Blog

Connect Claude to your email with MCP

How to connect Claude to a real mailbox over MCP: the claude.ai connector, the Claude Code command, what the tools can do, and how to scope the connection to one mailbox.

Jonathan S.Updated 13 October 20266 min read

Claude can read, search, send and tidy a mailbox on your domain once it has an MCP server to talk to. The setup is one URL in claude.ai or one command in Claude Code, and you choose which mailboxes it may see.

What you need before you start

Two things. A mailbox on a domain you own, because Claude needs a real address to read from and send as, not a forwarder. And a Faivelo plan that includes the MCP server, which is Growth ($12 a month) and up, or the 14-day trial. If the mailbox does not exist yet, create it in the dashboard first; it takes a minute, and the DNS for the domain is already done if the domain is verified.

You do not need an API key for either path below. Both sign in through your browser and hand Claude a short-lived token that it refreshes on its own. Keys are for agents with no browser, covered at the end.

The two ways in

  1. claude.ai. Open Settings, then Connectors, then Add custom connector. Paste https://faivelo.com/api/mcp and continue. Your browser opens the Faivelo sign-in page.
  2. Claude Code. Run the command below. The first time Claude Code uses the connector it opens the same sign-in page and mailbox picker.
  3. The consent screen. Tick the mailboxes Claude may reach. Everything you leave unticked stays invisible to it. Optionally allow it to create new mailboxes on your domains (capped at 10 a day). Approve.
  4. Check it works. Ask: "Summarise my unread mail in support@." A real answer means the connection is live. Every connection is listed under Settings, Connected Apps, with a one-click revoke.
Claude Code
claude mcp add --transport http faivelo https://faivelo.com/api/mcp
Signs in through the browser the first time and remembers the connection.

Cursor, VS Code, Codex, Gemini CLI and Windsurf take the same URL in their own config; the exact lines for each are in the coding agents guide. The Claude Code plugin (/plugin install faivelo@faivelo) adds a skill alongside the server so Claude also knows the setup path and the send API when it writes code for you.

What Claude can do once connected

The server exposes the mailbox as a set of tools, scoped to the mailboxes you ticked. Reading: list_messages, read_message, get_thread for a whole conversation, search_messages in one mailbox and search_all_mailboxes across every shared one, list_folders, get_attachment, get_contacts. Sending: send_message, with attachments and an optional sendAt for a scheduled send. Drafts: create_draft, update_draft and send_draft, so Claude can leave a reply in Drafts for you to read before it goes. Scheduled mail: list_scheduled and cancel_scheduled. Housekeeping: flag_message, label_message, mark_all_read, move_message, restore_message, mark_spam, delete_message, empty_trash. Aliases: list_aliases, create_alias, delete_alias, limited to routes into the shared mailboxes. And get_delivery_status to ask whether mail to someone bounced.

In practice you ask in plain words. "Find the invoice from the printer and reply that payment goes out Friday." "Move every newsletter from this month into a Newsletters folder." "Search support@ for refunds this week and give me a list." Claude picks the tools; you see the result in the chat and the changes in the mailbox, the same as if you had made them in webmail.

Claude gets a mailbox, not your account. The consent screen is where you draw that line.

Three choices that keep it safe

Scope it to one mailbox. Give Claude a mailbox made for the job, assistant@ or support@, and leave your personal address unticked. A connection can only see what was ticked, never other mailboxes, domains, billing or settings. If you want Claude to help with your own inbox, tick that one and nothing else.

Prefer drafts for anything that leaves the company. Ask Claude to draft replies rather than send them, and read them in Drafts before you press send. Sends through the connector are real and count against your plan's monthly allowance, and a destructive action (delete, move, mark spam) does exactly what it says. Claude asks before destructive steps, but a draft is the safer habit for customer-facing mail.

Revoke rather than wonder. Access tokens last an hour and refresh themselves; a replayed refresh token revokes the whole connection automatically. If you stop using a connection, remove it under Settings, Connected Apps, and access ends that second. Reconnecting later takes the same minute.

For an agent with no browser

A script on a server or an agent in CI cannot click through a sign-in. For those, create an API key under Settings, Developers, with the kind set to agent and the mail and mailbox scopes it needs, then send the key as a header. An agent key reaches only the mailboxes it created or was granted, which is the same fence the consent screen draws for a person.

Claude Code, with an agent key
claude mcp add --transport http faivelo https://faivelo.com/api/mcp \
  --header "Authorization: Bearer $FAIVELO_API_KEY"
Any MCP client that reads JSON config takes the same url and headers.

If you want the longer story of what MCP is and why a mailbox is a good fit for it, read what MCP means for email. The full tool list and the security notes live in Connect Claude, and email for AI agents covers agents that need an inbox of their own rather than access to yours.

Found this useful? Share it.

A mailbox Claude can work in

Start a trial, add your domain, create assistant@, and paste one URL into Claude. The trial includes the MCP server; after it, Growth is $12 a month for unlimited mailboxes and the server. Fourteen days, no card.

Start freeFree 14-day trial. No card needed.

Keep reading