Email inboxes for AI agents that screen every message before the model reads it
Pick Postfleet when your agent reads mail from strangers and a hidden instruction in one email could make it misbehave. It is cheap to start and honest about its limits, but it is a one-person service from 2026, so larger teams with compliance reviews should weigh AgentMail.
Fail-closed injection screening with a public red-team ledger that lists every attack that got through and every real email it wrongly flagged.
The lock-in is the screening record and the mailboxes an agent already uses, but the REST and MCP surface is thin and standard, so switching away is easy.
Main trade-off
A very young, single-operator service without SOC 2 or an EU region.
| Plan | Price |
|---|---|
| FreeCloud · 100 comprehended messages, 3 mailboxes, 3,000 sends a month (up to 100 a day), REST and MCP · United States | $0/mo |
| StarterCloud · 250 comprehended messages, 5 mailboxes, 5,000 sends a month, email support · United States | $7/mo |
| ProCloud · 1,500 comprehended messages, 10 mailboxes, 10,000 sends a month, custom sending domains · United States | $20/mo |
| ScaleCloud · 25,000+ comprehended messages, 150 mailboxes, 150,000 sends a month, multiple domains, priority support and shared Slack · United States | $200/mo |
Free plan at $0 a month with no card: 100 comprehended messages, 3 mailboxes and 3,000 sends a month. Starter $7 a month (250 comprehended messages, 5 mailboxes, 5,000 sends). Pro $20 a month (1,500 comprehended messages, 10 mailboxes, 10,000 sends, custom sending domains). Scale $200 a month (25,000+ comprehended messages, 150 mailboxes, 150,000 sends, annual billing available). A comprehended message is one model injection scan plus schema extraction; cleaning and the deterministic injection check run on every message and are not billed. Pro has a 30-day money-back guarantee.
No traffic charge. Plans cap comprehended inbound messages and sends per month (Free 3,000 sends, up to 100 a day; Scale 150,000).
Postfleet is an email API for AI agents that gives each agent its own mailbox over REST and MCP. Before the agent reads a message, it strips hidden text, screens for prompt injection and returns JSON. Its published red-team test caught 115 of 116 attack emails.
Prompt-injection screening
Every inbound message runs a deterministic check and, within the plan's allowance, a model scan; flagged mail fails closed and is never passed on as trusted content.
Hidden-content sanitising
Strips hidden HTML, zero-width characters and homoglyphs from every message, with each removal logged in a trust report.
Schema extraction
Send a JSON Schema and get typed fields back with a classification and a confidence score, instead of a raw MIME payload.
MCP server, hosted or local
Tools such as create_mailbox, send_email, reply_email, list_inbox and wait_for_email, through the hosted endpoint or the MIT-licensed npx package.
Safe sending
Idempotent sends through a client id, up to 10 recipients per message, allow and block lists, and drafts that wait for a person's approval.
Webhooks with signatures
message.received, message.sent, message.bounced, message.complained and attachment.scanned events, HMAC-signed, retried at least once, with a dead-letter redrive.
Comprehend for an existing inbox
POST /api/v1/comprehend screens mail you already receive elsewhere, without creating a Postfleet mailbox.
| Service | Regions | Pricing | Compliance | Checked |
|---|---|---|---|---|
| US | Free tier · Monthly · Annual | -- | 22 Nov | |
| US · Europe | Free tier · Monthly · Annual | SOC 2 | 6 Nov | |
| Global | Monthly · Subscription | -- | 6 Nov | |
| Global | Pay as you go · Monthly | -- | 6 Nov |
Postfleet has four plans: Free at $0, Starter at $7 a month, Pro at $20 a month and Scale at $200 a month. Each plan caps comprehended messages, mailboxes and sends, from 100 messages, 3 mailboxes and 3,000 sends on Free up to 25,000+ messages, 150 mailboxes and 150,000 sends on Scale. Pro carries a 30-day money-back guarantee, and Scale can be billed annually.
Yes, the Free plan needs no card and includes 3 mailboxes and 100 messages a month that get the full model scan and extraction. After those 100, mail is still cleaned and checked deterministically, and sending is capped at 100 a day with tighter limits on new recipients.
It is one inbound message that runs the model stage: a prompt-injection scan followed by extraction to your JSON schema, with a classification and a confidence score. That pair counts once against the plan. The cleanup pass and the rule-based check are free and unlimited.
No. Its privacy page says mail is processed only to comprehend it for you, and that its model provider, Anthropic, does not train on API traffic. The original message is deleted after seven days and cleaned mail after ninety days by default.
Yes, through MCP. Point the client at the hosted Postfleet endpoint with a bearer key, or run the local package with npx, and the agent gets tools to create mailboxes, send, reply, list the inbox and wait for an email. Both routes use the same key and the same plan limits.
Postfleet gives an AI agent its own email address, so it can send, receive and reply over a REST API or the Model Context Protocol (MCP), the standard way agents plug into outside tools. Its focus is what happens to a message before the agent reads it: every inbound email is parsed, cleaned of hidden text, zero-width characters and look-alike letters, screened for prompt injection, and handed back as JSON, optionally extracted to a schema you supply. It is built and run by its founder as a sole proprietorship in Rosemount, Minnesota; the domain was registered in July 2026 and the first MCP release shipped the same month.
The main difference from AgentMail, the agent inbox Postfleet compares itself with on its own comparison page, is that screening is the product rather than an add-on. Flagged mail fails closed: it reaches the agent only as cleaned text with the flag attached, never as trusted content. The scan runs on Claude Haiku 5.5 from Anthropic, and Postfleet publishes a living red-team record of 116 attack emails, most written by models from other labs and a public benchmark, with every miss and false flag listed. Sending is idempotent through a client id, and a mailbox can be set to hold outgoing mail as drafts until a person approves them.
It fits developers wiring email into an agent that reads mail from strangers: support triage, invoice intake, sign-up and verification flows, or an assistant built in Claude Code or Cursor that has to wait for a reply. The wait_for_email tool blocks until a matching message lands instead of polling, and the MCP server runs hosted or locally through npx, with the same key and limits either way. Agent frameworks such as OpenClaw can connect it like any other MCP server.
It is a young, one-person service, and that is the main reason to look elsewhere. AgentMail is broader today, with IMAP and SMTP access, WebSockets, a Python SDK and a SOC 2 report on its Startup plan; Postfleet states plainly that it does not claim SOC 2 and keeps all data in the United States with no EU region. Teams that mostly send transactional mail and only need a webhook for replies are better served by Amazon SES or SendGrid Inbound Parse, which forward inbound mail to a webhook without screening its content.
Our call: if your agent will read email from people you do not control, Postfleet is the safer default to start on, and the Free plan is enough to wire it up. If you need enterprise paperwork or IMAP today, start with AgentMail. The field is compared in the best email APIs for AI agents guide.
Plans, regions, the uptime commitment and the compliance list on this page are read from Postfleet's own pages and documentation. The verdict is Hokai's reading of them, not the provider's claim.