An AI agent that sends email from its own address
Most AI writing tools stop at a draft. The last step — actually sending it, to the right person, and knowing it left — is where the work is.
Ask a chat assistant to write an outreach email and you get an email. Then you copy it, open your mail client, find the address, paste, adjust, send. Repeat nineteen times. The writing was never the slow part.
An agent with its own mailbox closes that loop. It finds the contact, writes the message, sends it, and tells you what happened.
Sending as itself, not as you
There are two ways to give an agent email, and the difference matters more than it first appears.
| Agent uses your mailbox | Agent has its own address | |
|---|---|---|
| Access needed | Full access to your inbox and history | None to your mail at all |
| If it makes a mistake | Sent from you, to your contacts | Sent from an address that is obviously the assistant |
| Recipient sees | You | An assistant acting for you |
| Reputation risk | Your personal sending domain | Contained to the agent's address |
Handing a model your primary inbox means every email you have ever received is in scope for a mistake, and every message it sends carries your name. A separate address is a smaller, more honest surface: the agent can introduce itself as what it is, and a bad send is embarrassing rather than damaging.
On MyClawn every cloud clone gets its own address on its own machine and sends from that.
"Sent" has to mean sent
The failure mode that erodes trust fastest is an assistant reporting success it did not have. A language model will happily write "I have sent the email" because that is what a helpful reply looks like, whether or not a mail server ever accepted the message.
So the rule here is mechanical rather than promised: the agent reports what the mail server actually returned. If the message was accepted it says so; if it failed it says it failed and why. It will never claim something was sent when it was not — that is worse than useless, because you stop checking.
What is not enabled
Stated plainly so nobody builds a workflow on it: a clone can send mail today, but receiving is not wired up yet. Replies do not land in the agent's mailbox. If your job depends on the agent reading responses and following up on its own, that does not work right now. Use a reply-to address you control and handle replies yourself.
What to supervise
Email is the category where an agent's mistakes reach other people, so it deserves more supervision than research does.
- Watch the first batch. The screen is live. Read what it is about to send before it sends many of them, not after.
- Check who it found. Wrong-recipient errors come from the research step, not the writing step.
- Keep volume sane. A machine that can send two hundred messages should not. Cold-outreach rules and anti-spam law apply exactly as they do to you, and a fast sender is a blocked sender.
- Do not automate anything contractual. Quotes, commitments and agreements should leave with a human behind them.
What it is good for
- Supplier and vendor enquiries — the message is the same shape twenty times over
- Chasing a quote or a document someone owes you
- Introductions and scheduling, where the writing is routine
- Anything where the research and the sending belong in one step
How to try it
Signing up is free and takes about thirty seconds — no install, no card. A clone with its own machine and mailbox starts at $19/month; the tier table is on the homepage.