What you need first
Three things, and the second is the most delicate.
- Node.js installed on your computer.
- An OAuth authorization to Google. OAuth is the system by which you give an app permission to access your inbox without giving it your password. You create a project on Google Cloud, enable access to Gmail, and obtain the credentials. It's the most technical step: some servers have a guided procedure that simplifies it.
- Deciding the scope. It's best to start by granting permission to read and create drafts, not to send: that way the AI prepares but doesn't dispatch.
How to do it
- Get the OAuth credentials. Follow the procedure of the server you choose to authorize access to Gmail. Servers with "auto-auth" open a browser window where you confirm access, and do the rest on their own.
- Add the server to the configuration file. The operational syntax (Gmail server with automatic authentication):
{
"mcpServers": {
"gmail": {
"command": "npx",
"args": ["-y", "@gongrzhe/server-gmail-autoauth-mcp"]
}
}
}
- Restart the assistant and complete the sign-in when the Google window opens.
- Work on your mail in plain language. For example: "search for the emails from client Bianchi over the last month and summarize where we stand," or "prepare a draft reply to this last email, polite and proposing a call." The draft stays in your Drafts: you reread it and send it yourself.
A concrete example
Franca receives dozens of emails a day and loses time tracking down the threads of conversations. She connects the Gmail server granting read and draft permission, not send. In the morning she asks: "summarize this morning's unread emails grouping them by client and flag the urgent ones."
The AI gives her back an orderly picture in thirty seconds. For the three urgent ones she asks: "prepare the draft replies." Franca finds them in Gmail's Drafts, rereads them, corrects one, and sends them by hand. She has stopped digging through the inbox and writing from scratch, but every email goes out only after her own eyes. Control over sending stayed hers, and that's exactly where she wants it.
When it does NOT work (and how to fix it)
If the OAuth authorization fails
It's the most common obstacle. If the Google window gives an error, check that you've enabled access to Gmail in the Google Cloud project and that you're using the right credentials. Some assistants (for example, the desktop apps of certain plans) require a paid plan and a dedicated connector for email: check your own requirements before insisting.
If the AI wants to send on its own
Don't give it send permission until you trust it. Keep the scope at "read and drafts": the AI prepares, you dispatch. A wrong email sent automatically can't be recalled, and the risk isn't worth the convenience.
If the emails contain sensitive data
Your inbox is full of personal data, contracts, codes. When the AI reads your emails, that content passes through the model. Limit searches to what's needed ("only the emails from this client"), avoid having it comb through the entire inbox, and for the most confidential mail consider keeping it out of reach.
A tip from someone who actually uses it
Keep send permission off as a rule, and turn it on only for precise, repetitive tasks you trust (for example, a standard reply to identical requests), never for important correspondence. Mail is the place where an AI mistake shows the most and is recovered the least: the draft is your safety net, automatic sending takes it away.
Frequently asked questions
Does it work only with Gmail?
The most widespread and ready-made servers are for Gmail, but there are servers for other mail systems too. Look for an MCP server for your provider; if you can't find a ready-made one, Gmail remains the one with the most proven procedure. The principle (read, drafts, send under confirmation) applies to all of them.
Can the AI organize the inbox, not just read it?
Yes: email servers let you label, archive, and sort messages, in addition to reading them. Here too it's best to start slow and keep confirmation on actions that move or delete, so you don't end up with the inbox reorganized in a way you didn't want.
Is it safe to give your email to the AI?
Here's the misunderstanding to clear up: it isn't a flat yes or no, it's a matter of limits. Granting read and draft access, on targeted searches, with sending under your control, is reasonable for ordinary work mail. Granting full access with automatic sending to the entire inbox is a risk few should take. You decide the safety with the breadth of the permission; it isn't a property of the tool.