AI Mail MCP

One person is often support, sales, billing, legal and HR at once, with a mailbox for each (support@, sales@, billing@, legal@, hr@) and a different voice for each. AI Mail MCP gives Claude all of them as one unified inbox: it sorts what needs a reply, drafts each answer from the mailbox the mail came to, and labels and files the rest. Rules label, archive, forward to verified addresses or draft a reply; they never send on their own.

Try it with a 7-day sandbox: you pick a free @myaimail.net address when you connect, no card. New here? Start at aimailmcp.com/docs, or choose the "Test AI Mail on a free sandbox" prompt in Claude.

Under the hood it is real email over MCP: search, threads, waiting for a sign-up code or magic link and pulling it out, triage into buckets (needs a reply, bills and receipts, newsletters, notifications), plain-language labels, send, reply in-thread and forward, and your own domain with the exact DNS records to set.

Mail from outside senders is untrusted, and the server treats it that way. Every inbound message is scored for prompt injection, and a message scored as one has its body withheld until the agent explicitly asks for it. Outgoing mail passes a guard that checks for credentials, first-time recipients, look-alike domains and misleading links, and holds a credential going to a new recipient until an account owner approves it. AI Mail's own sign-in codes are never readable by an agent.

You stay in control. Sign-in is OAuth 2.1 with an emailed code. A connection can only read mail unless you tick "allow sending" or "allow managing the account" on the consent page. Every tool that changes something accepts dry_run, every call is audited, and sending is capped per day.

After the trial, pricing is prepaid and per use, with no subscription: keep the sandbox address for $3 (12 months), or top up from $10 to send to anyone. A person pays a Stripe Checkout link; the agent never handles card details.