agent-email-inbox

Agent Email Inbox

Most agents send mail from a human's account, or from a shared noreply@. Both break the moment a reply arrives: it lands in someone's inbox behind a filter rule, and nothing can answer it.

An Anima identity owns a real address. Mail addressed to the agent arrives at the agent.

Get an inbox

npm install -g @anima-labs/cli
anima init

init provisions an org, an agent and an inbox. Full send capability unlocks after the owner confirms a 6-digit code Anima emails them:

anima verify 

That gate is deliberate. An agent can create an identity on its own; it cannot start mailing strangers until a person says so. If a verification URL is printed, show it in full — never shorten it.

Send

anima email send --agent  \
  --to someone@example.com \
  --subject "Following up on the order" \
  --body "Plain text body."

--to repeats for multiple recipients. --cc, --bcc and --html are available. Rehearsing, or running in CI? Add --test and the server uses fixtures — nothing leaves the building.

anima email send --test --agent  --to you@example.com \
  --subject "Dry run" --body "Nothing was sent."

Read what comes back

anima email list --agent             # --limit, --label, --cursor
anima email get 
anima email search --semantic "invoice dispute"

Replies thread to the agent that sent them. That is what makes "which agent did this, and what happened next?" answerable months later.

Reply, and thread properly

Threading is an API field rather than a CLI flag:

curl -X POST https://api.useanima.sh/v1/email/send \
  -H "Authorization: Bearer $ANIMA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"agentId":"","to":["someone@example.com"],
       "subject":"Re: Following up","body":"...",
       "inReplyTo":""}'

inReplyTo accepts either an RFC Message-ID or an Anima message id, and the API emits In-Reply-To plus the full References chain, so Gmail and Outlook thread it the way a human's reply threads. An id that resolves to nothing is refused rather than quietly sent unthreaded.

A reply is also treated as a conversation rather than a campaign: it carries Auto-Submitted: auto-replied (RFC 3834, so vacation responders stay quiet and agent-to-agent loops still break) and no List-Unsubscribe. A one-click unsubscribe control beside a 1:1 answer makes a personal message read as marketing.

React to inbound

anima webhook create --url https://example.com/hook --events email.received

Verify the signature on your endpoint before trusting a payload.

What is enforced for you

• Suppression. Every send checks an opt-out list first. A recipient who unsubscribed or hard-bounced is never mailed again, whatever the agent asks.
• Loop protection. Outbound carries an RFC 3834 Auto-Submitted marker, and a velocity breaker holds a burst from one agent to one address for approval rather than letting a loop run.
• Own-address refusal. Sending to the agent's own address is rejected, not silently looped.

Also on the same identity

The inbox is one surface of an identity that can also hold a US phone number for SMS and voice, and an encrypted credential vault — same agent, same API, same bill. See agent-phone-sms and agent-credential-vault.

Free tier includes email, vault and MCP. No card. Docs: