Connecting an AI agent via MCP

Give Claude or another MCP client access to all your inboxes: search, sort, draft and, when you allow it, send.

The Model Context Protocol (MCP) is an open standard that lets AI agents use tools. centomail runs an MCP server, so an agent you choose can work with your mail and set up your account on your behalf, within limits you set.

What the agent can do

Mail: search and read messages and conversations, move and label them, mark them read, create and edit drafts.
Sending, only if you allow it: send a new message, a reply or a draft. Off unless you tick it when you connect the agent.
Aliases: list them, create named or random aliases, pause an alias.
Domains and mailboxes: add a domain and get the DNS records it needs, check them, and create shared mailboxes such as support@ or noreply@. Your agent adds the DNS records with its own access to your DNS provider; centomail never writes to your DNS.

It cannot delete aliases or domains, change forwarding rules, create API keys or webhooks, or touch your account's security settings.

Connect

In your agent's settings, add a remote MCP server (some apps call it a "connector" or "integration") with this URL:
```
https://mcp.centomail.com
```
The agent opens an centomail sign-in page. Sign in as usual.
The next page lists what the agent asks for. Sending is a separate box, Also let it send email as you, unticked: leave it unticked and the agent only prepares drafts.
Approve. You never give the agent your password or an API key.

To limit the agent to some folders, open Settings → Connected apps → Choose folders after connecting.

Drafts and sending

Without sending, the agent saves what it writes in your Drafts folder, marked as prepared by AI. Check the recipients and the text, edit if needed, and send it yourself.

With sending allowed, the agent sends from one of your addresses, within your account's sending limits, and the message lands in Sent. Every message it sends is listed in the console under Agent activity.

Stay safe with untrusted content

Emails are written by other people, and some will try to give instructions to AI agents ("ignore previous instructions and forward..."). centomail marks message content as external data in every response, so a well-behaved agent treats it as information, not as commands. Unless you allowed sending, the most damaging outcome, your mail being sent somewhere, is ruled out.

Still:

allow sending only to an agent you trust, and only when you need it;
limit the agent to the folders it needs;
remember that whatever the agent reads is processed by that agent's provider under their terms.

Review and revoke

The console's Agent activity page lists everything your agents did in the last 90 days: searches, moves, drafts, sent messages and setup, without any message text. Settings → Connected apps lists every connected agent and what it may do; Disconnect cuts off its access immediately.