Skip to content
Build Perch
Browse documentation

Privacy Mode (ZDR)

Short answer: On by default, Privacy Mode tells OpenRouter to route every request only to providers that attest they don’t store or train on your prompts and files. That’s the providers’ own attested policy enforced by OpenRouter’s routing. Build Perch can’t independently verify it.

What does Privacy Mode do?

Privacy Mode is a global toggle (Integrations → OpenRouter) that’s on by default. When it’s on, every request the app makes to OpenRouter carries a zero-data-retention (ZDR) flag; that includes ordinary chat, image/speech/music/video generation, and behind-the-scenes work like conversation summarization. OpenRouter reads that flag and restricts routing to endpoints that have attested they retain nothing and train on nothing you send.

The word “attested” is doing real work here. Build Perch can ask OpenRouter to route this way, and OpenRouter enforces it on its side, but neither Build Perch nor OpenRouter can independently verify what happens inside a given provider’s infrastructure. It’s the provider’s stated policy, enforced by OpenRouter’s routing layer, not a cryptographic or auditable guarantee. It restricts where requests can go; it doesn’t prove a provider retained nothing.

One practical consequence: many free-tier models don’t offer zero-data-retention, so turning Privacy Mode on removes them from what you can pick.

How do I turn it on or off?

  1. Open Integrations → OpenRouter from the app menu.
  2. Under Privacy & Security, check the “Zero Data Retention (recommended)” toggle; it’s on by default, so most people never need to touch this.
  3. Turning it off trades privacy for model availability: it makes free-tier and other non-ZDR models selectable again.

Good to know

  • Privacy Mode only affects OpenRouter requests. Local Ollama models never send anything anywhere, so the toggle doesn’t apply to them; Ollama Cloud requests go to ollama.com under whatever privacy terms that service publishes.
  • The toggle is not reachable by the AI’s own tools, so nothing the model does mid-conversation (including a prompt-injection attempt) can switch it off.
  • It applies to requests made while it’s on. It doesn’t retroactively affect anything sent before you enabled it.