Give your AI agent a Google Sheet to work from

A spreadsheet your agent can actually query

Agents work best against structured, up-to-date facts: a product list, a policy table, a roster, a pricing grid. A Google Sheet is where teams already keep that data. PasteSheet exposes it as a Model Context Protocol server, so the agent gets real tools instead of a pasted screenshot that goes stale the moment someone edits a cell.

Crucially, this is publish, not write-back: public sheets can be published from a URL, while restricted sheets can be selected through a file-scoped Google account connection. Either way, PasteSheet exposes one read-only source that any agent — yours or a teammate's — can consume, with no write tool and no Google Cloud project for you to configure.

The tools your agent gets

Connect the endpoint and the agent can call three read-only tools:

• list_tabs — discover the tabs in the sheet and the default one.
• get_schema — read each column's name, alias, and type before querying.
• query_rows — filter, partial-match, full-text search, sort, and paginate rows.

Connect a single sheet

Each endpoint has its own MCP URL. Copy it from the Connect via MCP panel and add it as a custom connector:

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

Querying beats stuffing the context window

The instinct with agent data is to load everything up front — dump the sheet into the system prompt and let the model sort it out. That works until it does not, and the failure is quiet: as rows accumulate, retrieval accuracy drops, arithmetic across hundreds of rows starts drifting, and token cost scales with sheet size on every single turn.

A query tool inverts that. The agent asks for what it needs, gets back the matching rows, and reasons over a small set. A 50,000-row sheet costs the same per turn as a 50-row one, because the sheet never enters the context window — only the answer does.

It also fixes staleness. A pasted table is a snapshot; a tool call reads the sheet as it is now, so an edit made two minutes ago is already reflected.

Designing a sheet an agent can use well

Agents read spreadsheets more literally than people do. A few habits make a large difference:

• One header row, plain text. Merged cells and decorative title rows above the headers break column detection.
• Name columns like fields, not labels. status beats Current Status ✅. Column aliases let you keep the pretty header for humans and a clean name for the agent.
• One fact per cell. A cell holding "active, enterprise, EU" cannot be filtered on. Three columns can.
• Keep types consistent down a column. A stray N/A in a number column is what turns a confident answer into a wrong one.
• Split unrelated data into tabs. The agent calls list_tabs first, so clear tab names are free navigation.
• Delete trailing empty rows. Otherwise counts and aggregates quietly include them.

Where the boundaries are

Two limits worth knowing before you build on this. First, the connection is read-only by construction — there is no write tool to call — so an agent cannot corrupt the source no matter what it is asked to do. That is the point, but it does mean write-back workflows need something else.

Second, keep source privacy and endpoint privacy separate. A restricted Google Sheet can be selected through Google Picker and read by PasteSheet without changing its Google sharing setting. You can then keep the endpoint behind a bearer key, or deliberately make the returned data public. A public endpoint is still public even when the source sheet remains restricted.

Read-only, so it is safe to hand out

Because the endpoint is read-only by design, it is safe to expose publicly or hand to a teammate: consumers can read and query the data but can never change the sheet. And there is no Google Cloud project, OAuth screen, or service account to set up — you paste a share URL and get a live API.

What it costs

Frequently asked questions

Sources

• Model Context Protocol specification — Model Context Protocol
• Configure the Google Workspace MCP servers — Google
• google/mcp — Google's open-source MCP servers — Google
• Installing and using an MCP server — Anthropic

Related use cases