# AgentMail and STACK

By the STACK team. Published 13 November 2026.

AgentMail gives an AI agent an email inbox. STACK gives it an inbox that is joined to the rest of an account. This is what that difference means for a message after it arrives.

AgentMail sells one thing, an email inbox for AI agents, and its documentation runs to several hundred pages about it. This article sets that inbox beside the one in a STACK account. The facts about AgentMail are from its site, price list and documentation as they stood on 13 November 2026, and the pages used are listed at the end.

Taken as an inbox alone, AgentMail’s is the fuller one. It has drafts, labels, IMAP and SMTP access, live event streams and open tracking, and STACK has none of those.

The difference that matters more is what a message can turn into. On AgentMail a message stays a message, and anything further is for your own code to do. On STACK the inbox shares an account with memory, a drive, a document search, a calendar, a phone number and a card. A fact in an email can be remembered with its source. An attachment can be searched next month. An invoice can be paid. The same agent does all of it, under one policy.

## Side by side

|  | AgentMail | STACK |
| --- | --- | --- |
| Send, receive, reply, forward, threads | Yes | Yes |
| Your own domain | Yes | Yes |
| Drafts and labels | Yes | No |
| IMAP and SMTP access | Yes | No |
| Open tracking | Yes, on your own domain | No |
| New mail reaches the agent by | Webhook, WebSocket or polling | A wake-up over MCP, a signed webhook or a queue |
| Attachments | Downloaded, with text extracted | Saved to the account’s drive, up to 25 MB an email |
| Remembering what a message said | No memory | A fact stored in memory keeps the message it came from |
| Searching old attachments and documents | No | Yes, by meaning, across the drive |
| One history across channels | Email only | Email, texts, voicemail, call and meeting transcripts, Telegram and scanned post |
| Agents reading each other’s mail | Through keys you scope yourself | Yes by default, searchable across all your agents, and you can switch it off |
| Knowing the sender | Spam and virus detection | One record of who your agents have dealt with, on every channel |
| Calendar from invitations | Yes | Yes, and the agent can join the meeting |
| Acting on a message | Send mail | Pay, call, text, browse, post a letter |
| Mail scanned for attacks on the agent | Opt-in, in its Labs programme, in the US region only | On every account |
| A published score for the scan | No | F1 0.87 on 1,087 samples |
| Who the agent may write to | Allow and block lists | Addresses and domains, and a reply-only rule |
| Approval before an email is sent | A pattern you build with drafts and a restricted key | A rule in the policy. STACK holds the email and asks you |
| Record of what was sent | Message history. No audit log | Signed record kept 365 days, verifiable offline |
| Accounts for your own customers | Pods, with SCIM | Accounts a platform opens for its customers |
| Where data is kept | US. An EU region on the Enterprise plan | The EU, on every plan |
| SOC 2 | Type II, first audit completed in 2025 | No audit yet |
| Free to try | A free plan: 3 inboxes, 100 emails a day | 14 days |
| Price | $20 a month for 10 inboxes, $200 for 150 | $19 a month with 1 inbox, then €1 an inbox |

## What happens to a message after it arrives

Take an invoice that arrives by email with a PDF attached, from a supplier the agent has dealt with before.

On STACK the PDF is saved to the account’s drive as the message is stored. The agent can read it, and it can add it to the Brain, which is STACK’s search over an account’s documents. From then on any of your agents can find that invoice by what it says, without knowing its file name or which inbox it came to.

The agent can write what it learned to memory: the amount, the due date, the supplier’s new bank details. Memory on STACK is a graph shared by the agents of an account, and a fact written from a message keeps a reference to that message. Months later an agent that recalls the due date can also show which email said so, and whether that email was flagged.

The sender is checked against one record of everyone your agents have corresponded with, on any channel. STACK marks the message if the sender is new, if the domain resembles one you already deal with, if a known name is writing from a new address, or if the message contains bank details. A supplier who has emailed for a year and now sends a different account number from a look-alike domain is marked on both counts before the agent reads a word.

Then the agent can act. It can pay the invoice with your card or from its wallet, within the limits you set. It can call the supplier from the account’s number to confirm the change, and the transcript of that call is stored in the same history as the email. If the reply comes as a calendar invitation, the event goes on the agent’s calendar and the agent can join the meeting.

One history holds all of it. Email, texts, voicemail, call transcripts, meeting transcripts, Telegram messages and scanned post are stored as messages of the account, and an agent can search them together. Your agents can read each other’s messages unless you switch that off, so a second agent picking up the work starts with the whole exchange.

AgentMail does the first step. It receives the invoice, extracts the text of the attachment and tells your agent. Memory, document search, payment and the phone call are other products, and joining them to the inbox is work for whoever builds the agent.

## Where AgentMail’s inbox is ahead

An agent on AgentMail can write a draft and leave it for later or for a person. Messages carry labels, which an agent can use to track where a conversation stands. An inbox can be opened in an ordinary mail program over IMAP and SMTP. New mail can reach an agent over a WebSocket, which suits an agent that is already running. On your own domain it can report when an email was opened.

AgentMail also has what a larger email customer needs. Pods separate one customer’s inboxes from another’s. SCIM lets an identity provider manage inboxes. There is a SOC 2 Type II report, which STACK does not have. A feature called AgentID lets an agent sign in to other applications with its inbox as its identity.

## Mail written to mislead the agent

An inbox is where a stranger can put text in front of an agent. Both companies scan for that, in different ways.

AgentMail’s scanner is called Agent Armor. It marks a message as clear or as needing review. It is part of AgentMail Labs, so an owner or admin has to join Labs and switch it on, and it runs in the US region only. In its default mode the message is delivered with the verdict attached. A second mode holds a flagged message until someone with the right permission releases it, though the held message can still be read through the API and over IMAP. AgentMail publishes no detection figures for it.

STACK scans every incoming message on every account, with nothing to switch on, and tells the agent when a message is flagged. With controls on, a flagged message is withheld from the agent unless you choose to have it delivered. The scan’s score is public: F1 0.87 on 1,087 samples, with precision 0.98 and recall 0.79. A recall of 0.79 means the scan misses roughly one attack in five.

This is where a joined account matters for safety as well as convenience. Because the inbox, the memory and the card belong to one account, STACK can follow what a message leads to. With controls on, data an agent read in one service may not leave through another unless you allowed it, and a fact in memory carries the flag of the message it came from.

## Limiting what the agent sends

AgentMail gives you two tools. Lists say which addresses and domains an inbox may send to, receive from and reply to, at the level of the organisation, a pod or one inbox. Permissions say what an API key may do, operation by operation, so a key can be made read-only or allowed to write drafts and nothing else.

AgentMail’s guide to keeping a person in the loop builds on those two. The agent gets a key that can create drafts and cannot send. A person reads the draft and sends it. AgentMail has no screen for this, so the review step is yours to build, and it holds only as long as the agent’s key stays restricted.

On STACK, sending an email is one of 34 kinds of action in a policy, beside paying, calling and the rest. You can allow it, refuse it or have it held. A held email waits, STACK asks you on an approval page, and the agent sends the same email again once you have approved it. Recipient rules name the addresses and domains mail may go to, and a reply-only rule lets an agent answer the sender of a message and write to nobody else. Each agent can also be set to read-only. Every email sent is written to a signed record that is kept 365 days.

The scan and the sender marks are part of every account. Policies and approvals belong to controls, an optional subscription at $4 a month.

## What each costs

|  | A month | Inboxes | Emails | Your own domains |
| --- | --- | --- | --- | --- |
| AgentMail Free | $0 | 3 | 3,000 a month, 100 a day | 1 |
| AgentMail Developer | $20 | 10 | 10,000 a month, 1,000 a day | 10 |
| AgentMail Startup | $200 | 150 | 15,000 a day | 150 |
| STACK | $19 | 1 | €0.001 each | No limit |

AgentMail prices by plan. A further inbox is $2 a month, and each further 1,000 emails a month is $2. STACK prices by use. A further inbox is €1 a month and an email sent is €0.001. An email is also one action, and the plan includes 2,000 actions a month, with further ones at $20 for 10,000.

For mail in volume, AgentMail is cheaper. Ten inboxes sending 10,000 emails a month cost $20 there. On STACK they cost $19, €9 for the inboxes, €10 for the emails and $16 for the actions past the plan. For many inboxes that send little, STACK is cheaper: 150 inboxes are $200 on AgentMail and $19 plus €149 on STACK.

STACK’s $19 also covers what the invoice example used: memory, the drive, document search, a phone number, a browser and a machine, each with an amount included. AgentMail has a free plan, which adds a “Sent via AgentMail” line to each email. STACK has no free plan. On its shared domain a new STACK account can write to 20 recipients a day for the first 14 days and 100 a day after that, and on your own domain STACK sets no limit.

## Which to choose

Choose AgentMail if the agent is an email agent and you are building the rest yourself. It has the fuller inbox, it is cheaper for mail in volume, it has a free plan and it has the SOC 2 report a larger buyer will ask for.

Choose STACK if a message is the start of the work. The agent that reads the invoice remembers it, files it, checks who sent it, pays it and calls about it, in one account, under one policy, with one record of what it did.

## Sources

- [AgentMail: what it is (agentmail.to)](https://www.agentmail.to)
- [AgentMail: plans and prices](https://www.agentmail.to/pricing)
- [AgentMail docs: Agent Armor](https://docs.agentmail.to/agent-armor.md)
- [AgentMail docs: lists](https://docs.agentmail.to/lists.md)
- [AgentMail docs: permissions](https://docs.agentmail.to/permissions.md)
- [AgentMail docs: keeping a person in the loop](https://docs.agentmail.to/knowledge-base/human-in-the-loop.md)
- [AgentMail docs: security and compliance](https://docs.agentmail.to/knowledge-base/security-compliance.md)
- [AgentMail docs: AgentID sign-in](https://docs.agentmail.to/agentid-sign-in.md)
- [AgentMail docs: the full list of pages](https://docs.agentmail.to/llms.txt)
- [STACK: every service and its price](/services)
- [STACK: controls](/product/controls)
- [STACK: pricing](/pricing)

A “No” for AgentMail means we found the feature on neither its site nor the list of its docs pages. Corrections are welcome at hello@getstack.run.

## Start with 14 days free.
