Email for Agents · September 30, 2026

Does ChatGPT Have an Email Address?

No. ChatGPT has no address of its own, but it can act inside your Gmail or Outlook. What OpenAI documents today, and what owning an inbox takes.

No, ChatGPT does not have an email address. None of the OpenAI help center pages we read on 30 September 2026 gives an address you can write to that reaches ChatGPT, or an inbox that belongs to it, and the contact page lists no support email either. What OpenAI does document is ChatGPT reading and sending mail from inside a Gmail or Outlook account that you connect. That is a different thing, and the difference decides who the mail is from, whose mailbox it lands in, and who is on the hook when it goes wrong.

This piece settles four things: what “ChatGPT” means in this question, what ChatGPT can do with email according to OpenAI’s own pages as of today, what it would take for any AI to genuinely own an address, and where the OpenAI support address confusion comes from. If you already know you want an agent with its own inbox and just need the steps, go straight to how to give an AI agent its own email address.

Every OpenAI capability below links to the page it was read on, with the date. OpenAI’s release notes change often, so treat anything older than a few weeks as a claim to re-check.

Does ChatGPT have an email address, or could an agent?

The question hides three different things: a model, an assistant product and an agent deployment. Only one of them could ever hold an address.

A model is a set of weights served behind an API. It has no persistent identity, no storage of its own, and no network presence between calls. It cannot own an address for the same reason a compiler cannot: nothing about it persists from one request to the next that mail could be delivered to.

An assistant product is ChatGPT as most people use it: an app, an account, a chat history, settings and connected apps. The account belongs to you. It signs in with your email address, which is why so many results for this query are really about finding or changing the email on your ChatGPT account. OpenAI’s How to change your email address page is about that address, the one you log in with, not one ChatGPT receives mail at.

An agent deployment is a specific process someone runs, with a job, tools and credentials. This is the only one of the three that can meaningfully own an address, because it is the only one with a stable identity someone is accountable for. An address is not a label. It is a mailbox that accepts mail from the internet at any hour, whether or not anyone has a chat window open, and something has to be sitting behind it.

So the honest answer is layered. The model cannot have an address. The product, as documented today, does not have one. An agent that you build and run can, but the address belongs to your deployment, not to OpenAI or to the model underneath it.

What owning an inbox actually requires

For an AI to own an email address in any sense that matters, five things have to exist. Leave any one out and you have something narrower than an inbox.

1. An address. A routable mailbox on a domain whose DNS points at a mail receiver.
2. A credential that authorises the agent to act on that mailbox and nothing else. Its reach should be narrower than the human who issued it, not equal to it.
3. A receive path. Mail arriving at the address has to reach the agent without a human forwarding it: a pollable event feed, a signed webhook, or both.
4. A send path that transmits from that address, and that is safe to retry, because agents retry.
5. A place a human can see and stop things. A view of every thread, a way to hold a send for approval, and a revocation that takes effect immediately.

The fifth is the one people skip, and it is the one that makes the other four safe to turn on. The model will eventually do something you did not intend, whether from a misread instruction or a prompt injection hidden in an inbound message. The question is whether a person finds out before the recipient does.

Most “AI email” features supply the send path and borrow everything else from your own mailbox. That is the core of the distinction that follows.

Can ChatGPT send email?

Yes, but only from an account you connect, and it is your address on the From line. The ChatGPT release notes entry dated June 8, 2026 says that if you have connected Gmail or Outlook, you can ask ChatGPT to draft and send emails from the same conversation: “ChatGPT will create a draft and you can choose to send it without leaving ChatGPT.” The same entry limits sending to the web and to certain paid plans; check the entry itself for which ones, because plan eligibility is the detail OpenAI’s pages revise most often. All of that was read on 30 September 2026.

Several details on OpenAI’s pages are worth knowing before you rely on this.

It sends as you. OpenAI’s Outlook Email and Calendar apps page (read 30 September 2026) says the app searches and retrieves messages “from the signed-in user’s mailbox” and lists the Microsoft Graph permissions it requests, including Mail.Send, Mail.ReadWrite and Mail.Send.Shared. Microsoft’s own permissions reference defines delegated Mail.Send as “Send mail as a user” and describes it as allowing the app “to send mail as the signed-in user.” The mail comes from your address, into your Sent folder, under your reputation.

Shared mailboxes are the nearest thing to a separate address. The release notes entry dated April 8, 2026 says the Outlook apps can read a shared mailbox and “send plain-text email from or on behalf of that mailbox” when you already have access. The Outlook page adds that “ChatGPT does not grant Microsoft sharing or delegation permissions”; access comes from the resource owner or your Microsoft administrator. The shared mailbox is still an address your organisation provisioned in Microsoft 365.

Approval is a setting, not a fixed rule. OpenAI’s Connected apps in ChatGPT page (read 30 September 2026) lists permission options including “Always ask”, “Allow read actions”, “Allow low-risk actions” and, for an eligible app or account, “Allow all actions”, which it says “carries elevated risk because supported actions may run without another confirmation.” It also notes that sensitive actions may require approval or be denied. What counts as low-risk for any particular email action is not spelled out there.

Agent mode is gone. The ChatGPT agent page, read 30 September 2026, now opens with: “ChatGPT agent is no longer available. Use ChatGPT Work for longer, multi-step tasks and finished deliverables.” Guides that tell you to use agent mode to handle your email describe a product OpenAI has retired.

If a chat tells you it emailed something and you have not connected Gmail or Outlook, check your Sent folder before believing it. The only documented send path goes through a connected account.

Can ChatGPT receive email? The closest thing is a trigger on your inbox

ChatGPT has no address that accepts mail, but it can now react when mail arrives in yours. OpenAI’s Scheduled tasks in ChatGPT page (read 30 September 2026) describes event-triggered tasks that “run in Work and respond to supported Gmail, Slack, or GitHub activity,” and says “Gmail tasks can respond to a new message or filter by sender or subject.” They need Work and an eligible paid plan; the page says Free and Go accounts cannot create them. The Connected apps page adds that such a task “can access only the connected account and information allowed by its existing app permissions,” and that if an action requires approval, “the task pauses until you review it.”

This matters because some of the top results for this query still say ChatGPT has no email triggers at all. As of today, OpenAI documents one. But look at what it is: a listener on your Gmail account. Anyone who wants the agent to act has to email you. The sender and subject filters decide what wakes the task up, but OpenAI’s pages describe no way to limit what it can then read to the messages meant for it. Our reading, not OpenAI’s wording: assume the task can reach whatever that Gmail connection can reach in your mailbox.

The same page also covers ordinary task notifications: after a task is created, you can choose push, email or both under Settings > Notifications. That is ChatGPT emailing you a notification, not ChatGPT holding an inbox anyone can write to.

Inside your mailbox versus a mailbox of its own

Here are the five requirements from earlier, set against each OpenAI surface documented today and against an agent-owned inbox. Everything in the OpenAI columns comes from the pages linked above, read on 30 September 2026.

The middle column needs explaining, because it is the one place OpenAI comes close to giving an agent its own mailbox. OpenAI’s ChatGPT Workspace Agents for Enterprise and Business page, read 30 September 2026, says that for each app connection a builder can choose an “End-user account”, where each person running the agent authenticates as themselves, or an “Agent-owned account”, where “the agent uses a shared connection.” It recommends a service account when possible and warns against using a personal account “unless you understand the risks.” It also describes Connector Action Constraints, with the example of only allowing “an email action to send messages to recipients at a specific domain.”

That is a real capability, and for a Business or Enterprise workspace that already runs Google Workspace or Microsoft 365 it may be all you need. The caveats: the page does not describe OpenAI providing that service account, so as we read it the mailbox is one your organisation creates and licenses in its own tenant; the agent inherits whatever that account can reach; and OpenAI’s page is explicit that constraints govern what the agent can ask a connector to do but “do not filter or restrict the data a connector returns.”

The right-hand column is the model this site is built around, explained in full in email for AI agents. The agent gets an address of its own, so a mistake lands in a mailbox containing only that agent’s mail rather than yours. The trade-offs of the two approaches are laid out in detail in an AI agent inbox versus Gmail.

What an agent-owned inbox looks like in practice

On Email for Agents, each inbox is a live address on agents.emailforagents.ai, created for one agent. Mapping it to the five requirements:

• Address: created with a single API call; the response carries a stable inbox ID. Test projects get a non-routable .mail.invalid address that cannot reach an internet recipient, so you can build the whole loop before anything real is sent.
• Credential: either an OAuth connection grant for an MCP client, or a scoped API key for your own code. A key cannot exceed the scopes of the human who created it, and authority is re-read on every request, so removing that human’s access narrows every key they minted on the next call.
• Receive path: message.received events, pollable or delivered to a signed webhook.
• Send path: sends accept an Idempotency-Key, so a retry does not become a second email.
• Human stop point: a console showing every thread, and an approval queue that emits approval.requested when a send is held and approval.rejected when a person declines.

Two details matter if you are thinking of wiring this to an MCP-capable assistant. First, the remote MCP endpoint at https://mcp.emailforagents.ai/mcp requires an OAuth connection grant, and an API key is rejected there by design, as covered in why an API key is rejected at the MCP endpoint. Second, the grant level shapes the tool list itself: the connection scopes are email.read, email.draft and email.send, and a read-and-draft grant produces a session in which send_message is simply absent. create_draft cannot transmit mail at all. Whether a particular MCP client can complete that OAuth flow is a question for that client; we do not certify or list any client as supported.

What this does not give you, today: custom domains require a paid plan, and paid plans are not purchasable yet. The free allowance is one live inbox and 100 email units a month per verified owner identity, which is enough to prove the loop and not enough to run volume.

What is OpenAI’s support email address?

OpenAI’s How can I contact support? page, read 30 September 2026, lists no support email address. It says to contact support “by opening the chat bubble icon displayed at the bottom-right of help.openai.com,” where a virtual assistant answers first and a human agent can join. Help center pages also carry an “AI Phone Call” option at 1-888-GPT-0090, described on OpenAI’s AI phone support page. That is a phone line, not an email channel.

Third-party pages that quote a support address are not quoting that page. If you need OpenAI to act on your account, use the chat on help.openai.com and include the email you sign in with, which the page asks for.

Two related points trip people up:

ChatGPT does have a phone number. OpenAI’s 1-800-ChatGPT page, read 30 September 2026, describes calling 1-800-CHATGPT (1-800-242-8478) from a US or Canadian number, and messaging the same verified contact on WhatsApp. The page describes calls and WhatsApp messages. It describes no email channel.

An address claiming to be ChatGPT is not ChatGPT. OpenAI’s Verifying Communications from OpenAI page, read 30 September 2026, lists the domains customers may receive OpenAI email from: openai.com and the subdomains ads., email., mail., sales. and tm.openai.com, plus c-openai.com. Those are addresses OpenAI sends from, not addresses that reach ChatGPT. A message from something like chatgpt@gmail.com asking you to share files is on none of those domains and is not from OpenAI. A From line showing one of them still deserves a check of its DMARC result, because the header on its own proves nothing.

If what you actually wanted was an AI that people can email directly, the answer is not to find ChatGPT’s address. It is to give an agent you run an address of its own, scoped to its job, with a person able to see and stop what it sends. The step-by-step setup starts in a test project, so the first message it ever sends is one nobody outside can receive.

foundations inboxes permissions research

Keep reading

All guides · Documentation · Email for Agents