Questions · Published 2026-11-18 · Updated 2026-11-18 · AgentiSend
How do always-on AI agents like OpenAI dots send email safely?
Always-on AI agents like OpenAI dots send email safely with read-only tools for background checks, and sends that are capped, held for approval and stoppable.
Always-on AI agents like OpenAI dots send email safely with read-only tools for background checks, and sends that are capped, held for approval and stoppable. The read-only checks run without a person; every send is capped per key, a risky one waits for a person, and one call stops them all. AgentiSend's connector marks each of its tools with read-only, destructive and open-world hints, so a host can let the checks run unattended and send everything else through approval.
How dots use an email app
OpenAI introduced dots on 2026-09-29 as agents that keep working when nobody is chatting with them. Two parts of that announcement decide how an email app behaves inside one. When idle, a dot does "proactive research" with connected apps through tools restricted to read-only, so it cannot send or change anything while it looks for problems. When it wants to act, "auto-review" checks the action against your instructions and Custom Rules and decides whether it can proceed, needs your approval, or is yours to do. Dots reach apps through the ChatGPT plugin directory, where AgentiSend is listed.
The checks a dot can run in the background
These connector tools are read-only, so they are available to proactive research:
| Tool | What a background check finds | REST equivalent |
|---|---|---|
list_emails with status bounced, failed or complained and since | Messages that did not arrive since the last look | GET /emails |
why_was_this_not_sent with no email_id | Whatever is holding sends on this key or account right now, each with is_normal_wait | none, MCP only |
get_trust_standing | The account's standing (ok, warning, throttled or paused) with each reason's value and threshold | GET /trust/standing |
get_agent_budget | Sends left in this key's budget and when it refills | GET /limits/keys/{id} |
explain_email | For one bounced or failed message: the verdict, whether a retry can work, and the steps that fix it | GET /emails/{id}/explain |
get_email, get_email_events | One message and its delivery events in order | GET /emails/{id}, GET /emails/{id}/events |
A dot can watch for failed messages the same way a script would:
curl -sS -X GET https://api.agentisend.com/emails \
-H "Authorization: Bearer $AGENTISEND_API_KEY"200
{
"data": [],
"has_more": false,
"next_cursor": "example",
"object": "list"
}And read the account's standing, which moves with bounce and complaint rates:
curl -sS -X GET https://api.agentisend.com/trust/standing \
-H "Authorization: Bearer $AGENTISEND_API_KEY"200
{
"domains": [],
"history": [],
"next_review_at": "2026-09-04T09:14:00.000Z",
"probation": false,
"probation_cap": -9007199254740991,
"probation_ends_at": "2026-09-04T09:14:00.000Z",
"reason_codes": [],
"reasons": [],
"sla_deadline_at": "2026-09-04T09:14:00.000Z",
"state": "ok",
"streams": {
"marketing": {
"bounce_rate": 1,
"bounced": -9007199254740991,
"complained": -9007199254740991,
"complaint_rate": 1,
"sent": -9007199254740991,
"state": "ok"
},
"transactional": {
"bounce_rate": 1,
"bounced": -9007199254740991,
"complained": -9007199254740991,
"complaint_rate": 1,
"sent": -9007199254740991,
"state": "ok"
}
},
"window_days": -9007199254740991
}What stops a send without a person
Every key carries a send budget the agent cannot raise, and a new key's ceiling is 600 requests in a 60-second window. The loop guard holds the send that would be the 4th near-identical message to one recipient inside 60 minutes, and answers approval_required instead of sending it. One call, POST /limits/kill-all, stops every key on the account, and POST /limits/keys/{id}/kill stops one.
What waits for you
A send the loop guard holds, or one the agent puts up itself with request_approval, becomes a row at GET /agent-actions in state pending. The agent is answered approval_required and its own key cannot approve. A person releases it with POST /agent-actions/{id}/approve or refuses it with POST /agent-actions/{id}/reject. That is the same shape as a dot's own approvals: the agent proposes, a person decides, and the decision is recorded with who made it.
Connect it
In ChatGPT, add AgentiSend from the plugin directory and approve the connection; the dot then sees the same connector. Any other assistant can connect to https://api.agentisend.com/mcp with OAuth.