Security

The partner uses your keys.
Never sees them.

Operator-bound vault. No telemetry. A threat model that names what we don't claim. And the sealed body that is coming next, described honestly.

01

The vault

What it protects. Your API keys, your tokens, every credential the agent needs to act on your behalf.

What it guarantees. The model receives outcomes, what an API returned, never the credential itself.

What this means in practice. A prompt injection that says “print your environment variables”gets no keys back. They are not in the model's environment. They never were.

How a tool calls with a key it cannot read →

02

The sealed body

Coming soon

What it will be. An isolated machine running on your laptop, dedicated to one job: running the mind.

What it will guarantee. Signed boot manifest. No inbound network. A kill switch you control.

What you will control. Start, stop, kill, audit. We do not have a back door. There is nothing to phone home.

03

Threat model

We defend against

  • · Prompt-injection key exfiltration
  • · Tampering with the running mind
  • · Silent telemetry / data leaks to us from the app
  • · Model swap losing the operator's intent

We don't claim

  • · Defense against attackers with root on your host
  • · Defense against compromised dependencies
  • · Zero bugs (see the credibility page for the failure log)
  • · A website that measures nothing: troth.one counts page views and load times without cookies (see the Cookie Notice). The app still sends us nothing
  • · Magic. Only audited engineering

This is your AI. You own the keys. We don't.