Gmail MCP server: connect AI agents to Gmail from the EU

Gmail is Google's email service, used privately and by organisations on Google Workspace. Google LLC is based in Mountain View, California.

Google's official Gmail MCP server runs at https://gmailmcp.googleapis.com/mcp/v1 and is in Developer Preview, so tools and behaviour may still change. It acts with the OAuth access of the signed-in user. It can search, read, label and draft, but it deliberately has no tool to send mail, which means a person always presses send. That makes it well suited to agents that prepare work for people rather than act alone.

A support agent can read new customer emails each morning, label them by topic and urgency, and draft replies using answers from your help centre and CRM. A finance agent can search for supplier invoices received this month and list the ones still missing from the accounting system.

Sales teams can ask "what did this customer last write to us, and what did we promise?" and get a summary of the thread. Assistants can draft follow-ups after meetings so they are waiting in the drafts folder for review.

Google's Workspace MCP servers need a Google Cloud project with the Gmail API and the Gmail MCP service enabled, and an OAuth client configured for that project. Your Workspace admin may need to allow the app.

Add the Gmail connector on inference.eu, enter your OAuth client details and sign in with the mailbox the agent should use. Choose the narrowest scopes that work. Tokens are stored encrypted in the EU and used only by the agents you allow; access can be revoked at any time in the Google account settings.

Mail is fetched from Google's own MCP endpoint, so data stays with Google, a US company, as it does for your other Gmail integrations.

On inference.eu, the model that reads your mail runs in EU data centres, and your Google credentials are stored encrypted in the EU. Email holds some of the most personal data a company has, including customer complaints, health details and contracts. Connect a shared or team mailbox rather than a personal one where possible, and tell the agent which senders and labels it should handle.

Because the server cannot send mail, the biggest risk of an agent emailing the wrong person is removed by design. Drafts and labels are still changes: a wrong label can hide a message from the team that should see it.

On inference.eu you can let searches and reads run on their own and require a person's approval for labelling and for creating drafts while you build trust in an agent. Be aware that emails can contain instructions written by outsiders; agents should treat message content as data, not as commands.