# Connect Google Sheets to Claude (private sheets, exact totals)

> Add a Google Sheet to Claude as a custom connector so it queries the rows instead of reading them. Works with private sheets and large tabs.

Source: https://datatomcp.com/guides/connect-google-sheets-to-claude · By Killian Helf · Updated 2026-10-19

Claude can read a spreadsheet you paste or attach, but it reads it as text. With a few thousand rows that means long context, slow answers and totals that are estimated rather than computed. This guide connects a Google Sheet to Claude as a **custom connector**, so Claude sends a query and gets back an exact result.

It works in Claude.ai, Claude Desktop and Claude Code, with private sheets as well as public ones.

## What you need

- A Google Sheet (any size up to 100,000 rows per tab on the free plan).
- A datatomcp account. Signing in with Google is simplest, because the same account later picks the sheet.
- A Claude plan that allows custom connectors.

## 1. Import the sheet

1. Sign in at [app.datatomcp.com](https://app.datatomcp.com/login) and open **Add data → Google Sheet**.
2. Click **Choose a Google Sheet**. Google's own file picker opens. Choose the spreadsheet; datatomcp gets access to **that file only**, not the rest of your Drive (Google's `drive.file` permission).
3. datatomcp reads the first tab with unformatted values, so `€1,200.50` arrives as the number `1200.5`, and builds the dataset.

You land on the **import report**: every column with its detected type, example values and warnings such as "number stored as text" or "mixed date formats". Nothing is converted silently; fixes are one click and create a new version.

A public sheet works too: paste its share link under **Paste a link** instead.

## 2. Add the connector to Claude

1. In Claude, open **Settings → Connectors → Add custom connector**.
2. Name it `datatomcp` and paste the URL `https://app.datatomcp.com/mcp`.
3. Click **Connect**. A datatomcp sign-in opens, then a consent screen listing what Claude may do (read and query your datasets). Allow it.

There is no API key to copy. Claude holds a token that you can revoke from datatomcp at any time.

**Claude Code** takes one command:

```bash
claude mcp add --transport http datatomcp https://app.datatomcp.com/mcp
```

Then run `/mcp` inside Claude Code, pick **datatomcp** and authenticate.

## 3. Ask a question

Start a chat and ask something your sheet can answer. With our 50,000-row test orders sheet:

> How many orders came from France in September 2026, and what was the revenue?

Claude calls `describe_dataset` to learn the columns, then one `aggregate` call:

```json
{
  "where": [
    { "column": "country", "op": "eq", "value": "France" },
    { "column": "order_date", "op": "between", "value": ["2026-09-01", "2026-09-30"] }
  ],
  "metrics": [
    { "fn": "count", "as": "orders" },
    { "fn": "sum", "column": "total", "as": "revenue" }
  ]
}
```

and answers **277 orders, €50,279.50**. The tool result is 32 bytes. Pasting the same sheet would have sent about 4 MB of text.

## Why this beats pasting the sheet

- **Exact numbers.** Counts and sums are computed by a database (DuckDB), not estimated by the model.
- **Big tabs work.** Claude only receives the rows or totals it asks for, so 100,000 rows cost the same as 100.
- **Real filter values.** Claude can call `distinct_values` first, so it filters on `"United States"` rather than guessing `"USA"`.
- **Every answer is traceable.** Each result says which version of the dataset it came from and when it was refreshed.

## Keeping it in sync

When the sheet changes, click **Refresh now** on the dataset page. On Pro, datatomcp re-reads the sheet every hour and only publishes a new version when something changed. If a refresh fails (the sheet was deleted, access was removed), agents keep getting the last good version and you get one email.

## Troubleshooting

- **"datatomcp can't open this sheet"** right after picking it: Google can take a few seconds to grant access. Try once more.
- **Claude doesn't use the tools:** mention the data explicitly ("use datatomcp to…") the first time, or check the connector is enabled for the chat.
- **Wrong column type** (a code like `00123` read as a number): change the type on the dataset page; the dataset is rebuilt as a new version.