Connect Google Sheets to any MCP client

What is a Google Sheets MCP server?

A Google Sheets MCP server exposes a spreadsheet to an AI client as a set of callable tools, so Claude, Cursor, or ChatGPT can list its tabs, inspect its columns, and query its rows on demand. PasteSheet hosts one for any sheet you paste — no Google Cloud project, no service account, and nothing to deploy.

MCP (the Model Context Protocol) is the open standard AI clients use to talk to external data and tools. An MCP server advertises a set of tools; the AI decides when to call them and how to use the results. A "Google Sheets MCP server" is simply a server that does this for your spreadsheet.

PasteSheet builds that server for you. For a link-shared sheet, paste its URL. For a restricted sheet, connect your Google account and choose the exact file in Google Picker. Either way you get a hosted MCP endpoint that any compatible client can add as a custom connector, with no Google Cloud project or server to run. (Google's own official and Workspace MCP servers do not cover Sheets as a hosted service — see how they compare.)

The tools your agent gets

Once connected, the AI can call three read-only tools against your sheet:

• list_tabs — discover the tabs in the sheet and the default one.
• get_schema — see each column's name, alias, and type before querying.
• query_rows — read rows with exact-match filters, partial (contains) matches, full-text search, sorting, and pagination.

Connect a single sheet

Each endpoint has its own MCP URL. Open the endpoint's Connect via MCP panel in PasteSheet, copy the URL, and add it as a custom connector. For a Desktop-style client the config looks like this:

{
  "mcpServers": {
    "pastesheet": {
      "url": "https://pastesheet.com/mcp/sheets/your-endpoint-id"
    }
  }
}

Connect from your client

The same endpoint works in every MCP-compatible client — pick yours for exact steps:

• Claude Desktop and Claude web — add the endpoint URL as a custom connector.
• Claude Code — one claude mcp add command in the terminal.
• Cursor — add it to the MCP settings.
• ChatGPT — connect the workspace server with OAuth.
• VS Code and Windsurf — paste the URL into the MCP config.
• Manus AI — add it as a custom MCP connector.
• n8n — use the MCP Client Tool node in a workflow.

Public vs private endpoints

Public endpoints are keyless — the MCP URL just works. For a private endpoint, append ?api_key=YOUR_KEY to the URL for web clients, or send an Authorization: Bearer YOUR_KEY header from CLI clients like Cursor and Claude Code.

Want one connector that exposes every sheet you own — or need to use ChatGPT? Connect the workspace server at https://pastesheet.com/mcp with an account key or OAuth. See the per-client guides below for exact steps.

Why does an agent need a cached endpoint?

Because agents are chatty. An agent exploring a question rarely reads a sheet once — it lists tabs, checks the schema, then fires several filtered queries in a row. Hitting Google directly for each of those burns quota fast.

Google's Sheets API allows 300 read requests per minute per project and only 60 per minute per user, and returns a 429 once you cross it (Google's published usage limits). PasteSheet reads your sheet once, caches the rows, and serves every subsequent tool call from that cache — so an agent's tenth query costs Google nothing. The cache window is yours to set, from 30 seconds up to an hour.

What a tool call actually looks like

You never write this by hand — the agent composes it — but seeing the shape explains why the connection is reliable. Ask "show me active enterprise customers, newest first" and the client emits roughly:

{
  "name": "query_rows",
  "arguments": {
    "tab": "Customers",
    "filters": { "status": "active", "plan": "enterprise" },
    "sort": "signed_up",
    "order": "desc",
    "limit": 20
  }
}

Why this beats pasting the sheet into the chat

The filtering happens in the endpoint, not in the model. That distinction matters more than it sounds. When you paste a spreadsheet into a conversation, the model has to hold every row in context and reason over all of them — which is expensive, degrades as the sheet grows, and quietly goes wrong on arithmetic across hundreds of rows.

With a query tool, twenty matching rows come back and the model reasons over twenty rows. A 50,000-row sheet answers just as accurately as a 50-row one, because the sheet size never enters the context window. And the data is current — a pasted copy is stale the moment somebody edits a cell.

Questions worth asking once it is connected

These all work without you naming a column or a tab — the agent reads the schema first:

• "What tabs are in this sheet, and what does each one hold?" — orientation, useful on a sheet you did not build.
• "Which rows have a blank email?" — the kind of data-quality sweep nobody does manually.
• "Total revenue by region for last quarter." — on Pro, this becomes one group_by + sum call rather than the agent adding numbers itself.
• "Find anything mentioning 'refund' anywhere in the sheet." — full-text search across every column.
• "Compare this month's signups to last month's and tell me what changed." — two queries and a comparison, composed by the agent.

What the agent can and cannot reach

The security model is deliberately narrow. The tools are read-only — there is no write tool to call, so no prompt, however confused or adversarial, can make the agent modify or delete your data. The worst case is a query returning rows you did not expect, not a mutated spreadsheet.

Scope is the other half. A per-sheet connector reaches exactly one endpoint; the workspace server reaches the endpoints on your account and nothing else. A restricted source can be connected through Google Picker, which grants PasteSheet file-scoped access to the exact sheet you choose. The app only reads that sheet and exposes no write path.

Keep two ideas separate: a private endpoint is not the same as a private sheet. The source can stay restricted in Google while the endpoint is public or protected with a bearer key. If a restricted source powers a public endpoint, the endpoint data is public to anyone holding its URL, so choose that setting deliberately.

Ways to get Google Sheets into MCP

There are four common ways to get a spreadsheet into an MCP client, and they are not equivalent — see the deep dives on Google's official MCP, a self-hosted server, and Zapier and Composio:

Why the schema step matters

Spreadsheets are untyped, and that is normally the thing that breaks automated access. A column headed Date might hold 2026-01-04, Jan 4, and an empty string in three adjacent rows, and a naive reader treats all three as text.

When you connect a sheet, PasteSheet infers each column's type and a clean alias, and get_schema hands that to the agent before it queries. So the agent knows signed_up is a date and can sort by it, rather than guessing from the header string. It also means you can rename a column for the API without renaming it in the spreadsheet, which keeps a messy human-facing header from leaking into every query.

If the sheet's columns later change shape, schema drift alerts tell you before an agent starts returning confidently wrong answers.

Open source, self-hosted, and in the MCP Registry

The server's docs, client config, and registry manifest are open on GitHub, and it is listed in the official MCP Registry as com.pastesheet/google-sheets — so any client that browses the registry can discover it.

Prefer to run your own? The hosted endpoint is the fastest path, but you have options: see the open-source Google Sheets MCP landscape, or self-host a Sheets MCP server on your own infrastructure and weigh the trade-offs against a managed endpoint.

What it costs

MCP is included on the Free plan. Connect any public endpoint over MCP with no Google Cloud project and no credit card. Private endpoints and the account-wide workspace server need a paid plan (from Starter ($9/mo)) for the keys and OAuth they authenticate with.

Pro adds full-text search and aggregation (count, sum, avg, group_by) that your AI agent can call through query_rows.

Frequently asked questions

Sources

• Model Context Protocol specification — Model Context Protocol
• Usage limits — Google Sheets API — Google

Related guides