadsheet; returns a URL.",
    "parameters": {"type": "object", "properties": {
      "title": {"type": "string"},
      "rows": {"type": "array", "items": {"type": "array"}}},
      "required": ["title", "rows"]}
  }
}]

def create_spreadsheet(title, rows):
    r = requests.post("https://openofficeai.com/api/v1/sheets",
        headers={"Authorization": "Bearer YOUR_API_KEY"},
        json={"title": title, "sheets": [{"rows": rows}]})
    return r.json()["url"]

Full walkthrough: OpenAI function calling integration.

Claude tool use

Claude proposes a tool_use block; you execute it and return a tool_result with the link.

Python (Anthropic)

tools = [{
  "name": "create_spreadsheet",
  "description": "Create a shareable spreadsheet and return its URL.",
  "input_schema": {"type": "object", "properties": {
    "title": {"type": "string"},
    "rows": {"type": "array", "items": {"type": "array"}}},
    "required": ["title", "rows"]}
}]
# On tool_use, POST to /api/v1/sheets and return the url as tool_result.

Full walkthrough: Claude tool use / MCP integration.

LangChain & CrewAI

Framework agents wrap the same call in their tool decorator.

LangChain (Python)

from langchain_core.tools import tool

@tool
def create_spreadsheet(title: str, rows: list) -> str:
    """Create a shareable spreadsheet; first row is the header. Returns a URL."""
    r = requests.post("https://openofficeai.com/api/v1/sheets",
        headers={"Authorization": "Bearer YOUR_API_KEY"},
        json={"title": title, "sheets": [{"rows": rows}]})
    return r.json()["url"]

Same idea in LangChain, CrewAI, LlamaIndex, PydanticAI, and the Vercel AI SDK — see the full integrations directory for copy-paste code per framework.

Four practices that make agent tool calls reliable

Keep the signature flat. A title + rows (array of arrays) shape calls reliably; deeply nested schemas cause malformed calls.
Write the description for the model. State that the first row is the header and the tool returns a URL — the model uses this to format data correctly.
Keep the API key server-side. The tool executes in your code, never inside the model, so the key stays in an environment variable.
Return just the URL. One short string is easy for the model to hand back; avoid returning large payloads it then re-summarizes.

Documents, not just spreadsheets

The same pattern produces formatted documents: add a second tool that posts to /api/v1/docs with a content array of headings and paragraphs, and your agent can return reports, memos, and proposals as shareable links — downloadable as PDF or DOCX. See the API docs for the full schema.

FAQ

Why not just have the agent write a CSV?

A CSV string still has to be saved, opened, and shared by the user, and large tables burn tokens and risk truncation. A document API returns a finished, hosted, editable artifact in one call.

Does this work with local models?

Yes — any tool-capable model works, including local ones via Ollama. The document call is the only thing that leaves your machine.

How do I start?

Grab a free API key (500 calls/month), pick your framework in the integrations directory, and paste in the tool.

Give your agent document output

Free tier: 500 API calls/month, no card required.

Get your API key Browse integrations