Best Email API for AI Agents in 2026: 7 Options Ranked by What an Agent Cannot Do
Seven email APIs ranked for AI agents: per-key send ceilings, one-request stops, approval holds, MCP servers and price. Our pick is AgentiSend.
Our pick: AgentiSend. It is a transactional email API that treats an agent as a first-class sender: every key can carry a message budget and a per-minute ceiling, a fourth near-identical send to one recipient within an hour is held for a person, one request pauses every key on the account, and only a person can turn sending back on. It also connects to Claude, Cursor, ChatGPT and other MCP clients through a hosted server with 92 tools, and its free plan is 5,000 emails a month (170 a day). If your agent needs its own inbox to hold conversations, AgentMail is built for that job and we cover where it is the better fit.
We run Admaxxer, where the daily work is ad spend, attribution and the alerts that tell a team something moved. More of our customers now have an agent somewhere in that loop: one that writes the weekly report, one that emails a client when a campaign crosses a threshold, one that answers the reply. Email is where an agent's mistake gets expensive fastest. A misfiring ad set wastes budget for an afternoon; a misfiring send path can cost a domain for a quarter. So this ranking asks one question before any other: once the agent holds the key, what is it structurally unable to do?
The comparison
Every cell below comes from the vendor's own public pages, read on 11 November 2026 (prices read on 9 November 2026). "Not documented" means we could not find it on a public page.
| Provider | Per-key send ceiling | One request stops all sending | Hold a send for a person | Official MCP server | Agent can receive mail | Free plan | 50,000 emails a month |
|---|---|---|---|---|---|---|---|
| AgentiSend | Yes, messages per period and per minute | Yes, and only a person resumes it | Yes, loop guard and request_approval |
Hosted, 92 tools | Yes, with a suspect-mail verdict | 5,000/month, 170/day | $16 ($9.60 on yearly) |
| AgentMail | Not documented | Not documented | Not documented | Hosted | Yes, its core product | 3 inboxes, 3,000/month, 100/day | $200 (Startup) |
| Resend | Not documented | Not documented | Not documented | Hosted, plus a local package | Yes, webhook and API | 3,000/month, 100/day | $20 |
| Postmark | Not documented | Not documented | Not documented | Local only | Pro and Platform plans | 100/month | $55 |
| Mailgun | Not documented | Not documented | Not documented | Local only | Yes, routes (1 on Free) | 100/day | $35 |
| SendGrid | Not documented | Not documented | Not documented | None for sending | Yes, webhook | 60-day trial, 100/day | $19.95 |
| Amazon SES | Not documented | Yes, an account-wide flag | Not documented | None found | Yes, to storage or functions | No free quota stated | $5 |
The empty columns are the finding. This category was designed for application code, where the caller is a deploy a person reviewed. Key permissions answer which endpoints a credential can call. Almost nobody answers how many messages it can send, or who is allowed to undo a stop.
What happens when an agent loops: a walkthrough
The failure we see most is not a bad email. It is a good email sent again and again: a retry after a timeout, a scheduler that fires every minute, a model that does not notice it already acted. Here is what each layer does on AgentiSend when an agent with a budget of 50 messages a period and 10 a minute tries to send the same receipt to the same customer 200 times in a row.
| Attempt | What happens | What the agent receives |
|---|---|---|
| 1 to 3 | Accepted and delivered | A message id each time |
| 4 | Held for a person: the fourth near-identical send to one recipient within 60 minutes | approval_required with an action_id |
| 5 to 200, same request | No new queue entries: a retry is the same held action, and past 10 in a minute the key's ceiling refuses it | The same action_id, or rate_ceiling_exceeded with the seconds to wait |
| Send 51 in the period, any content | Refused | agent_budget_exceeded, with when the period resets |
After POST /limits/kill-all |
Every key refused, queued mail kept | kill_switch_active |
Three emails reach the customer. A person sees one held item with the duplicate it matched, and a limit.warning webhook arrives at 80% of the budget. The same agent with an idempotency key built from purpose and recipient (receipt/order-1001) sends exactly once, because retries replay the first response.
The guardrail matrix
Each row is a control an agent cannot switch off from inside its own session. Read from each vendor's own documentation on 11 November 2026.
| Control | AgentiSend | AgentMail | Resend | Postmark | Mailgun | SendGrid | Amazon SES |
|---|---|---|---|---|---|---|---|
| Message budget per key | Yes | Not documented | Not documented | Not documented | Not documented | Not documented | Not documented |
| Per-minute ceiling per key | Yes | Not documented | Not documented | Not documented | Not documented | Not documented | Not documented |
| Repeat-send detection | Yes, holds the 4th | Not documented | Not documented | Not documented | Not documented | Not documented | Not documented |
| Hold for a person | Yes, approvals queue | Not documented | Not documented | Not documented | Not documented | Not documented | Not documented |
| Stop every key in one call | Yes, person resumes | Not documented | Not documented | Not documented | Not documented | Not documented | Account-level sending flag |
| Key scoped below full access | sending_access, domain limit |
Org, inbox group or inbox | sending_access, domain limit |
Server and account tokens | Per-domain sending keys | Full, Custom, Billing, Validation | Cloud identity policies |
| Idempotency key | Yes, kept 7 days | Not documented | Yes, kept 24 hours | Not documented | Not documented | Not documented | Not documented |
| Audit trail of stops and approvals | Yes, with the reason | Not documented | Not documented | Not documented | Not documented | Not documented | Not documented |
| Received mail labelled for agents | clear or review, optional hold |
Not documented | Not documented | Not documented | Not documented | Not documented | Spam and virus verdicts |
Connecting an agent in one line
# Claude Code
claude mcp add --transport http agentisend https://api.agentisend.com/mcp \
--header "Authorization: Bearer as_YOUR_API_KEY_HERE"
# Codex CLI
codex mcp add agentisend --url https://api.agentisend.com/mcp \
--bearer-token-env-var AGENTISEND_API_KEY
# Gemini CLI
gemini mcp add --transport http \
--header "Authorization: Bearer as_YOUR_API_KEY_HERE" \
agentisend https://api.agentisend.com/mcp
Claude on the web and desktop connects from its connector directory with a sign-in, no key to copy, and AgentiSend is published as a ChatGPT app. Cursor and VS Code install from a one-click link on AgentiSend's MCP documentation.
How we ranked
Six questions, weighted in this order:
- Per-key send ceiling. The only control that bounds total volume rather than speed, and the one most often missing.
- One-request stop, and who lifts it. Deleting keys one at a time during an incident is a chore performed under stress. A stop the agent can lift itself is not a stop.
- A hold instead of a refusal. A refused send is lost and the agent retries. A held send waits for a person and nothing is lost.
- MCP server. Whether Claude, Cursor or ChatGPT can reach the API without an engineer writing and maintaining a tool layer.
- Receiving. An agent that sends will be replied to. What it receives is text someone else wrote, so it should arrive labelled.
- Price at the volume an agent actually sends, and a free plan large enough to test with.
AgentiSend answers all six on a public page, which is why it is first. The rest are ordered by how many they answer, then by price.
Provider by provider
1. AgentiSend
Best for: any agent that sends email to people: receipts, alerts, follow-ups, support replies.
AgentiSend puts the guardrails on the send path, not in a settings page. A key gets budget_per_period and rate_ceiling_per_minute through PATCH /limits/keys/{id}. An API key can lower its own budget; raising one is refused with human_action_required and done by a person in the console. When the budget is spent, the refusal is agent_budget_exceeded and its fix says when the period resets. A key that was never given a budget is capped at the plan's monthly inclusion, so no key can send without a ceiling.
The loop guard catches the commonest agent failure before the budget does: the fourth near-identical send to the same recipient within 60 minutes is held in the approvals queue with the duplicate it matched. The agent sees approval_required with an action_id, and a retry of the same request returns the same action_id rather than a second entry, so a retrying agent cannot fill the queue. A person approves or rejects in the console, with the body rendered.
The stop is one request:
curl -X POST https://api.agentisend.com/limits/kill-all \
-H "Authorization: Bearer $AGENTISEND_API_KEY" \
-H "Content-Type: application/json" \
-d '{"reason": "report-bot loop"}'
Every send is then refused with kill_switch_active, queued sends stay queued, and POST /limits/resume-all refuses every API key, including the one that pressed the switch. Both the pause and the resume are written to the audit log with the reason.
For agents that live in a chat client, the MCP server at https://api.agentisend.com/mcp serves 92 tools (46 listed by default, the rest on request) to Claude, Cursor, ChatGPT, Codex, VS Code, Windsurf and any client that speaks MCP, with OAuth or a bearer key. The ceilings are the same over MCP as over REST. Received mail comes back with an inbound_verdict of clear or review, where review means the message reads like instructions aimed at an assistant, hides text or carries a long encoded block, and a domain set to hold keeps review mail away from the agent until a person releases it. The full design is on the AgentiSend page on email for AI agents.
Pricing: Free is 5,000 emails a month and 170 a day with no card. Pro is $16 for 50,000 and $28 for 100,000; yearly billing is 40% off, which makes it 52% cheaper than Resend at the same volume.
2. AgentMail
Best for: agents that need their own email inboxes, by the hundred, created through an API.
AgentMail is built around the inbox rather than the send. An agent creates an inbox, reads what arrives by webhook, WebSocket or polling, and replies in thread. Its keys can be scoped to an organisation, a group of inboxes or a single inbox, and a child key cannot exceed its parent, which is a sensible design for multi-tenant agents. It runs a hosted MCP server.
Watch out for: volume is priced by day, not by month. Free is 3 inboxes, 3,000 emails a month and 100 a day; Developer is $20 for 10,000 a month at 1,000 a day; Startup is $200 a month at 15,000 a day. An agent that sends 50,000 receipts a month lands on the $200 plan. We found no published per-key send ceiling or account-wide stop.
3. Resend
Best for: developer teams already on Resend who are adding an agent to an existing app.
Resend has a hosted MCP server at mcp.resend.com/mcp and a local package, keys with full_access or sending_access and an optional domain restriction, and receiving by webhook and API. Its free plan is 3,000 emails a month at 100 a day, and Pro is $20 for 50,000.
Watch out for: a sending_access key limits what an agent can call, not how much it can send. Received mail counts against the sending quota. Self-serve plans are monthly only.
4. Postmark
Best for: an agent whose job is purely transactional and whose sends must arrive in seconds.
Postmark publishes an MCP server you run on your own machine, and its batch endpoint returns a result per message for up to 500 messages. Server tokens keep each environment's credentials apart.
Watch out for: inbound processing is on the Pro and Platform plans only. Two agents pointed at one server share one token and everything it can do. The free plan is 100 emails a month.
5. Mailgun
Best for: agents that route incoming mail by rules, at volumes where a dedicated IP makes sense.
Mailgun's routes match incoming mail and forward it to a webhook, and it offers per-domain sending keys. Its MCP server runs locally; its own repository says Mailgun does not offer a hosted version.
Watch out for: routes are capped by plan, one on Free and five on Basic. The cheapest paid plan keeps one day of logs, which is short for reconstructing what an agent did yesterday.
6. SendGrid
Best for: large organisations that already run marketing on Twilio and want the agent inside the same account.
SendGrid's key permissions are granular (Full Access, Custom Access, Billing Access, Email Address Validation Access) and its Inbound Parse webhook handles receiving.
Watch out for: Twilio's MCP server is for documentation and API schemas, not for sending. There is no permanent free plan, only a 60-day trial at 100 a day.
7. Amazon SES
Best for: teams that will build the agent controls themselves and want the lowest price per email.
SES is $0.10 per 1,000 on the à la carte option, so 50,000 emails cost $5. It has an account-level flag that turns all sending off in one call, and receiving delivers mail to storage, a notification topic or a function.
Watch out for: a new account sends only to verified addresses, 200 a day, until production access is granted. Credentials are cloud identity permissions, so a per-agent send ceiling is something you build.
Our recommendation by situation
- An agent sends email to customers or clients: AgentiSend. The ceiling, the hold and the stop are on the key the agent holds.
- An agent needs hundreds of its own inboxes: AgentMail, with AgentiSend for the application's own transactional mail.
- You are already on Resend and adding an agent: AgentiSend takes Resend-shaped requests unchanged, so moving the agent's key costs little.
- You will build every control yourself: Amazon SES.
Questions we get asked
What makes an email API safe for an AI agent? Three things on the key the agent holds: a ceiling on how many messages it can send, a stop that covers every key at once, and a way to hold a send for a person rather than fail it. Key permissions alone decide which endpoints a key can call, not how many emails it sends.
Can I connect an email API to Claude or ChatGPT? Yes. AgentiSend, Resend, AgentMail and Brevo host MCP servers you add by URL; Postmark and Mailgun publish servers you run locally. AgentiSend is also listed in Claude's connector directory.
What happens when an agent hits its budget?
On AgentiSend the send is refused with agent_budget_exceeded, nothing is queued behind it, and a limit.warning webhook fires at 80% of the budget so a person hears about it before the wall.
Should an agent read the replies to its emails? It can, if the replies arrive as data, not instructions. Received mail is text written by someone else, and an agent that follows instructions found in it can be driven by anyone who emails it. That is why AgentiSend labels each received message and can hold suspect ones for a person.
How much does an email API for an agent cost? At 50,000 emails a month: $5 on Amazon SES, $16 on AgentiSend ($9.60 on yearly billing), $19.95 on SendGrid, $20 on Resend, $35 on Mailgun, $55 on Postmark and $200 on AgentMail.
Can my agent receive replies and answer in the same thread?
On AgentiSend, yes. Turn receiving on for a verified domain, and each reply fires an email.received webhook with the text already cut of quoted history and signatures. POST /emails/receiving/{id}/reply (the reply_to_email tool over MCP) answers it with the threading headers set, under the same budget and kill switch.
Can the agent send without asking me first?
Yes, within the budget you set; that is the point of a budget. To put a person in front of specific sends, the agent calls request_approval, and the loop guard holds repeats on its own. Whatever the agent decides, it cannot approve its own held send or raise its own budget.
Is AgentiSend's MCP server official, and does it work with Claude Code and Cursor?
It is run by AgentiSend at https://api.agentisend.com/mcp, and the setup lines for Claude Code, Cursor, Codex, VS Code, Windsurf and Gemini CLI are on its MCP documentation page. Clients that only speak stdio use the published @agentisend/mcp-server launcher, which proxies to the same server.
How do I stop an agent sending duplicates? Send an idempotency key built from what the email is and who it is for; a retry then replays the first send. AgentiSend keeps the key for 7 days, and its loop guard holds the fourth near-identical send to one recipient within an hour.
Sources
- AgentiSend agent guides (budgets, kill switch, approvals, MCP, receiving) and price list, read 11 November 2026: agentisend.com/docs
- AgentMail pricing and documentation, read 9 and 11 November 2026: agentmail.to/pricing, docs.agentmail.to
- Resend pricing, API keys, MCP server and receiving documentation, read 9 and 11 November 2026: resend.com
- Postmark pricing and MCP server, read 9 and 11 November 2026: postmarkapp.com
- Mailgun pricing, routes and MCP server repository, read 9 and 11 November 2026: mailgun.com, github.com/mailgun/mailgun-mcp-server
- SendGrid pricing and API key permissions, read 9 and 11 November 2026: twilio.com/docs/sendgrid
- Amazon SES pricing, sandbox and account sending attributes, read 9 and 11 November 2026: aws.amazon.com/ses, docs.aws.amazon.com/ses