Gmail MCP Server: What It Does and Whether It's Worth It
Google's Gmail MCP server lets AI search, read, label and draft email. How it compares with Claude's Gmail connector.

Gmail MCP is any server built on the Model Context Protocol (MCP) that lets an AI assistant like Claude or Cursor search emails in your inbox, read threads and draft emails for you. Google now ships an official one, and it deliberately stops at drafts: you send from Gmail yourself.
Below: what Google's server does, how to set it up, how it compares with Claude's Gmail connector and community servers, and the security risks of letting an AI read your mail.
Is there an official Gmail MCP server?
Yes. Google runs a remote Gmail MCP server at https://gmailmcp.googleapis.com/mcp/v1, over Streamable HTTP with OAuth 2.0, available to members of the Google Workspace Developer Preview Program (Google for Developers). It asks for two scopes, gmail.readonly and gmail.compose, and only reaches what the signed-in account can already see.
There are three ways to give an AI your Gmail today:
- Google's Gmail MCP server: official, narrow, still in preview.
- Claude's built-in Gmail connector: run by Anthropic, one toggle, Claude only.
- Community servers: open source, more tools, and you maintain them.
Google's Calendar server follows the same model, and the full lineup of vendor servers is in Best MCP Servers for Work.
What can the Gmail MCP server do?
Google's tool reference lists nine available tools, covering search, reading, drafts and label management:
Read tools
search_threads: search your mail with Gmail's search syntaxget_thread: read a full conversationlist_draftsandlist_labels
Write tools
create_draft: write a draft for you to review and send from Gmaillabel_message,label_thread,unlabel_messageandunlabel_thread
There's no send tool and no delete tool. That's a design choice: sending is irreversible and public, so Google keeps a person on that step. Google doesn't publish separate rate limits for the MCP server.
How do you set up the Gmail MCP server?
You build your own OAuth client in Google Cloud, then hand its credentials to your AI client. Google documents these steps:
- Join the Developer Preview. Your account needs membership in the Google Workspace Developer Preview Program.
- Create a Google Cloud project and install the gcloud CLI.
- Enable two APIs: the Gmail API (
gmail.googleapis.com) and the Gmail MCP API (gmailmcp.googleapis.com). - Configure the OAuth consent screen in the Google Cloud console. Set the app name to Gmail MCP Server, choose Internal if your Workspace allows it, otherwise External with test users, and add the two Gmail scopes.
- Create an OAuth client ID with the application type Web application, add your MCP client's redirect URI, and tick Use this client for an AI-powered agent. Without that box, requests get rejected.
- Add it to your client.
- Claude and Claude Desktop (Pro, Max, Team or Enterprise): add
https://claude.ai/api/mcp/auth_callbackas the redirect URI, then go to Settings > Connectors, click Add custom connector, name it Gmail, paste the server URL, and enter the client ID and client secret under Advanced settings. - Google Antigravity: add a
gmailentry undermcpServersin~/.gemini/config/mcp_config.jsonwith the URL and OAuth credentials. - Other MCP clients: server name
gmail, the URL above, Streamable HTTP transport and OAuth 2.0 authentication.
- Claude and Claude Desktop (Pro, Max, Team or Enterprise): add
If you get asked to sign in about every hour, the client isn't receiving a refresh token; Google's fix is adding access_type=offline and prompt=consent to the authorization request. Google's page doesn't cover Cursor or ChatGPT, so treat those as the generic "other clients" case.
How do you add Gmail MCP to Claude Code?
Add it in claude.ai, not in the terminal. If you're logged in to Claude Code with a claude.ai account, connectors you add at claude.ai show up in /mcp automatically. Anthropic's own Gmail connector doesn't support local OAuth from Claude Code at all, because Google only accepts the redirect URL claude.ai registered (Claude Code docs).
Gmail MCP server vs Claude's Gmail connector: which should you use?
In Claude, the built-in connector is the easier pick; Google's server matters when you want a Google-run endpoint any MCP client can use. Claude's connector works on every plan including Free, searches and reads mail, lists and creates drafts, and manages labels and threads. Since August 2026 (9to5Google) it can also send, reply and forward, asking for approval before each by default (Claude Help Center).
| Google's Gmail MCP server | Claude's Gmail connector | Community servers | |
|---|---|---|---|
| Made by | Anthropic | Independent developers | |
| Setup | Cloud project, two APIs, your own OAuth client | Sign in with Google, toggle it on | Local install, your own Google OAuth credentials |
| Who can use it | Developer Preview members | All Claude plans; owners enable it on Team and Enterprise | Anyone who runs it |
| Search and read | Yes | Yes (attachment metadata, not content) | Yes, often with attachments |
| Drafts and labels | Yes | Yes | Usually |
| Send | No | Yes, with approval by default | Often, plus delete |
| Clients | Claude, Antigravity, other Streamable HTTP clients | Claude, Claude Desktop, Claude Code via claude.ai | Mostly local clients |
On Team and Enterprise, an owner turns the connector on under Organization settings > Connectors, and a Google Workspace admin may need to mark Claude as a trusted app.
What about community Gmail MCP servers?
They trade guardrails for features, and maintenance is on you. The widely used GongRzhe/Gmail-MCP-Server offered send_email, delete_email and batch operations such as batch delete, and its owner archived the repository on 3 March 2026, so it's now read-only (GitHub). Before installing any community server, check that it's still maintained and which scopes it asks for.
Is Gmail MCP safe?
It's as safe as the scopes you grant, and email is untrusted input by definition: anyone can send you a message. Google's own setup guide warns about indirect prompt injection and recommends screening content with Model Armor.
The attack is simple. Someone emails you hidden instructions, your AI reads the thread during a task, and the email tells it to forward your last ten threads somewhere. Models resist this kind of planted instruction poorly: in the MCPTox benchmark of 45 real MCP servers and 353 tools, poisoned tool descriptions got o1-mini to comply 72.8% of the time, and no model refused more than 3% of attacks (Wang et al., 2025).
- No per-message permissions. Every tool call can access data across the account: contracts, invoices and password-reset links included.
- Keep sending gated. Google's server can't send; with Claude's connector, leave per-send approval on.
- Separate readers from senders. Don't run a server that reads outside mail next to one that can write or send somewhere private in the same session.
More in Is MCP Safe?
Is Gmail MCP worth it?
Yes, for live inbox work: finding a thread, summarizing a conversation, drafting a reply in context. If you use Claude, start with the built-in connector; set up Google's server if you need it in another client or want Google's narrower, draft-only boundary.
What it won't do is understand your work. Ask "where did we land with the Acme contract" and it retrieves every thread that mentions Acme. It doesn't know which reply superseded the last one, or that the deal moved to Slack two weeks ago, so the model rebuilds your situation from raw mail every session. That's the gap between a connector and a context layer, explained in Connectors vs MCP vs a Context Layer.
Where does Unabyss fit next to Gmail MCP?
Unabyss does the understanding part. Gmail is one of its live integrations, connected with the read-only gmail.readonly scope: it reads your messages and threads and turns them into personal context alongside Slack, Notion, Calendar and your other apps. It serves that context to Claude, ChatGPT, Cursor and other MCP clients from one connection, and it never drafts or sends email. In an August 2026 Unabyss benchmark against a baseline with the same tools connected, adding Unabyss cut cost by 36%, time by 38% and follow-up prompts by 60% (What Structured AI Context Is Worth).
Many people run both: a Gmail connector to pull an exact thread or draft a reply, and Unabyss so every AI tool already knows who Acme is and where the contract stands. See how it works.