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

How it works

# From inbox to sealed reply

Envolt Mail separates what an agent *wants* to do from what it's *allowed* to do, and keeps a person accountable for the difference.

Two roles have names of their own here. The [Director](https://envoltmail.com/docs/glossary#director) is the accountable person. A [Courier](https://envoltmail.com/docs/glossary#courier) is an agent working under that person. Everything below is about what passes between them.

[Running in staging] Built and running on our staging environment, on synthetic test mail only. [Designed] Specified in detail, not built or not switched on yet. [What runs today →](https://envoltmail.com/status)

1.

## Inbound mail is quarantined [Running in staging]

Every message is parsed in isolation and stored encrypted with its own key. HTML becomes plain text. Text a person couldn't see, whether hidden by styling, tucked in comments or written in invisible characters, is removed and counted, so agents only ever get a safe, text-first view. Links, attachments and hidden text are data. They can't carry instructions that change what an agent may do. [Watch it on a hostile email](https://envoltmail.com/lab#xray).
2.

## Rules triage before any model runs [Running in staging]

Duplicates, newsletters and bulk mail are recognized and filed by rules, without spending anything on a model. Only mail that needs thought goes further.
3.

## A quarantined reader describes the message [Running in staging]

A reader with no tools turns the message into a structured case: intent, urgency, a short summary, and any commitments or decisions in it. Card, bank account and US Social Security numbers are redacted before it reads anything. Simple detectors run first, and the reader can only add to their warnings, never clear them. If it misbehaves, the case fails closed to a person.
4.

## The agent drafts a plan, not a send [Designed]

An agent working the conversation will propose a plan: who to reply to, what to say, what to attach. It holds no password and no standing permission to send.
5.

## The Seal policy checks the plan [Running in staging]

A policy engine will check every plan against rules your team owns: who the recipients are and whether they're known, what's in the content, rate limits, and whether a person must approve. Anything unclear goes to a person by default.
6.

## A person decides, or the agent has earned it [Running in staging]

New agents draft for approval. Approving will take one tap from a queue that shows the recommendation, alternatives and risk ([see the tour](https://envoltmail.com/tour)). Over time an agent can earn the right to send routine replies on its own, from its measured track record. Any serious mistake takes that right away automatically.
7.

## The message goes out labeled and sealed [Running in staging]

Approved messages will be sent with a single-use capability, so one approval can never be reused. Each will carry the agent label, a signed Agent Seal header, a plain-language disclosure and a link to a public receipt. [What recipients will see](https://envoltmail.com/docs/recipients).
8.

## Everything is on the record [Running in staging]

Every event is written to a hash-chained ledger that is anchored daily. Before each anchor, recently active conversation threads are re-verified, and quiet ones on a rotation. A broken chain is sealed, recorded and never anchored, so you can prove what happened, and what didn't. [Try to tamper with one](https://envoltmail.com/lab#ledger).

Every step is designed to be explained after the fact: why a message went where it went, who decided, and what the record shows. The [console tour](https://envoltmail.com/tour) shows how that reads to the person in charge.
