CarlyEmail

CarlyEmail gives an agent a real email address it can work from — not a simulation and not a relay through your personal mailbox.

Create an inbox in a single call and it starts receiving straight away. Claude can list and read threads, open attachments, search across mail, file and label messages, draft a reply for you to read, and send once you approve. Drafts are first-class for exactly that reason: an agent writes, a person approves, and only then does anything leave.

Every tool declares what it does before it runs. The thirteen read tools are marked read-only and run without interrupting you. Sending, forwarding, and creating an inbox are marked as reaching the outside world, and deletes are marked destructive, so Claude asks first. Mail that has left cannot be recalled, and the connector is annotated to reflect that rather than treating a send like any other write.

Access is scoped. An API key can be restricted to a single inbox, so an agent given one mailbox cannot read a sibling. Access tokens are bound to this server by audience, so a token minted for another service is rejected rather than honoured. Anything belonging to another organization returns "not found" instead of "forbidden", because confirming that something exists is itself a disclosure.

Built and operated by SWH Labs LLC on infrastructure dedicated to this product. Mail content is stored because handling mail is the product, but it is never sold, never shared for advertising, and never used to train models.