Skip to main content
OrcaRouter does not log, store, or retain the content of your prompts or model outputs.

​
What this means, concretely

  • Requests (prompts, messages, tool call payloads, uploaded audio and images) are forwarded to the destination provider and discarded in memory after we receive the response.
  • Responses (generated text, tool results, generated images, TTS audio) pass through our servers back to you and are not written to any persistent store.
  • Error logs capture a truncated error message from the upstream (e.g., “rate limit exceeded”, “context length exceeded”) for debugging — but never the prompt or response content that triggered the error.

​
What we do keep

See Data Handling for the full list. In summary: timestamps, token counts, latency, and HTTP status codes — the metadata necessary to bill correctly and detect abuse. Never content.

​
Why this is the default (not a per-request opt-in)

Some API platforms let you toggle retention per-request. We made non-retention the default because:
  1. The overwhelming majority of commercial and personal use cases don’t benefit from having prompt content stored.
  2. A default-on flag is an attack surface — misconfiguration leaks prompts.
  3. Zero retention is a differentiator from direct-provider use: OpenAI retains 30 days of abuse logs; Anthropic retains similarly. OrcaRouter does not add a second retention layer on top.
If you need content retention for your own observability or evaluation, capture prompts and responses in your own application before sending them. OrcaRouter will never hold a copy.

​
Caveat: upstream providers still receive your data

OrcaRouter is a pass-through. Your prompts and responses are seen by the upstream provider (OpenAI, Anthropic, Google, etc.) under their own terms and retention policies. If you need end-to-end retention guarantees that span the upstream, check that provider’s policy — or choose a provider that offers explicit ZDR itself.