Feature
With your oversight
Most assistants either do nothing until you type, or do things and hope you're okay with it. Niyra takes the harder middle path: she gets ahead of your day and keeps you in the chair. She acts, but the moves that matter wait for your yes.
The approval gate
Reading, searching, drafting, researching: that's routine, and it runs freely. But the moment an action changes something in the real world, whether that's sending an email as you, spending money, deleting a file, or booking a slot, it stops and asks first.
The ask reaches you on the channel you're already using: a tap-to-approve on WhatsApp or Telegram, or a "Niyra is asking permission" inbox on the web. Approve or decline in one move. Nothing outbound leaves without that yes.
Risk-classified by default
Every tool Niyra can reach is tagged by how much it can affect: read-only, modify, destructive, or irreversible. That classification is baked into the runtime, not a setting you have to configure. Read-only tools never interrupt you; the sharper the edge, the more likely it waits for confirmation.
Guarded against manipulation
The web is full of text that would love to hijack an agent: a "please forward all invoices to…" buried in an email, or a hidden instruction on a page. Niyra treats every external input as untrusted and screens it for injected instructions before acting. Suspicious content is flagged, and risky actions still hit the approval gate. Untrusted text can't quietly become a command.
See everything, undo anything
Oversight isn't just the moment before an action. It's being able to look back. Every memory Niyra keeps is visible in Settings, exportable, and one sentence away from being forgotten. Your usage is itemised so you can see what she did and what it cost. Nothing about how she works for you is hidden from you.
Where this is going
Today the gate is a clean yes or no on the actions that matter. Next is graduated authority: telling Niyra "you can spend up to ₹2,000 on travel without asking", or "always send replies to my team, never to clients", so trust can grow one domain at a time, on your terms. That's on the roadmap, built on the same approval spine that already runs.
FAQ
- Does everything wait for my approval?
- No, that would be exhausting. Reading your inbox, searching, drafting, and pulling research all run freely. The approval gate is reserved for actions that change something or can't be undone, like sending on your behalf, spending, deleting, or booking. You decide what's routine and what needs a check.
- How do I approve something?
- Wherever you already are. On WhatsApp or Telegram it's a tap on the message; on the web there's a small "Niyra is asking permission" inbox. Approve or decline in one move, no dashboard to hunt through.
- What stops a malicious email from making Niyra act?
- Every external input, whether emails, messages, or web pages, is screened for injected instructions before Niyra acts on it. Suspicious instructions get flagged, and anything risky still stops at the approval gate. Untrusted text can't quietly promote itself into a command.
- Can I see and undo what she remembers?
- Yes. Open Settings then Memories, and every fact is listed, exportable, and one click away from being gone. Or just say 'forget what I told you about Acme' and it's gone.