Gmail MCP Server Review 2026: Email Automation for AI Agents

What is the Gmail MCP Server?

If you have ever caught yourself asking Claude to "reply to that investor from yesterday" and felt the pain of the model saying it cannot actually read your email — you are the target audience for this review. The Gmail MCP server closes that gap. It is an open-source Model Context Protocol server that gives your AI agent six first-class tools for operating Gmail: send, read, search, labels, drafts and attachments. Behind the scenes it wraps the official Gmail API. On the surface, your agent just… uses your inbox.

The community standard in 2026 is GongRzhe/Gmail-MCP-Server, which crossed 42,000+ installs on Smithery in April and remains the most-installed email MCP of any kind. It ships with auto-OAuth, sensible default scopes, structured JSON responses tuned for AI context windows, and one-line install across every major MCP client.

Since Anthropic donated MCP to the Linux Foundation in December 2025, adoption has accelerated fast. OpenAI, Google, Cursor, Windsurf, and Claude Desktop all speak the same protocol now. The Gmail MCP server — originally a weekend project — is quietly the single most-deployed piece of that ecosystem.

Key Features

1. Send Email — with attachments and HTML bodies

The send_email tool takes a to/cc/bcc, subject, body (plain or HTML), and optional attachments array. Agents can construct and send real, RFC-822 compliant messages from your account. In hands-on testing the whole round-trip — Claude reasoning → tool call → Gmail API → message in recipient's inbox — averages under 1.8 seconds.

2. Read Inbox — threads, bodies, headers

The read_email tool returns a structured JSON view of any message: sender, recipients, subject, date, label list, and the full decoded body. The server strips quoted replies by default so AI agents don't get context-poisoned by the same conversation echoing down a thread.

3. Search — with full Gmail operators

The search_emails tool accepts the full Gmail query syntax: from:, to:, label:, has:attachment, before:, after:, is:unread, subject:. In practice this means an agent can write from:investors@example.com has:attachment newer_than:7d and get back only the right threads. No more dumping 10,000 tokens of inbox into context.

4. Labels — create, apply, remove

The modify_email tool adds or removes any Gmail label, including system labels (INBOX, STARRED, UNREAD) and user labels. This is the trick for automation: label a thread AI/Needs-Reply and your agent can pick it up on the next tick.

5. Drafts — save, edit, send later

The draft_email tool is the safety valve I use everywhere. Instead of letting an AI send directly, I have it draft the reply, label the thread AI/Review, and leave the actual "send" button for a human. This pattern single-handedly makes email automation safe.

6. Attachments — inline download and upload

Attachments come back as base64 blobs with filename and MIME type metadata. The agent can save them locally, pass them through OCR, or re-upload into a reply. Upload path works the same way in reverse.

How to Install and Configure

The full setup — from zero to a working Claude Desktop that can send email — takes about five minutes. Here's the exact sequence I used.

Step 1 — Enable the Gmail API in Google Cloud Console

Create a new project (or reuse one), open APIs & Services → Library, search "Gmail API", and click Enable. Then go to Credentials → Create Credentials → OAuth client ID, pick Desktop app, and download the JSON.

Step 2 — Stage the OAuth credentials

# macOS / Linux
mkdir -p ~/.gmail-mcp
mv ~/Downloads/client_secret_*.json ~/.gmail-mcp/gcp-oauth.keys.json

# Windows PowerShell
New-Item -ItemType Directory -Path "$env:USERPROFILE\.gmail-mcp" -Force
Move-Item "$env:USERPROFILE\Downloads\client_secret_*.json" `
          "$env:USERPROFILE\.gmail-mcp\gcp-oauth.keys.json"

Step 3 — Run the auto-OAuth flow

npx @gongrzhe/server-gmail-autoauth-mcp auth

A browser window opens. Pick your Gmail account, click Continue → Allow, and the command writes credentials.json next to your keys file. The refresh token inside is what the MCP server uses from then on — you will never see the OAuth screen again on this machine.

Step 4 — Register with your MCP client

For Claude Desktop, open ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows) and add:

{
  "mcpServers": {
    "gmail": {
      "command": "npx",
      "args": ["@gongrzhe/server-gmail-autoauth-mcp"]
    }
  }
}

For Claude Code, it is a one-liner:

claude mcp add gmail npx @gongrzhe/server-gmail-autoauth-mcp

Restart the client, and you will see send_email, read_email, search_emails, modify_email and draft_email show up in the tool list. Ask it to "search my Gmail for unread emails from this week, labeled Clients" and you are live.

Quick install via Smithery

If you prefer a managed install, Smithery handles everything in one command:

npx -y @smithery/cli install @gongrzhe/server-gmail-autoauth-mcp --client claude

Pricing

The Gmail MCP server itself is free forever. MIT license, self-hosted, no account required beyond your existing Gmail. The only costs sit underneath the integration:

• Google Cloud OAuth client — free for standard personal use. Workspace admins can issue one centrally.
• Gmail API quota — 1 billion quota units per day, with 250 units per user per second. A typical agent uses 5–50 units per action. You will not hit this limit in normal use.
• Your AI provider — Claude, GPT, Gemini tokens. Same as any other MCP server.

For 99% of users, total incremental cost of adding the Gmail MCP server is $0/month.

Pros and Cons

Alternatives Compared

Outlook MCP

The Microsoft 365 equivalent. Same six-tool shape (send, read, search, labels/folders, drafts, attachments), uses Microsoft Graph under the hood instead of Gmail API. Setup is slightly longer because Azure AD app registration involves more steps than a Google Cloud OAuth client. About 11,000 Smithery installs as of April 2026 — strong, but a third of Gmail MCP. Pick Outlook MCP if your mailbox is Outlook; there is no reason to run both.

SendGrid MCP

A different use case entirely. SendGrid MCP is about outbound transactional email — product emails, marketing blasts, API-triggered notifications. It does not read your inbox. If your agent needs to send a thousand personalised onboarding emails, SendGrid MCP is correct. If your agent needs to operate a human inbox, Gmail MCP is correct.

Resend MCP

Same category as SendGrid — transactional-only. Resend is the developer-favourite 2026 successor (cleaner API, React Email integration, generous free tier). Great for sending. Useless for reading or searching an inbox. Use it alongside Gmail MCP if your agent does both developer-ops email and human-ops email.

FAQ

Final Verdict — 4.5 / 5

The Gmail MCP server is one of those small pieces of software that changes how you relate to a daily tool. Email stops being a tab you have open and becomes a surface your agents can operate. Summarising a day of correspondence, triaging leads, drafting polite follow-ups, filing receipts — all of it moves off your todo list and onto a loop. The half-point I held back is for the lack of built-in sandboxing: a misconfigured agent with gmail.send can do real damage to real relationships, and the reference implementation does not help you prevent that beyond what your MCP client offers.

For every Claude Code, Cursor, and Windsurf user I know, this is on the default MCP install list. It is free, it is five minutes, and it unlocks an entire category of agent workflow. If you are reading this and do not yet have it wired up — go fix that now.

Reviewed by Wayne MacDonald, AI Tools Editor at Skiln. Last tested April 2026 against Claude Desktop 0.12, Claude Code 1.2, Cursor 0.45 and the GongRzhe/Gmail-MCP-Server v1.4 reference build. See also: Context7 MCP Review, Google Workspace CLI Review, Firecrawl Review.