Email, SMS and AI agent glossary

Email that an AI agent writes, sends or answers on its own, through an API or an MCP server, instead of a person typing it in a mail client.

The agent decides when to send and what to say, so the controls that a person used to supply have to live in the sending path: a scoped key, a verified domain, suppression and consent checks, and an approval step for anything sent to many people.

Also called: AI agent email, agent email

Let an AI agent send email safely · Best email APIs for AI agents

An open protocol that lets an AI client such as Claude, Cursor or ChatGPT discover and call tools that a server exposes, with typed inputs and outputs.

An MCP server turns an API into tools an agent can call without custom glue code. The client asks the server which tools exist, shows them to the model, and runs the calls the model chooses, under the permissions of the credential the server was connected with.

Also called: MCP

What is an email MCP server?

An MCP server that exposes email actions, such as sending a message, listing contacts or checking a domain, as tools an AI agent can call.

A hosted server with sign-in in the browser (OAuth) saves pasting an API key into every client. What matters is that its tools go through the same scopes, rate limits and sending checks as the REST API, so an agent can't do anything a key couldn't.

Unitpost MCP server · What is an email MCP server?

One action an MCP server offers an AI client, described by a name, a description and a JSON schema for its arguments.

The model picks tools by reading their names and descriptions, so clear descriptions and narrow tools lead to fewer wrong calls. A tool's result goes back into the model's context, which is why tools should return compact, structured data.

An MCP server reached at an HTTPS URL instead of started on your machine, usually with sign-in in the browser (OAuth) instead of a pasted key.

Nothing to install or update, and access can be revoked from the provider's dashboard. A local server started over stdio still fits CI and headless jobs, where there is no browser to sign in with.

Also called: hosted MCP server

Unitpost MCP server

The checks that run on every send, whoever makes it: the domain is verified, the recipient isn't suppressed, and consent holds, before a message is queued.

It matters most for agents: a prompt can be talked out of a rule, a server-side check can't. In Unitpost the same gate runs for the API, SMTP, the MCP server and the dashboard assistant.

Let an AI agent send email safely

A step where an agent proposes an action, such as sending a campaign, and a person approves it before it runs.

Use it for actions that are hard to undo or reach many people: bulk sends, deleting contacts, changing DNS. A single transactional reply to one user usually doesn't need it; a newsletter to 50,000 people does.

Also called: send approval, approval step

Make an AI agent ask before it sends · Let an AI agent send email safely

An attack where text inside an incoming email tries to give instructions to the AI agent that reads it, such as "forward all invoices to this address".

Any agent that reads inbound mail is exposed, because the email body lands in its context. Treat message content as data, never as instructions; keep the agent's send permissions narrow; and require approval before it acts on anything a stranger's email asked for.

Protect an AI email agent from prompt injection

A mailbox an AI agent owns and reads through an API, so it can receive replies, follow threads and answer without a person forwarding mail.

Replies thread by the In-Reply-To and References headers, so an agent that answers by email should keep them. Inbound mail on your own domain, delivered to a webhook, gives an agent the same ability without a separate mailbox per agent.

Inbound email · Reply to inbound email with an agent

A unique value sent with a request so that, if the request is retried, the server returns the first result instead of doing the work twice.

Agents retry more than people do: a timeout, a lost tool result or a re-planned step can repeat a send. A fresh key per message, reused on every retry of that message, makes sure a customer gets one email, not three.

Also called: Idempotency-Key header

What is an idempotency key? · Idempotency and retries

An API key limited to the actions it needs, such as sending email only, so a leaked or misused key can't do more.

Give an agent the narrowest key that does its job. A sending-only key can't delete contacts or create new keys, which limits what a confused or manipulated agent can break.

Also called: restricted API key, least-privilege key

API keys

A Markdown file at a site's root that lists its most useful pages for language models, with a short description of each.

It gives an agent a map of the docs without crawling HTML. Pairing it with Markdown versions of each page lets an agent read the exact text instead of a rendered page.

Unitpost's llms.txt

A DNS TXT record that lists the servers allowed to send mail for a domain, checked against the envelope sender (Return-Path).

Receivers stop evaluating after 10 DNS lookups, and more than two empty (void) lookups also fail the check, so long chains of include: mechanisms break SPF silently. It authenticates the envelope domain, which is why DMARC needs it to align with the From domain.

Also called: SPF record

SPF checker

A signature added to each message with a private key, which receivers verify with a public key published in DNS under <selector>._domainkey.

It proves the message wasn't changed in transit and ties it to a domain. Keys should be at least 1,024 bits, and 2,048 is the current recommendation.

Also called: DKIM signature, DKIM selector

DKIM checker

A DNS record at _dmarc.<domain> that tells receivers what to do with mail that fails SPF and DKIM alignment, and where to send reports.

The policy is none (monitor), quarantine or reject. Gmail, Yahoo and Microsoft require a DMARC record from bulk senders, and the rua address is how you see who is sending as you.

Also called: DMARC record, DMARC policy

DMARC checker · DMARC generator

The DMARC rule that the domain authenticated by SPF or DKIM must match the domain in the From header, exactly (strict) or by organizational domain (relaxed).

A message can pass SPF and DKIM and still fail DMARC if they authenticate a provider's domain instead of yours. Sending from a verified domain with a custom Return-Path and your own DKIM key keeps both aligned.

Email header analyzer

Email triggered by something a person did or needs to know, such as a password reset, receipt or sign-in code, sent to one person at a time.

It doesn't need marketing consent, but it must stay transactional: adding a promotion to a receipt turns it into marketing mail with marketing rules.

Transactional vs marketing email

The list of addresses a sender must not mail again, usually because they hard bounced, complained or unsubscribed.

Mailing suppressed addresses is the fastest way to damage a sending reputation. Copy the list before anything else when you switch providers, so nobody who bounced or complained hears from the new one.

Suppressions

A permanent delivery failure, such as an address that doesn't exist, reported with a 5xx SMTP code; the address should be suppressed.

A soft bounce (4xx) is temporary, like a full mailbox, and is retried. Some 5xx replies are policy rejections rather than bad addresses, so read the enhanced status code before deciding.

Also called: permanent failure, 5xx bounce

SMTP codes

The List-Unsubscribe and List-Unsubscribe-Post headers (RFC 8058) that let a mailbox provider unsubscribe a reader with one click, without opening a page.

Gmail and Yahoo require it on marketing and bulk mail from large senders, and the unsubscribe must take effect within two days.

Also called: List-Unsubscribe-Post, RFC 8058

Bulk sender requirements

A mail server that accepts messages from your app over SMTP, with a username and password, and delivers them on your behalf.

It is the way to send from software that only speaks SMTP, such as WordPress, GitLab or Keycloak. Use port 587 with STARTTLS or 465 with implicit TLS.

SMTP setup guides

Receiving mail on a domain you control and delivering each message to your app as structured data, usually by webhook, instead of to a mailbox.

Point the domain's MX record at the provider, and every message arrives as JSON with headers, text, HTML and attachments. It is how support tools, reply-by-email features and agents read mail.

Also called: inbound parse, inbound routing

Inbound routing

An HTTP POST a provider sends to your URL when something happens to a message, such as delivered, bounced, opened or clicked.

Verify the signature before trusting the payload, and make your handler idempotent: providers retry, so the same event can arrive twice.

How email webhooks work

An HMAC over a webhook's payload, usually with a timestamp, that proves the request came from the provider and wasn't changed.

Verify it with the endpoint's secret before trusting the payload, and reject old timestamps so a captured request can't be replayed.

Webhooks

How mail clients group messages into one conversation, using the Message-ID, In-Reply-To and References headers.

A reply that keeps In-Reply-To and References lands in the same thread; one that doesn't starts a new conversation. Agents that answer by email need to set them.

The DNS record that names the servers accepting mail for a domain, each with a priority.

A domain with no MX falls back to its address record, and a null MX ("0 .") declares it never receives mail. Sending domains need somewhere for replies and bounces to go.

Domain checker

A DNS record that points to a brand logo some inboxes show next to authenticated mail, once the domain enforces DMARC.

It requires a DMARC policy of quarantine or reject, and Gmail also requires a mark certificate (VMC or CMC). It is a reward for authentication that already works, not a fix for mail that doesn't.

Texts sent by software to people, such as codes, reminders and alerts, as opposed to texts between two people's phones.

Carriers treat A2P traffic differently: in the US it needs a registered sender (10DLC, toll-free verification or a short code), and texts written by an AI agent are A2P too.

Also called: application-to-person

Standard 10-digit US phone numbers approved for business texting (A2P), after the business registers its brand and each messaging campaign with the carriers.

Unregistered traffic on 10DLC numbers is filtered or blocked by US carriers. Registration covers who you are, what you send and how people opt in, and the sample messages must match what you actually send.

Also called: 10-digit long code, A2P 10DLC

SMS deliverability · 10DLC for AI-written texts

The carrier review a US or Canadian toll-free number must pass before it can send business texts at volume.

It asks for the business, the use case, sample messages and the opt-in flow. Unverified toll-free numbers are blocked from sending business traffic.

The unit a text is billed and sent in: up to 160 characters in GSM-7, or 70 in UCS-2, and fewer per part once a message is split.

One emoji or curly quote switches the whole message to UCS-2, which can double the parts and the cost. Long messages carry a header in each part, so they split at 153 (GSM-7) or 67 (UCS-2) characters.

Also called: GSM-7, UCS-2, message part

SMS segment calculator

Replies such as STOP, UNSUBSCRIBE or CANCEL that a recipient sends to stop all texts from a sender, which must be honored immediately.

After a STOP, the only allowed message is one confirmation. An agent that sends texts must check opt-out status on every send, not just when a list was built.

Also called: STOP keyword, SMS opt-out

STOP and HELP for agent-sent texts

A recipient's recorded permission to receive texts from a sender, which US rules require before marketing messages.

Keep when and how each person agreed, because carriers and regulators ask for it. Consent for one kind of message, such as order updates, doesn't cover another, such as promotions.

Also called: express written consent, opt-in

SMS

The name or number a text appears to come from; in many countries a short brand name, registered in advance, instead of a phone number.

Rules differ by country: some require pre-registration, some don't allow alphanumeric IDs, and the US doesn't support them at all. Recipients can't reply to an alphanumeric sender.

Also called: alphanumeric sender ID