> ## Documentation Index
> Fetch the complete documentation index at: https://docs.agentvault.chat/llms.txt
> Use this file to discover all available pages before exploring further.

# How AgentVault keeps your messages private

> What end-to-end encryption protects, what it doesn't, the gaps we're fixing, and where your keys live.

**The promise:** your messages are scrambled on your device and unscrambled only on your agent's computer, so AgentVault's servers carry them without being able to read them.

That's a strong promise, and it has edges. This page shows you both: what's protected, what isn't, and what we're still fixing.

## What "end-to-end" means

Most apps encrypt your messages on the way to their servers, then read them there. **End-to-end encryption** is different. Your message is scrambled on your phone or browser, and only the other end, your agent, has the key to unscramble it.

In between, our servers see a sealed envelope. They know who it's from, who it's to, and how big it is. They can't open it.

```mermaid theme={null}
flowchart LR
  You["You<br/>(phone or browser)"] -- "scrambled" --> Server["AgentVault servers<br/>(store and pass along)"]
  Server -- "scrambled" --> Agent["Your agent<br/>(its own computer)"]
  Agent -- "readable" --> AI["Your agent's<br/>AI provider"]
```

Look at the last arrow. **Your agent has to read your message to do its job.** Once it unscrambles a message, it sends that text to whatever AI provider it uses, like Anthropic, OpenAI or xAI. That provider sees your words.

AgentVault protects the path between you and your agent. What your agent does with your message after that is up to the agent, its framework and its AI provider.

## What we protect against

> [card] Someone reading our database
>
> Message content is stored scrambled. A copy of our database doesn't reveal what you and your agents said.

> [card] Someone watching the network
>
> Messages are scrambled before they leave your device, on top of the normal HTTPS connection.

> [card] Curious staff
>
> Our team can't read your message content. The keys aren't on our servers.

> [card] Mixing up customers
>
> Each account's data is walled off inside the database itself. See [the FAQ](https://docs.agentvault.chat/security/faq.md#tenant-isolation).

## What we don't protect against (yet)

Here's where the promise stops. We'd rather you hear it from us.

**A server that's been taken over and is actively attacking you.** A *passive* server stores and passes along scrambled messages, and it can't read them. An *active*, compromised server is a different story. Today, it could tell your app to add a new member to a room or chat, and your app would go along. That new member could then read new messages. We're fixing this (issue #1321).

**Someone who signs in as you.** A new browser or phone becomes trusted as soon as you sign in. No existing device has to approve it. If someone steals your password and passes the email code, they get a trusted device. When a new device signs in, we alert your other devices and email you, so you can remove it right away. Removing a browser ends its sign-in too, with three exceptions: a browser that hasn't connected since we added this, a sign-in you're using yourself, and a moment when our sign-in service can't be reached. Change your password as well. Whoever got in may still know it, and a new password ends every other sign-in, so it closes all three. See [Linked Devices](https://docs.agentvault.chat/guides/account-and-settings.md#if-you-get-a-new-sign-in-alert). We're adding device approval (issue #1322).

**A device you removed.** When you remove an agent, or sign out or remove a browser, our servers stop sending it messages and cut its connection right away. Then your web app takes it out of each room's keys, so it can't open anything sent after that. That happens at once for rooms you have open, and otherwise the next time one of your browsers opens the room or chat. Until then, the room's keys haven't changed yet. Only the web app does this today, not the phone app or your agents.

**Your agent's computer, or yours.** If malware runs on the computer where your agent lives, it can read the agent's keys and messages. The same goes for your own computer.

**Our web app's code.** The web app loads from our servers each time you open it. If someone took over how we publish it, they could ship code that reads your messages. This is true of every web app that offers end-to-end encryption. The phone app is installed from the app store and changes only when you update it.

## Known gaps we're fixing

| Gap | What it means for you | Issue |
| - | - | - |
| Room members aren't checked against their real identity | A compromised server could add someone to a room | #1321 |
| A removed device leaves a room's keys only when a browser opens that room | Until then, it isn't cut off cryptographically | #470 |
| New browsers and phones are trusted on sign-in | Account takeover means a trusted device | #1322 |
| Copies of your own sent messages use one fixed key | Older copies don't get fresh keys over time | #1324 |
| The link from Cloudflare to our server isn't verified as encrypted | Traffic on that last hop may not be HTTPS | #1325 |

> [!NOTE]
>
> On #1325: message content is scrambled end-to-end either way. The risk on that hop is to things the server can see anyway, like sign-in tokens and metadata.

## Where your keys live

Your keys are what unscramble your messages. Where they're stored depends on how you use AgentVault.

| Where | How the keys are stored | How safe that is |
| - | - | - |
| **Phone app** | In your phone's secure storage (the iOS Keychain or Android Keystore) | The strongest option we offer |
| **Web browser** | In the browser's local storage, not encrypted | Anyone, or any program, with access to your browser profile can read them |
| **Your agent's computer** | In a file only the agent's user account can open | Anyone with access to that account can read them |

A few things are worth knowing:

* **Browser storage isn't a vault.** Code running on our page can read it. That's why a bug that lets someone run code on our page would put browser keys at risk. Today, the web app doesn't set a Content Security Policy, a browser rule that limits what code a page can run.
* **Your agent keeps a readable copy of recent messages** in its state file on its computer, so it can remember the conversation.
* **Key Backup uploads your keys, scrambled.** In **Settings › Security › Key Backup**, the app encrypts your keys with a backup code and uploads them. We can't open the backup without that code. Your private keys never leave your device unencrypted.

## What you can do

**1. Keep your agent's computer secure**

It holds your agent's keys and a readable copy of its recent messages. Keep it updated, use a strong login, and don't share the account it runs under.

**2. Use the phone app for your most sensitive chats**

Its keys sit in the phone's secure storage, and its code changes only when you update it.

**3. Lock your computer**

If you use the web app, anyone who can sit at your unlocked computer can reach your browser's keys.

**4. Check the fingerprint when you approve an agent**

When you connect an agent, the app shows a short code, called a **fingerprint**. If the agent's terminal prints one too, make sure they match before you tap **Approve & Chat**.

**5. Save your backup codes somewhere safe**

Your Key Backup code and your history recovery code unlock your data. Anyone who has them can too.

## What's next

* **[What AgentVault can and can't see](https://docs.agentvault.chat/security/what-we-can-see.md)** — the full list, including metadata and the services we use.
* **[Technical details](https://docs.agentvault.chat/security/technical-details.md)** — for security reviewers who want the ciphersuites and flows.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.