Skip to content
Call wiki

What's the best way to feed a spreadsheet to an AI agent?

Last updated 2026-08-17 · Confidence: documented — Anthropic’s help center and docs plus its public xlsx skill; the pasted-link failure is tried, the one-export-per-tab scope inferred.

Download the sheet as CSV and hand the agent the text file — that works on every surface. Pasted links and the connector are patchier than they look, and raw .xlsx pays off only where the agent can run code.

Section titled “Pasted links: not a load path for Sheets”

The advice that a pasted Drive link is “automatically loaded in some form” failed a live test — a Sheets link pasted into Claude chat loaded nothing (tried — no Drive connector was active). The documented paste-a-URL flow sits inside the “Add from Google Drive” picker, and the Drive integration page marks Google Sheets “Not currently supported” there — that route live-syncs Google Docs only. A Sheets URL in the message body is just text Claude can’t open on its own.

The connector reads Sheets — as tool calls

Section titled “The connector reads Sheets — as tool calls”

The Google Workspace connectors article lists “Read Sheets, Slides, PDFs, images, and MS Office files” among the Drive connector’s capabilities. So with the connector on, asking outright — “read this tab through the Drive connector” — works: a per-request tool call, text extraction only (see Doc freshness for the mechanics). The two Anthropic pages disagree on Sheets because they describe different routes: connector reads versus Docs-only synced knowledge.

CSV: text in, text out

Section titled “CSV: text in, text out”

File → Download → CSV yields a plain text file — one table of comma-separated records — and claude.ai accepts CSV uploads directly. One table per file means one export per tab of a multi-tab workbook. Same principle as Markdown formats: agents do best on formats they can read as text.

xlsx: opaque in chat, fine for coding agents

Section titled “xlsx: opaque in chat, fine for coding agents”

An .xlsx is a zip archive of XML parts, not readable text, and claude.ai accepts .xlsx uploads only with code execution and file creation enabled. Agents with a shell simply script over it: Anthropic’s public xlsx skill reads and edits workbooks with pandas and openpyxl, recalculating formulas via LibreOffice. “xlsx is much harder” holds for plain chat, not for coding agents. Writing results back to a live Sheet is its own problem — see Docs write path.

Sources

Section titled “Sources”