Google Docs MCP

google-docs-mcp

A read-only Model Context Protocol server for Google Docs. Its job is to explain a document's comments to an AI assistant (Claude or any MCP client) in full detail: who wrote each one, when, on which piece of text, whether it is open or resolved, the complete reply thread, and who resolved it.

• Read-only by design. It only requests the documents.readonly and drive.readonly scopes. There is no tool that creates, edits, resolves or deletes anything, and every tool is advertised with readOnlyHint: true.
• OAuth 2.0 sign-in. Users authenticate with their own Google account through the desktop-app flow with PKCE. Tokens are stored locally and refreshed automatically (see the seven-day exception for External apps in Testing mode).
• Zero configuration for the assistant. Tools accept a document ID *or* a full Google Docs URL.

Built by José Antonio Garrido. I write about the developer tools I build on my blog, josegarrido.dev.

Tools

Example: list_comments

# Q4 project plan

• Document ID: 1AbC...
• Link: https://docs.google.com/document/d/1AbC.../edit
• Owner: Ana
• Last modified: 2026-09-22 08:00 UTC by Jose (you)

Comments (all): 2

[1] 🟢 OPEN — Ana · 2026-09-22 08:00 UTC

Quoted text (what the comment refers to):

delivery date

Surrounding paragraph (#7):

The delivery date will be shared with the client once the scope is final.

Comment: The delivery date is missing.

_No replies yet._

Comment ID: AAAAxyz

[2] ✅ RESOLVED — Ana · 2026-09-20 10:32 UTC

... Thread (2 replies):

• Jose (you) · 2026-09-21 09:00 UTC: Yes, moving it.
• Jose (you) · 2026-09-21 12:00 UTC — marked as RESOLVED

Resolved by Jose (you) on 2026-09-21 12:00 UTC.

Comment and reply authors appear by display name: the Drive API doesn't populate their email address. A document's owner and last editor show an email when Google makes it visible to you.

Quick start

Prerequisites

• Node.js >= 20
• A Google Cloud project with the Google Drive API and Google Docs API enabled

1. Create an OAuth client in Google Cloud

1. Open the Google Cloud Console and create or select a project.
2. APIs & Services → Library: enable *Google Drive API* and *Google Docs API*.
3. APIs & Services → OAuth consent screen: configure it. Use *Internal* for a Workspace organization; for *External* apps in testing mode, add your account as a test user.
4. APIs & Services → Credentials → Create credentials → OAuth client ID, application type Desktop app.
5. Download the JSON and save it as ~/.config/google-docs-mcp/credentials.json.

External apps in Testing mode have to sign in again every seven days. With an *External* consent screen whose

publishing status is *Testing*, Google

issues a refresh token that expires in 7 days,

so automatic renewal stops working and tools report that the sign-in expired until you repeat step 2. The running

server picks up the new sign-in without a restart. An *Internal* consent screen (Workspace organizations only) doesn't

have this limit. Neither does publishing the app (*In production*), but drive.readonly is a restricted scope, so

Google then

shows an unverified-app warning and caps the project at 100 users

unless the app is verified.

You can use environment variables instead of the file:

export GOOGLE_CLIENT_ID="...apps.googleusercontent.com"

export GOOGLE_CLIENT_SECRET="..."

2. Sign in with Google

npx -y -p @yeyo11/google-docs-mcp google-docs-mcp-auth

Your browser opens (or the URL is printed so you can paste it). Google asks for read-only access to Drive and Docs. Tokens are saved to ~/.config/google-docs-mcp/token.json with 0600 permissions. Run the same command again to switch accounts.

3. Register the server with your MCP client

No clone or build step is needed; npx fetches the package on first use.

Claude Desktop (claude_desktop_config.json):

{

"mcpServers": { "google-docs": { "command": "npx", "args": ["-y", "@yeyo11/google-docs-mcp"] } } }

Claude Code:

claude mcp add google-docs -- npx -y @yeyo11/google-docs-mcp

Any other MCP client that supports stdio servers works the same way. If you use environment variables for the OAuth client, add them to the env block of your client configuration.

Running from source instead

git clone https://github.com/yeyo11/google-docs-mcp.git

cd google-docs-mcp npm install npm run build npm run auth

Then point your MCP client at node /absolute/path/to/google-docs-mcp/dist/index.js.

4. Ask away

"Summarize the open comments in https://docs.google.com/document/d/…/edit and tell me which ones are waiting on me."

Configuration

How it works

MCP client ──stdio──▶ google-docs-mcp ──OAuth2 (read-only)──▶ Google Drive API (files, comments)

└▶ Google Docs API (document text)

• Comments come from the Drive API (comments.list / comments.get) with every field requested explicitly, so nothing is silently dropped: quoted text, replies, resolve/reopen actions, deleted flags.
• The "surrounding paragraph" is found by locating the quoted text inside the document body fetched from the Docs API (all tabs are traversed, including tables).
• Google API errors (401/403/404/429) are returned as MCP isError results with an actionable message instead of breaking the session. Logs go to stderr; stdout is reserved for the protocol.

Development

npm run dev        # run the server from src/ with tsx

npm run typecheck # type-check without emitting npm test # smoke test: tool listing, error handling without credentials, comment formatting

See CONTRIBUTING.md for the workflow and commit conventions.

Security

• Only read-only scopes are requested; Google never grants this client write access.
• If the stored token lacks the expected scopes, the server refuses to use it and asks you to sign in again.
• credentials.json and token.json are git-ignored. Never commit them.
• To revoke access, remove the app at https://myaccount.google.com/permissions and delete token.json.

To report a vulnerability, see SECURITY.md.

Limits

• Read-only. It can't create, reply to or resolve comments, or edit documents.
• Google Docs only. Sheets, Slides and other file types are refused, by URL and by bare ID.
• Suggestions, assignments and mentions aren't read as such. A comment is its text, status and thread: an assigned comment looks like any other and a mention is just part of the text. Depending on your access, text pending as a suggestion can appear in get_document as ordinary text.
• Comment authors are told apart by display name, since Drive doesn't return their email. In get_comments_summary, two people with the same name count as one.
• `get_document` returns plain body text from all tabs, tables included: no formatting, images, smart chips, headers, footers or footnotes. 20,000 characters by default, 500,000 at most (max_chars).
• The surrounding paragraph is found by searching for the quoted text. It takes the first match, so a repeated phrase can point to the wrong paragraph, and it may be missed if the text was edited after the comment. Quotes under 8 characters get no paragraph, and paragraphs over 600 characters are trimmed.
• `list_comments` pages by 100. Use page_token for the rest. The status filter applies to each page, so a page can come back short or empty while more remain. get_comments_summary reads every page.
• One Google account per installation. Run the sign-in command again to switch.
• Local stdio only. There is no remote or HTTP transport.
• Tokens are stored unencrypted in token.json, with 0600 permissions.
• `drive.readonly` covers all of Drive the account can see, not only Docs. The tools read Docs only, but that is what you grant on Google's sign-in screen and what the stored token allows.
• What a tool returns reaches the assistant. Document text, comments, names and email addresses go to whichever assistant and model provider you use. Text written by other people arrives as is, so treat it as untrusted.

Roadmap

• Optional write tools (reply, resolve) behind an explicit opt-in scope
• Multiple accounts / profiles
• Streamable HTTP transport for remote deployments

License

MIT