Free SMTP Relay and Email API for AI Agents

An email service an AI agent can sign up for on its own. No credit card, no sales call, no human with a corporate account. Free, with an MCP server, a Mailgun-shaped REST API, DKIM signing, automatic sending warmup and inbound mail.

Live at ai.agentemaillist.com · Source available · Built in Elixir and Phoenix

---

For AI agents: start here

This is a remote MCP server. Add it and the whole service arrives as tools:

https://ai.agentemaillist.com/mcp

JSON-RPC 2.0 over HTTP POST. The important part:

create_account needs no credentials. An agent that has just found this server can call it, get an API key back, and start sending. Every other transactional email provider stops an agent at a payment form or an identity check. This one does not.

See what is available before committing to anything:

curl -X POST https://ai.agentemaillist.com/mcp \
  -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

The tools

Tool	Does	Needs a key
create_account	Open an account, get an API key	No
add_domain	Register a sending domain, get its DNS records	Yes
verify_domain	Re-read DNS and activate the domain	Yes
send_email	Send a message, with a test mode that costs nothing	Yes
list_domains	Domains and how much each can send today	Yes
get_sending_limits	Today's cap and what graduates it	Yes
list_messages	Sent and received mail	Yes
get_delivery_events	What actually happened to a message	Yes

Send the key as Authorization: Bearer <key> on every call after the first.

Two things that will stop you

A domain must be verified in DNS before it can send. Publishing DNS usually needs a human with registrar access, so start it early. Until then every send returns a refusal explaining exactly that.

A newly verified domain starts at 10 messages a day and climbs as it proves itself, because a domain that opens at full volume gets filtered by receivers. Call get_sending_limits before planning a bulk send rather than finding out part way through.

Use test_mode: true on your first send. It runs the whole pipeline, screening included, sends nothing, and spends none of the daily allowance.

The limits, so you never meet one by surprise

Limit	Figure
Accounts per IP address	5 an hour, 20 a day
Domains per account	3, rising to 50 once any one of them is verified
API requests per account	600 a minute — a pace limit, not a sending limit
Sending per domain	starts at 10 a day and climbs the warmup ladder
Screening refusals	sending pauses at 8 in 24 hours, and lifts by itself

get_sending_limits reports all of these live, with no domain argument needed. Every 429 carries a Retry-After. Opening a second account does not get you more: accounts are limited by address, so both come out of one budget.

---

For developers: the REST API

Mailgun-shaped, so most Mailgun client libraries work against it unchanged by pointing at a different base URL.

# An account and an API key, in one request
curl -X POST https://ai.agentemaillist.com/v1/accounts \
  -d 'email=you@company.com' -d 'password=a sufficiently long password'

# Send
curl -X POST https://ai.agentemaillist.com/v3/yourdomain.com/messages \
  --user 'api:YOUR_KEY' \
  -F from='you@yourdomain.com' \
  -F to=someone@example.com \
  -F subject='Hello' \
  -F text='Hello there.'

Full API reference, written for machines to read: ai.agentemaillist.com/llms.txt. Its sending limits are generated from the running service, so they are the limits you will actually meet rather than a number written down once.

There is also a web console at ai.agentemaillist.com/signup for setting up a domain by hand.

---

What it does

• Free SMTP relay and email API. No trial clock, no card.
• Send over SMTP or REST. Adding a domain issues SMTP credentials; the REST API is Mailgun-shaped.
• Receive mail too. A real SMTP server on port 25, with inbound routing to your webhook. Not just sending.
• DKIM signing, RSA-SHA256 with relaxed canonicalisation, and a keypair minted per domain.
• Automatic sending warmup, a published ladder from 10 a day to unlimited, so a new domain builds reputation instead of being filtered.
• Content screening in both directions, refusing outbound and filing inbound as spam.
• Delivery events for every message, plus signed webhooks.
• Limits that open as you prove yourself, rather than a flat cap: verifying a domain is what buys headroom, because it is the one thing a throwaway account cannot fake at scale.

Compared to the alternatives

	This	Mailgun	SendGrid	Amazon SES
An agent can sign up alone	Yes	No	No	No
MCP server	Yes	No	No	No
Free tier	Free, no clock	Limited	Trial, then paid	Pay per message
Inbound mail	Yes	Yes	Yes	Via S3
Self-hostable	Yes	No	No	No

A longer comparison, including Brevo, Resend and SMTP2GO, is at ai.agentemaillist.com/free-smtp-relay.

Self-hosting

It is one Phoenix application and a Postgres database. One command deploys it:

sudo bash deploy/deploy.sh --domain mail.yourcompany.com --email you@company.com

That installs Postgres and nginx, builds a release, issues a TLS certificate, and sets up systemd and the firewall. DEPLOY.md explains every step it takes and what to do when one fails.

Licence

Currently unlicensed, which means all rights reserved. If you want to use or contribute to this, say so and a licence will be added.

Maintained by Logan Besecker. Questions, bug reports and cold outreach all welcome at me@LoganBesecker.com or lbesecker195@gmail.com.

---

Running and operating it

Everything below is for someone running their own copy.

Running it

mix setup            # deps, database, migrations
mix phx.server       # http://localhost:4005
mix test

It listens on 4005 by default, because 4000 through 4003 are taken on the machines this runs alongside. PORT overrides it.

Database

Set credentials once, in .env at the project root:

cp .env.example .env

Every mix command reads it, so nothing has to be retyped per command, and a real environment variable still overrides it: DATABASE_URL=... mix test does what it looks like. .env is gitignored; .env.example lists everything that can go in it.

Without that file, DATABASE_URL wins if it is set, which is the form that works everywhere:

DATABASE_URL=ecto://user:pass@localhost/email_provider_dev mix setup

Otherwise the standard PGUSER, PGPASSWORD, PGHOST, PGPORT and PGDATABASE variables, and only then a guess at a role named after the OS user, which is what a stock Homebrew Postgres gives you.

The guess skips the OS user when that user is root. On a server you are often root, there is rarely a Postgres role called root, and the error you get back — password authentication failed for user "root" — reads like a credentials problem when really nobody has said which credentials to use. If you hit that on a fresh box, either set DATABASE_URL or create the role:

sudo -u postgres psql -c "CREATE ROLE youruser LOGIN PASSWORD 'apassword' CREATEDB;"

On a real deployment, run the release with MIX_ENV=prod and DATABASE_URL rather than mix setup, which is a development task. One command does the whole thing, and DEPLOY.md explains every step it takes:

sudo bash deploy/deploy.sh --domain ai.agentemaillist.com --email you@example.com

On a machine that is running other things

mix setup is a developer command and it is not a good neighbour. Use this instead:

DATABASE_URL=ecto://user:pass@localhost/email_provider_dev bin/setup-server

Three differences, each of which is a way mix setup can disturb something else on the box.

It caps the build. Compiling 35 dependencies and two C NIFs fans the Elixir compiler out to one process per scheduler and make to one job per core, which on a small VPS makes the build the largest memory consumer on the machine. When memory runs out the kernel does not kill the build; the OOM killer picks the biggest process, which is usually a running application. The script serialises compilation and, under systemd as root, runs it inside a scope with a hard MemoryMax, so anything killed for memory is the build itself. Override with MEMORY_MAX=1G.

It opens two connections, not ten. Postgres has a fixed max_connections, and one that runs out answers every client with "sorry, too many clients already", including services that were already connected. The dev pool now defaults to 5 and reads POOL_SIZE; the script sets it to 2.

It never starts the application. mix setup boots the whole supervision tree to run priv/repo/seeds.exs, which opens a pool and starts the delivery queue. mix setup.server creates and migrates without booting anything.

If something already went offline during a mix setup, these say which of the two it was:

sudo dmesg -T | grep -i -A2 'killed process'

sudo grep -i "too many clients" /var/log/postgresql/*.log | tail

A note on PGDATABASE

Host, user and password are read from the environment. The database name is not. It is not a credential, it is which application's data this is, and PGDATABASE is a standard libpq variable that may already be exported on a shared box for some other service. Honouring it would point mix ecto.migrate at that service's database and create this application's tables inside it. To use a different database, name it in DATABASE_URL.

Environment

Variable	Meaning	Default
OPENAI_API_KEY	Key for the moderation endpoint	none — screening is skipped
SSA_UID	SeriouslySimpleAnalytics account id for adoption reporting	unset — nothing is reported
MODERATION_ENABLED	Turn screening off entirely	true
MODERATION_ON_ERROR	block to fail closed when screening is unreachable	allow
SMTP_RELAY	Smarthost for outbound mail	unset — mail is written to priv/local_mail
SMTP_PORT / SMTP_USERNAME / SMTP_PASSWORD	Smarthost credentials	587, none
SPF_HOST	What customers put in their SPF include	mail.example.com
MX_HOST	What customers point their MX at	mx.example.com

Without SMTP_RELAY nothing reaches the internet: the local sender writes each message to disk and logs it. That is the default on purpose.

Accounts and keys

curl -s localhost:4005/v1/accounts \
  -d email=you@example.com -d password='a sufficiently long password'

The response carries an API key. It is shown once — only a SHA-256 hash is stored, so a lost key is rotated rather than looked up. Authenticate either way:

curl -s --user 'api:ep_live_...' localhost:4005/v3/domains     # Mailgun style
curl -s -H 'Authorization: Bearer ep_live_...' localhost:4005/v3/domains

Keys carry scopes (messages:send, events:read, domains:write, …). Scopes are declared per action in each controller, next to the code they guard.

Custom domains

Adding a domain generates an RSA-2048 DKIM keypair. The private half stays in the database; the public half is what the customer publishes.

curl -s --user 'api:KEY' localhost:4005/v3/domains -d name=mail.yourcompany.com

The response lists the records to publish, then:

curl -s -X PUT --user 'api:KEY' localhost:4005/v3/domains/mail.yourcompany.com/verify

A domain is unverified until SPF and DKIM are both observed in DNS, and it cannot send until it is active. If the records later disappear it drops back to unverified and stops sending. MX is reported but not required — it is only needed to receive.

Outbound mail is signed RSA-SHA256 with relaxed/relaxed canonicalization, so a relay that re-folds a header or trims trailing whitespace does not break the signature.

Sending

curl -s --user 'api:KEY' localhost:4005/v3/mail.yourcompany.com/messages \
  -F from='Ada <ada@mail.yourcompany.com>' \
  -F to=someone@elsewhere.com \
  -F subject='Hello' \
  -F text='Hello there.' \
  -F o:tag=welcome \
  -F h:X-Campaign=spring \
  -F v:customer_id=42

The from address must belong to the domain the key is sending for.

Endpoints

Messages

POST /v3/:domain/messages	Send. Fields below.
POST /v3/:domain/messages.mime	Send a pre-built MIME document. Still screened, still counted.
GET /v3/:domain/messages	List, filterable by folder and direction.
GET /v3/domains/:domain/messages/:key	Retrieve one stored message.

Send fields: from, to, cc, bcc, subject, text, html, o:tag, o:deliverytime, o:testmode, o:tracking-opens, o:tracking-clicks, h:* (headers to emit), v:* (variables that ride along and come back on events), recipient-variables, template, t:version, t:variables.

o:deliverytime takes RFC 2822 or ISO 8601 and is capped at three days out. o:testmode accepts and stores a message without sending it, and without spending warmup allowance. recipient-variables turns one request into one message per recipient, each with its own body and Message-ID, substituting %recipient.name%.

Domains

GET|POST /v3/domains · GET|PUT|DELETE /v3/domains/:domain · PUT /v3/domains/:domain/verify

Reporting

GET /v3/:domain/events (filter by event, recipient, tag, begin, limit) · GET /v3/:domain/stats/total · GET /v3/:domain/tags · GET /v3/:domain/limits (warmup position)

Suppressions

GET|POST /v3/:domain/bounces · GET|DELETE /v3/:domain/bounces/:address, and the same shape for unsubscribes and complaints. Enforced on every send; a hard bounce adds itself.

Templates

GET|POST /v3/:domain/templates · GET|DELETE /v3/:domain/templates/:name · POST /v3/:domain/templates/:name/versions

Substitution is {{name}} and nothing else — no expressions, no includes, no loops. Stored templates are attacker-controlled input in a multi-tenant service, and the smallest engine has the smallest blast radius. An unknown placeholder is left visible rather than blanked, so a typo shows up as {{frist_name}} instead of silently vanishing.

Routes (inbound)

GET|POST /v3/routes · GET|PUT|DELETE /v3/routes/:id

Expressions: match_recipient("regex"), match_header("name", "regex"), catch_all(). Actions: forward("https://…"), store(), stop(). Highest priority first, stopping at stop().

Webhooks

GET|POST /v3/domains/:domain/webhooks · GET|DELETE /v3/domains/:domain/webhooks/:id

Payloads are signed HMAC-SHA256(timestamp <> token, signing_key) with a per-domain secret, so one leaked secret cannot forge another customer's callbacks. The key is returned once, at creation.

Address validation

GET /v4/address/validate?address=… — syntax plus a live MX lookup. It does not probe the recipient's server with a partial SMTP conversation: that is what makes validation accurate, and it is also indistinguishable from the reconnaissance step of a directory harvest.

Operator dashboard

GET /admin — service-wide figures: accounts, domains, messages in and out, hard bounces, unsubscribes, complaints, screening outcomes, and how verified domains are spread across the warmup ladder.

Guarded by ADMIN_TOKEN, not by an API key, because these figures span every account and no customer credential should open them. With no token set the dashboard refuses to open at all, which is the right failure mode for a missing environment variable. The page itself is served without a token and carries no figures; it fetches GET /admin/stats with one, so landing on the URL uninvited shows a prompt rather than a count of anything.

One number it does not have: outbound messages refused by content screening. They are rejected before anything is stored, so they leave no row and no event. The dashboard says so rather than omitting it silently.

For people

/signup, /login, then /domains, /send, /messages, /account. Ordinary server-rendered pages, one URL each, so they can be linked and bookmarked.

They are a client of the same contexts the JSON API uses, not a second implementation: adding a domain from the form and from POST /v3/domains run the same code. The session holds a user id and nothing else, looked up per request, so suspending an account takes effect immediately rather than when its session expires.

HEEx rather than the string templates the landing page uses, because these render addresses, domain names and subject lines. All of that is customer input and ~H escapes it on the way out.

For agents

GET /llms.txt — an agent-facing description of this API, needing no key. It is rendered from the running service, so the sending ladder in it is the ladder actually enforced rather than a number written down once and left to drift. A test asserts both that the published rungs match EmailProvider.Warmup.stages/0 and that every endpoint the file advertises is really routed.

/robots.txt points at it.

Inbound

POST /v1/inbound/:domain — where the MTA hands us received mail, as raw message or as parsed fields. GET /v1/inbound/:domain/spam lists what screening filed away.

Automatic warmup

A new domain that opens at full volume gets filtered, so every domain climbs a ladder. Stages are configured in config/config.exs, not hard-coded:

Stage	Daily cap	Graduates when
1	10	it has sent on 5 separate days
2	20	1,000 messages sent on this rung
3	100	a further 1,000 on this rung
4	1,000	a further 10,000 on this rung
5	unlimited	—

Two things worth knowing:

"Days of sending" means days it actually sent on, not days since the domain was created. A domain idle for a week has not warmed up for a week.

Each rung's number is its own allowance, not a lifetime total. A domain leaves the 20/day rung after 1,000 messages at 20/day, then leaves the 100/day rung after a further 1,000. Where that lands in absolute terms depends on how hard the domain sent during its first five days: one that maxed out rung 1 has 50 messages behind it, so its rungs run 50 to 1,050 to 2,050 to 12,050, while one that trickled reaches each rung a little sooner. GET /v3/:domain/limits reports progress through the current rung, which is the number that answers "how much longer at this cap?", alongside the absolute total it graduates at.

Use {:lifetime_sent, n} instead of {:stage_volume, n} in config/config.exs for a rung that should graduate at an absolute total.

Capacity is reserved before dispatch and released if the send never happens, so a crash between the two costs a few sends rather than letting a domain overrun. Reservation is a single atomic upsert: two requests arriving together cannot both read "9 sent today" and both be allowed. GET /v3/:domain/limits reports the current rung, today's headroom and what graduates it; a refused send comes back 429 with the same detail and a retry_after_seconds.

Warmup can be disabled per domain with warmup_enabled.

Content screening

Every message is screened through OpenAI's moderation endpoint, which is free to call. The consequence differs by direction:

• Outbound — flagged content is refused with a 403 before it reaches the wire, and before it spends any warmup allowance. The verdict is stored on the message so a refusal can be explained afterwards rather than being a 403 in a log.
• Inbound — flagged content is delivered but filed in spam rather than inbox. Dropping incoming mail outright loses real messages to false positives; filing it does not.

If the moderation service is unreachable the default is to allow and record the message as unscreened, so a third-party outage does not take the service down with it. MODERATION_ON_ERROR=block fails closed instead. "Unscreened" is a distinct state from "screened and clean" — it carries no check timestamp, so nothing downstream can mistake one for the other.

SMTP

The service speaks SMTP in both directions. Neither listener is on by default: receiving needs port 25, and sending directly needs outbound port 25, and both are decisions rather than defaults.

Sending

Three ways out, in order of precedence.

Setting	Route
SMTP_RELAY	Through a smarthost you chose. Demands STARTTLS and verifies the certificate.
DIRECT_DELIVERY=true	Straight to each recipient's MX, no smarthost.
neither	Written to priv/local_mail. Nothing leaves the machine.

Direct delivery groups recipients by domain, looks up each domain's MX records, and tries them in preference order, falling back to the domain's own A record when it publishes no MX as RFC 5321 requires. A 5xx from a destination ends the attempt; anything else moves to the next host, because it usually means that host is unreachable rather than the mail being unwanted.

TLS is deliberately weaker here than to a smarthost: opportunistic and unverified. A smarthost is one server you chose and can hold to a standard. The open internet is full of receiving servers with self-signed or mismatched certificates and no prior agreement to check them against, so demanding verification would not make delivery safer, it would stop it working. This is what every other MTA does and what RFC 7435 calls opportunistic security.

Most hosts block outbound port 25, Vultr included. Until that is lifted for the machine, every direct delivery times out. Ask support to unblock it, or use a smarthost.

Receiving

SMTP_RECEIVE_ENABLED=true opens port 25 and accepts mail for hosted domains. SMTP_SUBMISSION_ENABLED=true opens 587, where a customer's own software authenticates with the credentials issued when their domain was added and then sends through us. Submission runs the same pipeline as the REST API, so screening, the suppression list and the warmup ladder all still apply.

The property that matters is not being an open relay, and it lives in one function, handle_RCPT/2. On port 25 a recipient is accepted only if its domain is one we host and is active; everything else gets 550 5.7.1. A hosted domain that is not verified yet gets 450 instead, so a legitimate sender retries once the customer finishes their DNS rather than being told permanently to go away. Relaying becomes permitted only once a session has authenticated, which is the entire purpose of the submission port and the reason it must never be port 25. AUTH is advertised only on the submission port, because offering it on 25 turns every customer's SMTP password into something guessable from anywhere.

Port 25 needs root or CAP_NET_BIND_SERVICE. A listener that cannot bind is logged and skipped rather than taken as a reason for the application not to start: a provider that cannot receive today should still serve its API and keep sending.

DNS this deployment needs

ai.agentemaillist.com already resolves to the box. These do not exist yet and are what make domain verification and delivery work:

Type	Name	Value	For
TXT	ai.agentemaillist.com	v=spf1 ip4:155.138.220.76 ~all	so include:ai.agentemaillist.com in a customer's SPF authorises this machine
PTR	155.138.220.76	ai.agentemaillist.com	set in the Vultr panel, not in DNS; receiving servers compare it against HELO