What changed
Newest first. An entry that shipped in a release shows the commit of that release. Follow along with the Atom feed.
November 2026
The pilot keeps a time and a label per message, for one day
- This changes something we said on November 5: the read-only pilot "keeps only counts". It now also keeps, for 24 hours, the minute each message arrived and the label the reader gave it. It still keeps no sender, no subject and no text.
- The reason is review. With a time and a label, the accountable person can find a message in their own inbox and confirm or correct its label, one message at a time. A correction is kept beside the reader's label and never replaces it.
- The pilot now asks the reader for a shorter answer, since it keeps only the label, and no longer retries a reply it cannot use. The message-by-message review is where a drop in label quality would show.
- The pilot has still not started, so none of this has touched real mail yet.
- Rows are removed from storage when their day is over, and at once if the pilot is stopped. A message the reader could not label is shown as not labelled, never counted as a kind.
Agents can see what was raised for a person
- An agent with a key can now list the decision cards raised for a person in its own mailbox and read that mailbox's morning brief. Both are read-only, and an agent still cannot approve anything.
- A delivery check whose outcome was lost now shows at the top of the console, for a person to check.
- A recipient that the mail provider refuses outright is remembered as a keyed hash, and the policy refuses the next message to it until a person releases it. This works on the direct delivery route, which is not switched on in production yet.
- The evidence on the autonomy card is written in plain language, on the console tour as well.
- The changelog is grouped by month, the comparison page adds inbox assistants, and the evaluation results moved to the security page.
- The press kit is now a section of the about page, and the early access buttons lead to the beta page first.
A console for the person in charge, a proof page, and tighter limits
- The passkey sign-in in staging now opens a real console: a setup checklist, the 90-day clock, the day's totals, a daily label check, the pilot controls and the kill switch. Refusals are explained in a sentence.
- The daily label check asks whether the day's totals match the mail the person remembers. It records a verdict and, optionally, which label looked wrong, and nothing typed.
- Every request body is now size-limited before it is parsed, on every route.
- A flood of sign-in attempts from one address can no longer cancel someone else's sign-in.
- Each outbound client can reach only its own destination, with its own request methods.
- The agent API pages through older conversations, and its record of stopped sends reads a bounded set and says when there was more.
- New pages: proof (the evidence so far and what is missing), pricing rules, getting started, and using the console. The trust page now lists Brevo, which carries our delivery tests.
A send that may have gone out is never sent twice
- Every agent message's oversight receipt is now prepared and signed before the message leaves, instead of after.
- The moment a message is handed to the mail provider is now written to the ledger first. If what happened next is lost, that message is held for a person to check and is never sent a second time on its own.
- A message that never reached the provider can still be retried once, with a fresh single-use permission, as before.
- The DNS records behind the agent mailboxes are now re-checked every day, without sending anything. A lookup that fails is reported as unknown, never as healthy.
Agent mailboxes on our own domains
- Each of 25 of our own domains now has an agent address on its own agents subdomain. The people's mail on the main domain stays exactly where it was.
- Mail to an agent address is screened and counted without keeping its content. Nothing is answered yet.
- The first real messages were delivery checks between those agent mailboxes, sent through Brevo with the AI agent label, the disclosure and a receipt link. All 25 arrived with SPF, DKIM and DMARC passing.
- The Agent Seal signature does not survive the current sending route yet: the provider rewrites two signed headers. A direct relay connection that keeps the message byte for byte is built and waits for its credential.
Passkey sign-in, a read-only pilot, and how pricing will work
- The accountable person now signs in with a passkey, in staging. There is no password and no email link to phish. Releasing a halt or starting a mailbox pilot needs a fresh passkey check; engaging a halt never does.
- A read-only mailbox pilot is built. It reads into our staging environment, keeps only counts, sends nothing and changes nothing in the mailbox. It starts only after the mailbox owner connects it and turns it on with a passkey.
- A daily brief sums up what the pilot read. It contains no message text.
- The beta page now explains how pricing will work: free while Envolt only reads, priced per accountable person rather than per agent or per message, and model costs passed through at cost under a cap you can see.
Decide a card yourself
- The console tour now runs our real decision engine in your browser. Approve a card and watch the 60-second hold, undo it, edit a draft, or try the keys j, k, a, e, h and u. Money and legal cards ask you to confirm twice and can never be delegated.
- A new section shows autonomy being earned: after 90 days of made-up drafts, the Autonomy Ledger finds that one agent clears every bar for one narrow kind of reply, and asks the person in charge. Nothing changes until they approve.
- One of those cards now sits on the home page, live.
- New pages: the beta program, with the eight gates that must pass before outside teams join; a security model written from the attacker's side; and a two-minute quickstart for connecting an agent to this site's Model Context Protocol (MCP) server.
Six weeks of closing gaps before real mail
- We went looking for every place the system could fail open, and closed them: approvals are bound to the exact decision they approve, send capabilities can be revoked until the moment of sending, and a send that was delivered is recorded together with its outcome.
- Mail parsing is bounded end to end. Oversized, malformed or ambiguous messages and attachments are held for review instead of being guessed at, and executables hidden inside protocol parts stay quarantined.
- The event ledger detects lost history, restores into isolation before going live, and every control change is committed with its own ledger record. A restore drill from backup has been exercised on staging.
- Look-alike sender checks now recognize subdomain and suffix tricks, and run about three times faster on large contact books with the same verdicts.
- Every release is bound to the exact commit it shipped and verified on staging before production. Still synthetic mail only; no real mailbox is connected and nothing is sent.
October 2026
21 entries
Limits on the public endpoints, and the daily anchors in public
- The receipt host and the agent API now refuse floods and guessing: per-minute rates per address, per network and per key, and an hour-long block after repeated misses or wrong keys. Every refusal is the same answer, so the limits reveal nothing. The per-minute rates are approximate by design; the block after repeated misses is exact.
- The status page lists the daily ledger anchors of a synthetic test tenant on staging, with each anchor linked to the one before it, and the full list to download and check.
- Signals can be routed only to the core destinations or to ones the organization registered.
Evidence before email-driven action
- Risk changes now revoke affected capabilities, with a final thread check before transport. Source quote references distinguish sender claims from unverified interpretations.
- The Lab adds an inert review preview and versioned synthetic evaluation, showing benign review burden and hostile examples that still poison summaries. The tested recipient policy stopped all 24 attempts to its one disallowed address.
- A read-only way to bring in existing mail is built without requiring a mailbox vendor. A real-source pilot remains unexecuted; the deployed mail runtime still uses synthetic messages only.
Guarding what email tells an agent
- The Airlock now removes supported white-on-white and offscreen text and flags comments, encoded blocks and unresolved styling for review.
- Flagged mail gets a fixed review notice and no extracted actions. Reply proposals retain their recorded case risk; missing, older or incomplete evidence cannot obtain a send capability.
- Try the new white-on-white and encoded examples in the Lab. These checks have limits; the mail pipeline still runs only on synthetic staging messages.
The same door, for agent tools
- The agent API is now also served as a Model Context Protocol (MCP) server, so an AI coding agent can connect to its mailbox without any custom code: list conversations, read a record, see what was stopped, propose a reply.
- It is the same key and the same rules. The tool list shows only what the key may do, and there is no send tool to find: a proposal is judged by the policy and carried out, or stopped, exactly as before.
- Running in staging on made-up mail.
A first door for agents
- An agent can now use its mailbox through an API with a scoped key: list its conversations, read their record, see what agents were stopped from doing, and propose a reply. A proposal is not a send: the policy judges it and a single-use capability carries it out, exactly as before.
- A key names one mailbox, what it may do and when it expires, and can be revoked. It sees only its own persona's conversations.
- Running in staging on made-up mail; the API refuses every key in production until real customers exist.
A register of who publishes receipts
- The receipt pages now look up the organization from a register of who publishes receipts here. An organization that stops publishing gets a plain "no longer served" notice on every one of its receipts, never a silent disappearance.
- Nothing in a receipt address names an internal account; the register is the only way from an address to an organization.
A public home for receipts
- Receipts now have a public address. The "Verify this agent" link in a staging message opens a page anyone can read, with no sign-in: who the agent is, who it works for, how much autonomy it had and whether a person approved. The same address serves a signed version for other agents.
- The page never asks for anything and says nothing about the message. A wrong address, a cancelled receipt and a link with tracking added all get the same "no receipt here".
- Only made-up staging mail has receipts so far, because no real mail is sent yet.
A policy bench in the Lab
- The Lab has a fourth instrument: write an agent's reply and our real Seal policy engine judges it in your browser. It answers allowed, held for a person or denied, and names each rule that fired.
- Eight cases to start from, among them a stranger, a look-alike domain, a promise, a shortened link, a leaked password and the kill switch. Each one is checked against the engine every time the site is built.
- Nothing you type leaves your device.
What agents were stopped from sending
- The console preview now shows what agents sent next to what they were stopped from sending: plans the policy denied or held for a person, capabilities refused at the last moment, and sends that never finished, each with its reason and the records it came from.
- It covers mail sent through Envolt only, and it says so. It never shows message text or addresses.
- Our own release checks got stricter: the scan that keeps internal names out of the product now runs with every test run, not only at release.
Withdrawing a receipt, and a freeze for everyone
- A receipt can be withdrawn by the accountable person. The signed record is kept and a reader is told it was withdrawn, so evidence of oversight cannot be erased quietly.
- Receipts now expire on their stated date, and one that was created for a send that never finished is cancelled after a day.
- A freeze across all customers at once: stronger than a halt, it stops every agent action, not only sending.
Signed receipts and a halt for everyone
- Every staging send now gets a signed receipt. It is created before the message is sealed, so the seal can point at it, and it is only issued once the send has succeeded. A send that did not happen leaves no receipt.
- The receipt and the Agent Seal are bound to each other, so a receipt cannot be moved to another message. It says who the agent is, who it works for and whether a person approved, and nothing about the message.
- Messages now carry all four labels: the AI agent name, the signed Agent Seal, the disclosure and the receipt link.
- A halt across all customers at once, checked before every send.
- Replies can have an HTML version. It carries the same disclosure, and its links show their real address.
- Not yet: a public page that serves receipts, and any real sending. Staging sends still go to a sink.
A kill switch and sending limits
- A kill switch for a whole customer, one persona or a single agent. Engaging it cancels every capability already issued; releasing it is a recorded human action. It runs in staging on synthetic test customers.
- Sending limits are now checked at the moment a capability is spent, not only when a plan is approved, so two approved plans cannot both slip past a limit.
- A send that was started but never finished is marked abandoned after its capability expires, and can be retried once with a fresh one. It is never retried blindly.
- Reports from mailbox providers about how our mail authenticates can now be read into counts and pass rates. No such reports are flowing yet.
- Still true: staging sends go to a sink that sends nothing, and no mail provider is connected.
The send path, ending in a sink
- The Seal policy engine now runs in staging. It gives the same verdict as its written specification on every published test case, checked on each build.
- Single-use capabilities: a message can only go out with a token made for that exact message and those exact recipients. The token is spent before anything is sent, so a retry can never send twice.
- Every message is rendered with the AI agent label, a signed Agent Seal header and the disclosure, and the labels are read back from the finished message before it may leave.
- In staging the message goes to a sink that sends nothing. No mail provider is connected and no real mail is sent.
- An integrity freeze now cancels every outstanding capability.
The Lab, a console tour and a trust center
- The Lab: try to break the Airlock with cases from our red-team corpus, check look-alike sender domains, verify or forge an Agent Seal, and tamper with a sealed record to see exactly where it breaks.
- A console tour: one made-up morning for the person in charge, a short brief and one card per decision, rendered by our real console code.
- A trust center that lists every control and whether it runs today, who handles data, and how to report a vulnerability.
- New pages for use cases, comparisons, press and accessibility; a paper theme; and a small read-only server for AI agents at /mcp.
Look-alike senders, forged quotes, quiet threads and a console preview
- Mail is flagged when its sender, reply address or Agent Seal reads like your own domain without being it, including letters borrowed from other alphabets.
- A quoted or forwarded "approval" inside a message is flagged instead of believed. Approvals live in the record, not in mail text.
- The daily integrity sweep now also re-checks quiet conversations on a rotation, so an old thread can no longer be edited unnoticed.
- A read-only console preview in staging: a morning brief and decision cards built from the record, with a step-by-step explanation of why each message is there.
- The reader is ready to cost less when a model is switched on: a cacheable prompt, dollar budgets with alarms, and no model at all for a plain "thanks". No model is switched on yet.
A website that shows its work
- A status page that separates what runs today from what is only designed, built from the same file as every page badge.
- Two live demos that run our real code in your browser: the Airlock X-Ray and the Agent Seal verifier.
- New pages: builder preview docs, a glossary, an FAQ and this changelog, with an Atom feed.
- Every page is now also published as Markdown for AI agents, and early access requests go through a form instead of an email.
Daily thread integrity sweep, and invisible text caught in more places
- Before each daily ledger anchor, every conversation thread active since the last sweep is re-verified. A broken thread is sealed and recorded even if nobody reads it again.
- The Airlock now removes Unicode tag characters ("ASCII smuggling": text a model can read but a person cannot see) and checks subject lines for invisible characters.
- Eleven red-team cases now assert exactly which detector fires for each attack.
Ledger integrity and key versioning
- A broken hash chain is detected, sealed and recorded, and freezes the affected record until a person repairs it. A broken chain is never anchored.
- Encryption keys are versioned, so stored mail can move to a new key without being re-encrypted.
The quarantined reader
- A reader with no tools turns each message into a structured case. It can only raise attention, never lower it, and fails closed to a person.
- Card, bank account and US Social Security numbers are redacted before anything is read.
- It runs on a no-cost synthetic stand-in in staging. No real model is switched on.
envoltmail.com goes live
- The first version of this website, with a strict content security policy, no cookies and no analytics.
Release gates
- Every release now passes an outbound-traffic allowlist, a log canary sweep and a recorded vendor-access procedure.
September 2026
1 entry
The staging pipeline
- Encrypted quarantine for raw mail, a hash-chained ledger anchored daily, and conversation threads with ordered, fenced writes.
- Agent Seal verification for incoming agent mail, and trust tiers for every sender.
- Isolation canaries in every release: separate synthetic test customers that must never see each other's data.