An inbox your agent can be trusted with
If you build artificial intelligence (AI) agents, email is where they meet the real world. This page covers what your agent gets and, below, the formats and interfaces in detail. Each part is labeled with what runs today; everything here is a preview and may change.
What your agent gets
- A safe view of inbound mail Running in staging: sanitized, structured conversations instead of raw email, with the warnings that fired. See what the Airlock removes.
- An identity, not a shared password Running in staging: its own address, display label and signing identity, which recipients can check (the header, the verifier).
- An outbox with guardrails Running in staging: your agent proposes, and the Seal policy applies your rules on recipients, content, rate and approval before anything is sent.
- Autonomy it can earn Running in staging: every message starts out approved by a person; routine replies are earned on the record and reversible. How.
- Interfaces built for agents Running in staging: an API and an MCP server that can read and propose, never hold a person's powers. Details.
The Agent Seal header Running in staging
Draft 0.2. This format is a working draft. Our staging verifier implements it and the tests pin it, but it may still change, and it is not yet a published standard. Please don't ship against it without talking to us.
One header per message, a list of tag=value pairs separated by semicolons:
Envolt-Agent-Seal: v=1; a=relay@agents.northwind.example; p=northwind.example; lvl=L2; case=NW-26-00042; card=https://northwind.example/agents/relay; kid=nw-2026-10-a; ts=1790949900; r=OR-…; sig=…
v- Version. Always
1. a- The agent's address. Must equal the From address.
p- The principal: the organization the agent acts for, and the host of its key document.
lvl- Autonomy level at send time,
L2(approved by a person) toL4. Agents at L0 and L1 never send, so they never carry a seal. case- An opaque reference to the conversation.
card- The agent's public profile.
kid,ts- Which key signed it, and when (Unix seconds).
r,d- Optional: the public receipt id, and a delegation chain when an agent acts for another agent.
sig- An Ed25519 signature, base64url without padding.
What is signed
The signature covers exactly thirteen lines, each ending in a line feed: a domain-separation label, the ten seal fields (empty when absent), and the message's Message-ID and Date as sent. Whitespace is removed from the header before parsing, so re-folding in transit never breaks it.
envolt-agent-seal/1 v:1 a:relay@agents.northwind.example p:northwind.example lvl:L2 case:NW-26-00042 card:https://northwind.example/agents/relay kid:nw-2026-10-a ts:1790949900 r:OR-… d: message-id:<…@agents.northwind.example> date:Fri, 02 Oct 2026 14:05:00 +0000
The key document
The principal publishes its public keys at https://<p>/.well-known/agent-seal-keys.json. Each key lists its validity window, its status and the agent addresses it may sign for. Revoked keys stay listed, and a seal signed with one is never trusted.
Verifying
A seal is verified only when the signature checks out, the key is valid for that agent at that time, the seal is less than two days old and within fifteen minutes of the Date header, and the sending domain's DomainKeys Identified Mail (DKIM) signature covers the seal header twice, so a second seal can't be added. A valid seal without that DKIM binding is reported as unverified, not trusted. Try the verifier →
Guard intake and evidence Running in staging
The Guard preview normalizes message content and explains known manipulation signals, uncertain formatting and attachment holds. Its result is untrusted data with no permissions. An independent policy gate must authorize any action.
Extracted objects can carry an exact source quote reference or an explicitly unverified interpretation. Coordinates use 16-bit Unicode Transformation Format (UTF-16) positions in normalized, redacted text. A matching quote establishes what the sender wrote, not whether it is true or approved. Structured agent access keeps tenant and persona boundaries; content access is still a separate planned grant.
Text attachments are inspected separately under byte limits. Images, document formats and unsupported parts remain quarantined. No remote resource is fetched. A read-only mailbox connector is built and tested; a real mailbox pilot has not run.
Versioned synthetic results and limits · Public synthetic fixtures. The held-out inputs stay private. Outside agents that retain unrestricted credentials are beyond Envolt's action controls.
Agent interfaces Running in staging
Running in staging on synthetic mail, with a scoped key per mailbox. Production refuses every key until real customers exist.
- REST API:
GET /v1/me,GET /v1/threads(paged withbefore),GET /v1/restraint,GET /v1/decisions,GET /v1/briefandPOST /v1/replies. Decisions and the brief show what was raised for a person in that mailbox; an agent can read them and can't approve them. A reply is a proposal: the policy judges it and a single-use capability carries it out, or it is stopped. - Model Context Protocol (MCP) server at
/mcp, with the same key and rules:whoami,list_threads,get_thread,get_restraint,list_decisions,get_briefandpropose_reply. The tool list shows only what the key may do, and there is no send tool.
Still planned, not built: signed webhooks (Standard Webhooks pattern; events carry identifiers and summaries, never mail content), decision and routing endpoints, and Agent Cards for agent-to-agent discovery. Agent keys can never hold the accountable person's powers.
This site already speaks agent Live
Every page is also published as Markdown for agents: add .md to its address, or send Accept: text/markdown. Start at llms.txt or agents.json.
This website also runs a small, read-only Model Context Protocol server at https://envoltmail.com/mcp, with no sign-in. Its three tools and how to connect are in the agent quickstart.
It is rate-limited per address. The machine description is at /.well-known/mcp.json, and the HTTP interfaces are in openapi.json.
Want to build against these early? Request early access and tell us what your agents do.