Security and privacy

How Twarita keeps you in control

An agent that can send, submit and pay for you needs firm limits. Here is what Twarita does today, and where its limits are.

The safeguards

  • A separate Guardian checks every outside action

    Web requests, file writes, messages, form submissions and payments all pass through the Guardian, a separate program that the agents can't change. It applies your rules, which the AI can't rewrite, and anything it doesn't recognise stays blocked.

  • Risky steps wait for you in Inbox

    By default Twarita asks before it pays, sends a message, submits a form, deletes something, installs software or shares your personal data. Payments need your approval at every safety level. Each card shows exactly what will be sent, and where.

  • Agents reach only allowed websites

    Each task can reach the sites it was started for and the ones you've allowed in Trust Centre. Everything else is blocked and shown to you. Allowing a new site is itself an approval in your Inbox.

  • Passwords and keys stay in a vault

    They go straight into a vault on your device (Windows Credential Manager on Windows, your system keyring on a Linux desktop, an encrypted vault on your server) and are never shown again. Website passwords are typed in by the Guardian, only on their own site and after you approve, so the agents never see them.

  • Outside text is screened

    Web pages, emails and documents are screened for hidden instructions aimed at an agent, and suspicious text is removed before an agent reads it.

  • Pause all, on every screen

    One click stops every agent and every task. Nothing runs again until you choose Resume.

  • A signed receipt for every action

    Activity keeps a receipt, signed by the Guardian, for each action Twarita takes, so you can see exactly what it did and when.

  • Your data stays on your device

    Your profile, history and files live on your computer or your own server. Twarita sends no telemetry. The AI provider you choose receives the prompts each step needs; with a model on your own computer, nothing leaves it.

  • No open ports on your server

    On your own server, Twarita is reachable only from your devices over Tailscale, and each device needs a pairing code or a passkey.

  • Downloads you can check

    Every installer is listed with its SHA-256 checksum, so you can confirm the file is the one we published.

What to keep in mind

  • The installers aren't code-signed yet. Windows shows a SmartScreen warning the first time. Check the SHA-256 before you run the installer.
  • No safety check is perfect. Screening can miss a cleverly hidden instruction, and an AI can misread a page. That's why the risky steps come to you, and why it's worth reading each approval before you say yes.
  • Your AI provider sees your prompts. What you send to a cloud AI is handled under that provider's own policy. Use a model on your own computer if you'd rather nothing left it.
  • You're the last check. Twarita acts on what you approve. If you're unsure about an approval, deny it; the agent skips that step.

Found a security problem? Please tell us privately: see reporting a security issue.