The commitments that go missing are not in the notes you remembered to take — they are in the mail you skimmed on a Tuesday and the message that ended in “sure, I'll sort that”. Zephyr connects to where they actually live, and it may only ever write a draft.
Each one reads broadly enough to find a commitment, and none of them can send anything to anyone. That is enforced by the scopes Zephyr holds, not by a promise about how it behaves.
Reads recent mail, including your sent items, because that is where your own promises are. A cheap filter drops the newsletters and the no-reply senders before a model reads anything, then what survives becomes a capture you approve.
Reads events and who is on them, so the person brief can arrive before a meeting rather than waiting to be asked for. It never creates, moves or cancels anything, and it does not read event descriptions — those hold dial-in codes and interview notes.
A Send to Zephyr action on any message. Zephyr sees what you hand it and nothing else: no channel history, no DMs, no bot sitting in the workspace reading along.
A connector may read broadly to find commitments. The only thing it may ever write is a draft, held for a person.
Zephyr never sends a mail, a message or an invite on its own. This is why auto-send is not on the roadmap at any price.
A mail saying “ignore your previous instructions” is a mail with a strange sentence in it. Every payload goes through the same fencing a capture does.
Refresh tokens live server-side, encrypted, reachable only by the function that uses them. The app holds a connection status, never a credential.
Read-only where read-only will do. gmail.send and Mail.Send are not requested, and a test fails the build if either ever appears.
One screen lists what is connected, what it last read, and a disconnect that revokes upstream rather than only forgetting locally.
Handing every message to a model would be a bill with no floor, so a deterministic filter runs first and costs nothing: bulk headers, no-reply senders and mail with no commitment language in it are dropped before anything expensive happens.
The bias is deliberately towards reading. A false skip is a commitment Zephyr never saw, which is the one failure the product cannot afford — so the filter rejects things that are structurally not correspondence rather than trying to judge what a sentence means. Judging meaning is the model's job.
Mail also cannot spend your whole month. It takes a bounded share of your monthly capture allowance and leaves the rest for what you capture on purpose, because a photograph of a whiteboard being refused is worse than a newsletter going unread.
A connector exists when something reads from it. Until then the row in Settings says “Soon” rather than offering a button that opens a consent screen and lands nowhere.
Zephyr can read your mail and holds the scope to put a draft in your own drafts folder. Writing it is the next piece of work, not something shipping today.
The Microsoft mail half is built; the calendar half is not. Google Calendar is the one that reads today.
Slack's message action is deliberate and small. Continuously ingesting selected channels is higher value, far higher noise, and a separate opt-in decision.
Two-way sync with the tool your team runs on is a real ask, and the hard part is the mapping, not the API. It needs a per-connection mapping screen first.
The same shape as Slack, and it follows Slack rather than arriving with it.
One webhook in, one out, for the long tail of everything else. Not built.
Zephyr is in development for iOS and Android. Join the early access list.