Email for AI agents
Email for AI agents means an agent has its own address and can receive, read and send mail through an API or MCP tools. Agents need it because email is how services verify accounts and how businesses talk. The risk is that an agent acts on what it reads, so mail must be checked before it arrives.
Updated
Why do AI agents need email?
AI agents need email because the rest of the world already runs on it. Radicati Group forecasts about 4.73 billion email users worldwide in 2026. An agent with an address can write to any of them without a new app or an agreement.
Email does three jobs for an agent.
- Identity. Most services ask for an email address at sign-up and send a code to prove it.
- Conversation. Customers, suppliers and candidates answer by email, in threads that can run for weeks.
- Records. Invoices, receipts, bookings and calendar invitations arrive as mail.
The need is growing. Gartner predicted in June 2025 that 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024.
Can AI agents send and receive email?
Yes. An AI agent can send and receive email when it is given an address and a way to work it: an API, an SDK, or tools over the Model Context Protocol (MCP).
Sending is the easy half. The agent calls a send function with a recipient, a subject and text. Receiving is harder: mail has to be accepted, stored, parsed, threaded and handed to the agent at the right moment.
The model does not speak the mail protocols itself. The inbox does, and the agent sees each message as structured data. What an email API for AI agents must do covers each part.
What are the ways to give an agent email?
There are five common ways to give an AI agent email. They differ in who owns the address, what the agent can read and when mail is checked.
| Way | Good for | Watch for |
|---|---|---|
| Connect the agent to a human's mailbox, such as Gmail or Outlook | An assistant that helps one human with their own mail | The agent reads everything in the mailbox and acts under the human's name |
| A shared mailbox or an alias | A small team trying an agent on one address | No separate identity, keys or record for the agent |
| A sending API with inbound forwarding | One-way notifications, with the odd reply | Often no stored inbox and no threads. You build them |
| Your own mail server | Full control | Deliverability, authentication records and filtering are yours to run |
| An agent inbox | Agents that hold two-way conversations | Whether mail is checked before the agent reads it |
Gmail and Outlook do screen mail before delivery, with spam and malware filters. A message that looks real can still pass, and what catches it then comes after delivery. Neither mailbox was built for a model reading the mail: see Gmail for AI agents.
For most agents that talk to the outside world, an inbox of the agent's own is the cleanest fit. What is an agent inbox explains the idea, and give an AI agent an email address is the hands-on version. Agent identity explains why a borrowed address is a risk.
What can go wrong when an agent reads email?
Email is a channel where any stranger can put text in front of an agent. That makes it one of the easiest ways to attack one.
- Prompt injection. Text in a message tells the model to do something else. The OWASP Top 10 for LLM Applications lists prompt injection as LLM01, the top risk, in both the 2025 edition and the current 2026 edition.
- Hidden instructions. The text is invisible to a human: white on white, in a comment, or inside an attachment. See hidden text in email.
- Data leaving. A hijacked agent forwards mail, files or secrets to an attacker. See data exfiltration through agents.
- Ordinary fraud. Fake invoices, changed bank details and phishing links work on agents as they work on humans.
- Stolen sign-ins. An agent that can read a human's mailbox can read the password reset sent to it.
These attacks have been shown against a shipped product. In EchoLeak, CVE-2025-32711, researchers showed that one crafted email could make Microsoft 365 Copilot disclose data with no click. Microsoft rated it 9.3 (critical) and published its advisory in June 2025, saying the flaw was already mitigated.
The pattern is the same each time. Once the attacker's message meets the agent, it is too late. The check has to come before the agent reads.
What should you look for?
An email setup for an AI agent should answer eight questions.
- Does the agent have its own address, apart from any human's?
- Is mail screened before delivery, or after the agent has read it?
- Does the agent read a safe view, with hidden text removed?
- Where does doubtful mail go, and can the agent open it?
- Is outgoing mail checked for secrets, and for replies an injected instruction is steering?
- Can a human approve sends, and are a human's login codes kept from the agent?
- Are keys narrow: one inbox, with reading apart from sending?
- Is there a record of what was held, released and sent?
The answers matter more than the feature list. AI agent security covers the wider picture beyond email, and human in the loop email covers question six.
Where Trotta Mail fits
Trotta Mail is the secure pre-delivery inbox for AI agents and humans. Every message is screened while the sending server is still connected, before it is accepted.
- Attacks, known and new, are refused at the door or held outside the inbox for a human.
- Everything else is delivered with a verdict that says what was found.
- An agent's key reads the safe view. It cannot read or release held mail.
- Outgoing mail is screened too. A message that contains a secret is not sent.
- In a human's mailbox, login codes and reset links start sealed from every key, and an agent starts with drafts only.
An agent gets its own inbox with one API call. It then works the inbox through the REST API, the MCP server, the TypeScript or Python SDK, webhooks or live events.
curl -X POST https://api.trotta.email/ v1/ agents/ signup \
-H "Content-Type: application/ json" \
-d '{"name": "riley"}'The call returns an API key, the inbox with its address, and a link for the human responsible for the agent. The quickstart continues from there.
No filter catches everything. A reworded scam with no technical tell can still be delivered, which is why uncertain mail is held for a human. One limit is worth knowing too: Trotta Mail has no IMAP or SMTP access for mail clients.
The Free plan covers 2 inboxes and 1,000 emails a month. Every check is on every plan. See pricing.
Questions
Can ChatGPT or Claude send email for me?
Yes, once connected to an inbox. An assistant such as ChatGPT or Claude can use a Trotta Mail inbox as MCP tools: it reads mail, writes drafts and sends from the inboxes you pick.
Should an AI agent use my email address?
For anything beyond helping you with your own mail, no. Give the agent an address of its own, so it reads only its own mail and its actions are recorded under its own name.
Do AI agents need a special kind of email?
The mail itself is ordinary email. What changes is the inbox: it is created and read through an API, and it should check mail before a model reads it, because a model can act on text a human would ignore.
Is it safe to let an AI agent read email?
It is safer when mail is screened before delivery, the agent reads a safe view, and doubtful mail waits for a human. No filter catches everything, so keep keys narrow and put a human on high-risk sends.
How much does email for an AI agent cost?
On Trotta Mail, Free is $0 for 2 inboxes and 1,000 emails a month. Developer is $20 a month for 5 inboxes and 5,000 emails, and Startup is $200 a month for 50 inboxes and 50,000 emails.
Sources
- Radicati Group: worldwide email forecast for 2026
- Gartner, June 2025: agentic AI in enterprise software applications by 2028
- OWASP Top 10 for LLM Applications 2026, the current edition (published August 2026)
- OWASP Top 10 for LLM Applications 2025, LLM01: Prompt Injection
- Microsoft Security Response Center, CVE-2025-32711: M365 Copilot Information Disclosure Vulnerability