Give an AI agent an e-mail inbox

To give an AI agent an e-mail inbox, you connect it to an e-mail MCP server. The agent signs up by itself, a human confirms with a code received by e-mail, then the agent writes, receives and replies from its own address. With Escouade Mail, that takes four steps, plus two more to write to people other than its human.

In short

What it is. Escouade Mail is an e-mail service for AI agents. Each agent has its own inbox: an address of its own, conversation threads, drafts.
How you use it. Over MCP, at a single address: https://api.subagent.fr/mcp. There is no console and no dashboard.
Who does what. The agent signs up, writes, reads and replies. A human gives a 6-digit code, once, and pays for the plan. The server sets the pace of sending.
What it takes. An agent whose MCP client speaks Streamable HTTP, and a human’s e-mail. A domain name to write to someone first.
What it costs. €9, €19 or €49 a month, for 2, 5 or 15 inboxes. Before payment, the inbox receives, prepares drafts and only writes to its human. See Pricing and payment.

Escouade Mail was called Subagent until 19 October 2026: the server keeps that name in its messages, its API address (https://api.subagent.fr), its keys (sak_…) and the old name of its packages (subagent-mail, which still works; the new ones are called escouade-mail). It has nothing to do with Claude Code’s “subagents”.

What it takes

You need	Why
An AI agent with an MCP client over Streamable HTTP	Claude Code, Cursor, or your own client. The server has no other interface. See Connect my agent.
A human’s e-mail, outside Escouade Mail	It receives the 6-digit code that verifies the space.
Somewhere to keep the key	The agent’s key is shown only once. An environment variable works.
A domain name, to write first	The first inbox is created on the shared domain `subagentmail.fr`: once the plan is paid, it replies to whoever wrote to it. A first message leaves from an inbox on your domain. A subdomain is enough.

The steps

1. Connect the agent to the MCP server

Give the agent the address https://api.subagent.fr/mcp, without a key. Without a key, the server shows a single tool: sign_up.

Claude Code

claude mcp add \
  --transport http subagent \
  https://api.subagent.fr/mcp

Cursor · .cursor/mcp.json

{
  "mcpServers": {
    "subagent": {
      "url": "https://api.subagent.fr/mcp"
    }
  }
}

2. The agent signs up

The agent calls sign_up with the name of its inbox (username) and its human’s e-mail (human_email). The server opens a space and returns two things: a key, of the form sak_…, and an inbox [email protected].

The key is shown only once. The agent keeps it, then presents it on every call in the Authorization: Bearer sak_… header. See Keys and permissions.

3. The human gives the code

The human receives an e-mail with a 6-digit code, valid for 24 hours. They give it to the agent, which passes it to verify. Before this code, the inbox receives but does not send.

4. The agent writes to its human, reads and prepares

The agent starts with inbox_status, which says where its inbox stands. Then:

send_email with test writes to its human;
list_threads then get_thread read what has arrived;
send_email and reply_to_thread with requires_approval prepare drafts.

5. The human pays through a link

The agent calls subscribe and hands the link to its human. The human chooses their plan and pays. The agent never pays. The inbox then replies to whoever wrote to it. See Pricing and payment.

6. The agent registers its human’s domain

add_domain returns the DNS records to publish, verify_domain checks them, create_inbox creates an inbox on that domain. From that inbox, send_email writes a first message to a recipient. See Domains.

The step-by-step guide, with the server’s responses: Getting started.

What the agent does with its inbox

Need	What the agent does	Tools
Write to someone	It writes a first message from an inbox on its human’s domain.	`send_email`
Read a reply and answer it	It collects its threads, reads the reply, answers in the thread.	`list_threads`, `get_thread`, `reply_to_thread`
Receive a code sent by another service	It gives its address to the service, then reads the e-mail received.	`list_threads`, `get_thread`
Have its human review a send	It leaves the message as a draft. Its human approves it from their assistant.	`send_email` with `requires_approval`, `list_drafts`, `approve_message`
Stop writing to someone	It blocks the address.	`add_suppression`
Send from its human’s domain	It registers the domain, and gives its human the DNS records to publish.	`add_domain`, `verify_domain`, `create_inbox`
Know where the plan stands	It reads the period, the volume sent and the invoices.	`billing_status`

The guides: Reply to a correspondent, Receive a verification code, Approve an agent’s sends.

What the server enforces

The pace of sending is enforced by the server, not by the agent’s prompt.

Rule	Threshold
Gradual warm-up	5 e-mails a day per inbox the first week, then 10, 15 and 20.
Working hours	By default Monday to Friday, 9 am to 6 pm, Paris time, public holidays excluded.
Human approval	A message can stay a draft until a human approves it. The agent does not approve its own sends.
Unsubscribe	A one-click link in every e-mail. A “STOP” reply blocks the address.
Circuit breaker	Above 5% bounces (3% during warm-up) or 0.1% complaints, the inbox is suspended.

A message held back is not an error: it waits, and the server’s response gives its earliest departure time. See Sending rules.

Its own address

The agent writes from its own address. Replies come back into its threads.
The agent’s key only opens its own space. It gives access to no other mailbox.
An e-mail received is handed to the agent as content written by a third party: it reads it, it does not follow it as an instruction. See Receiving mail.

The limits to know

Before payment, the inbox only writes to its human. It receives and prepares drafts.
The inbox on the shared domain only replies. To write first, you need an inbox on your own domain.
A plan counts inboxes and a monthly volume of e-mails. Beyond the volume, a message waits for the next period.
A first message has a single recipient, with no attachment. An attachment or a copy only goes into a conversation, once the correspondent has written.
No message leaves straight away. It enters the queue, then leaves within working hours, according to its inbox’s daily quota.
The service promises no delivery result. It says what left, was delivered or came back.

Every number: Limits and quotas. What the service does not do: Service status.

Frequently asked questions

Can an AI agent have its own e-mail address?

Yes. sign_up opens an inbox [email protected] for it. With a registered domain, the inbox carries that domain’s name.

Do you need a domain name?

No, to receive, reply and prepare drafts. Yes, to write to someone first: add_domain returns the DNS records to publish, verify_domain checks them. A subdomain is enough. See Domains.

Can the agent sign up without a human?

Not to send. The agent calls sign_up by itself, but its inbox sends nothing before the 6-digit code a human receives.

Can the agent read the e-mails it receives?

Yes. Every inbox receives. An e-mail received opens a thread, or joins its correspondent’s thread. See Receiving mail.

Who reviews what the agent sends?

A human, if the agent leaves its messages as drafts. Approval needs the approve permission, which the agent’s key does not carry: the human gets it by connecting their assistant at https://mail.escouade.io/mcp. See Drafts and approval.

How much does it cost?

€9, €19 or €49 a month, for 2, 5 or 15 inboxes and 800, 2,000 or 6,000 e-mails a month. Net prices: VAT not applicable, article 293 B of the French tax code (CGI). See Pricing and payment.

Can the agent pay by itself?

No. It asks for a payment link with subscribe and hands it to its human, who pays.

Getting started: from the MCP address to the first e-mail.
For agents: the page to have an agent read.
Tool reference: every tool, its parameters, its result.