The Privacy Tradeoff Behind Free Stealth Models
Zero-dollar inference is useful for testing, but it is not a privacy guarantee. Ox Alpha is an anonymous preview, so the correct question is not “is it free?” but “which provider receives this request, and what does that provider say it does with it?”
Default rule for an anonymous preview
Assume prompts may be visible to the active provider until the exact route's terms say otherwise. Use synthetic repositories, public documents, and redacted examples for first-pass testing.
What each access surface currently says
OpenRouter
OpenRouter's model catalog identifies the provider namespace as stealth, but the API record does not publish an underlying lab identity. Read the model-specific terms and do not infer that OpenRouter's general platform policy overrides the upstream provider's policy.
OpenCode
OpenCode's public site says the OpenCode product does not store code or context, while Zen documentation describes a curated gateway with provider-specific exceptions. The live alias is not publicly named Ox Alpha, so check the current provider and terms in the product before sending proprietary context.
Read OpenCode provider docs →Command Code
Command Code's official documentation says “your code stays yours” and that Command Code does not train on or store it. That is a product-level claim; still confirm the current upstream route and use the available zero-data-retention controls when your workflow requires them.
Read Command Code security docs →Five-minute preflight checklist
- Remove API keys, passwords, customer identifiers, private URLs, and credentials from the prompt.
- Replace proprietary code with a minimal reproducible example.
- Confirm the exact model ID and upstream provider shown by the client.
- Check whether a zero-data-retention mode is available and whether it changes routing.
- Set a fallback model and a spending/rate limit before letting an agent run unattended.