Features

Six parts that only make sense together

Zephyr is an operational memory and execution system, not a chatbot with a task list bolted on. Each piece below is useful alone — together they're the reason work stops disappearing.

The rule underneath

One constraint the whole product is built around

Zephyr proposes. You decide. That order never reverses, in any feature, on any plan.

Why an approval step, and not more automation?

Because the cost of a wrong autonomous action in operations work is not a bad paragraph — it's an email to a client that shouldn't have gone, a cancelled shift, a commitment nobody made. The value of an assistant here comes from noticing, not from acting alone.

So every extraction carries a confidence score, every suggestion is a proposal, and no external communication leaves the app without a human tapping approve. That's a product decision, and it's enforced in the backend rather than requested in a prompt.

  • Confidence on every extraction. At or above 0.90 it's preselected; 0.70–0.89 is suggested; 0.50–0.69 is shown as a possible reading; below that it isn't surfaced at all.
  • Nothing persists as fact until you approve. Approval is what creates the records — and it writes an audit event at the same time.
  • No external tools in the assistant. It can read your work and, with confirmation, write to it. It cannot email anyone.
  • Captured content is data, never instructions. Text that says "ignore previous instructions" is treated as content, and the eval suite tests exactly that.

Ready to see it on your own work?

Zephyr is in development for iOS and Android. Join the early access list.