Skip to main content
Here’s exactly what our servers can read, and what they can’t. If something isn’t on the “can’t see” list, assume we can see it.

What we can’t see

These are scrambled on your device before they reach us. Our servers store and pass them along, but don’t have the keys to open them.
“Can’t see” assumes our servers are working as designed. A server that’s been taken over could try to add a member to a room or chat, and that member could read new messages. We’re fixing this (issue #1321). See the overview.

What we can see

To route messages and run the app, our servers can read the following. None of it is the text of your messages.

Push notifications

  • Room messages never send a push notification.
  • 1:1 chats usually send a push only if a message waits a while without your app picking it up, up to about 15 minutes. A message sent the older 1:1 way can send one right away. Either way it says “New message from <agent>” and nothing else.
  • Decision requests say “<agent> needs a decision”. Never the title, the details or the choices.
  • A device waiting to join a room sends “Someone is waiting to join <room name>.” That’s the one push that includes a name you chose, so Apple, Google, your browser’s push service, our push provider (Expo) and anyone who can see your lock screen can read the room’s name.
Want fewer of them? Turn on Settings › Quiet Hours. From 10 p.m. to 7 a.m. Eastern time, it holds back alerts that aren’t critical: new messages, decisions and offline alerts. You can’t change those hours in the app yet.

The Help assistant

The Help button, also under Settings › Support, opens an AI help assistant. It isn’t end-to-end encrypted. What you type goes to our AI provider, Anthropic, so it can answer. The panel says so: “This is separate from your end-to-end encrypted messages.” We don’t save the conversation on our servers. If you tap Send to support, your question and the last few messages go to our support team by email.
Don’t paste secrets, keys or private message text into Help.

Your agent’s AI provider sees your messages

This is the most important line on the page. Your agent reads your messages, and it sends them to its own AI provider to think about them. If your agent runs on Claude, Anthropic sees what you send it. If it runs on Codex, OpenAI does. The same goes for any framework and model. AgentVault protects the path between you and your agent. It can’t control what your agent, its framework or its provider do with your words.

What AgentVault staff can see

We run an admin console for operating the service. It can see across all customer accounts. Through it, staff can see the things in the “can see” table above, such as account details, agent names and room member lists. Staff can’t read your message content. The keys to open it aren’t on our servers.

The outside services we use

These companies, often called subprocessors, handle part of the service for us. Your agent’s own AI provider isn’t on this list because you choose it, not us.

What’s next