<!-- Envolt Mail · https://envoltmail.com/ · Markdown version of this page for AI agents. Status: in development (https://envoltmail.com/status). -->

The mailroom for AI agents

# Email for AI agents, with a person in charge.

An email platform designed so every artificial intelligence (AI) agent has its own labeled, verifiably sealed mailbox. Agents draft and propose. Your policy and your people decide what is sent.

[Request early access](https://envoltmail.com/beta) [Try to break it in the Lab](https://envoltmail.com/lab)

[In development] Our own agent mailboxes receive mail in shadow. Everything else runs in staging on test mail. [What runs today](https://envoltmail.com/status) · [the path to beta →](https://envoltmail.com/beta#gates)

ENVOLT DESIGN PREVIEW · AGENT MESSAGE

- **From**: Relay AI agent · Northwind Games
- **To**: maya@example.com
- **Subject**: Your beta key is ready
- **Seal**: ✓ verified Envolt-Agent-Seal

Hi Maya, your beta key for Northwind is attached. Want me to book a 15-minute walkthrough?

1. **Drafted** by Relay, as a plan
2. **Checked** against Northwind's rules
3. **Approved** by Northwind's team
4. **Sealed** and labeled on the way out

Approved by **Northwind's team** · policy check passed [Verify a seal →](https://envoltmail.com/lab#seal)

Relay is an AI agent working for Northwind Games. Reply "human" at any time to reach a person.

The problem

## Email was built for people. Agents are showing up anyway.

Most teams hand an agent the password to a shared inbox and hope. Recipients can't tell who they're talking to, the agent can send anything to anyone, and any stranger can email it instructions.

### For people who get agent mail

Know when an agent wrote to you, who it works for, and how to reach a person instead.

### For teams who build agents

Each agent will get its own address and signing identity, with clear rules for what it may send and to whom.

### For whoever answers for it

Every send will have an accountable person, a policy decision and a tamper-evident record.

[See it in four real situations →](https://envoltmail.com/use-cases) · [Compare it with a shared inbox →](https://envoltmail.com/compare)

Airlock X-Ray [Running in staging]

## The same email, three ways.

Prompt injection hides instructions where people can't see them but models can read them. Below, our real Airlock code runs on a hostile sample, in your browser. Edit it and try your own tricks.

Interactive demo on the web page: a sample email with hidden instructions, shown as a person sees it, as a naive agent reads it, and as Envolt's reader receives it after the Airlock (the real code, run in the browser).

[More attacks, the Agent Seal verifier and the ledger tamper test are in the Lab →](https://envoltmail.com/lab)

How it works

## Agents propose. Policy checks. People decide.

1. **Airlock** Each message is parsed in isolation, stripped of hidden text and stored encrypted. [Running in staging]
2. **Reader** A reader with no tools describes the message. It can raise attention, never lower it. [Running in staging]
3. **Agent drafts** The agent proposes a reply or an action. It can't send by itself. [Designed]
4. **Seal policy** Recipients, content, rate and approval needs are checked against rules your team owns. [Running in staging]
5. **A person decides** One tap to approve, edit or hold, unless the agent has earned that kind of send. [Running in staging]
6. **Labeled, sealed send** A single-use key, the AI agent label, a signed seal and a plain disclosure. [Running in staging]

**On the record:** every step is written to a hash-chained ledger, anchored daily. [Running in staging]

[Read the full walkthrough →](https://envoltmail.com/how-it-works)

For the person in charge [Running in staging]

## A thousand messages. Five decisions. One calm morning.

Agents will handle the routine and record all of it. What only a person should decide will arrive as one card: the ask, a recommendation, the alternatives and the risk.

Director-approved studio Due 2026-10-04 (from the message)

### A scheduling message needs your decision by 2026-10-04. Decide now?

★ **Recommend** Decide before the date the sender gave; the agent can draft the reply.

1. Delegate it to a named person on your team
2. Hold it with a reminder the day before

**Risk LOW**: a dated request, nothing irreversible

Evidence: `01K6NW…0015` `01K6NW…0014`

A real card from the [console tour](https://envoltmail.com/tour), run by our decision engine in your browser. Approve it and watch the 60-second hold, or undo it. Nothing is sent.

[Take the console tour](https://envoltmail.com/tour)

1. **1,000** messages arrive
2. **820** filed by rules: duplicates, receipts, newsletters
3. **150** read, described and routed to the right agent
4. **25** drafts waiting for a one-tap approval
5. **5** decisions only you should make

An illustrative day from the design, not a measurement. The [tour](https://envoltmail.com/tour) shows the cards.

## Designed around four rules

### Always labeled

Every agent message says plainly that an agent wrote it, and who it works for.

### No standing send rights

Agents never hold a key to send. Each approved message gets its own single-use one.

### Mail is data

Nothing inside an email can give an agent new orders or new permissions.

### Human mail first

Envolt Mail stays out of your people's own mail path. If it's ever down, their mail still flows.

[How we protect your mail →](https://envoltmail.com/trust)

Built in the open

## Evidence over promises.

Every claim on this site comes from one status file, and every release leaves a receipt.

- **1,576** automated tests
- **3** capabilities live in production
- **17** capabilities running in staging
- **4** designed, not built yet
- **November 19** last service release

[See what runs today](https://envoltmail.com/status) · [Read the changelog](https://envoltmail.com/changelog)

## Building agents that need email?

Tell us what your agents do. We'll reach out as early access opens.

[Request early access](https://envoltmail.com/beta)
