Skip to content
LUNAROUTEDocs

Zero data retention

LunaRoute is zero data retention for inference: request and response bodies are processed in memory and never stored. Your prompts, your code, your agent’s variable names: gone the moment the response ships.

The only thing recorded from an inference request is billing metadata:

  • timestamp
  • model
  • token counts
  • cost

That is exactly what your usage ledger shows — and it is all there is. No bodies, no headers, no secrets.

Enterprise deployments can add private, single-tenant inference — dedicated servers with no shared GPUs, no shared caches, and no co-mingled traffic.

Zero data retention covers inference request and response bodies. A few features store something by design, and only when you use them:

  • Generated images are retained for up to 7 days so you can fetch or edit them — see Images → Retention.
  • Document conversion stores the Markdown artifact for 7 days when you ask for a URL instead of inlined text; with inline (embed) output, nothing is stored.
  • Billing metadata (above) is kept to run the ledger and reconcile charges.
  • Third-party providers. When a request is routed to a model that is not hosted by LunaRoute — for example BYOK traffic sent with your own provider key — the upstream provider processes it under its own data retention policy. Zero data retention covers only traffic served on LunaRoute-hosted inference.

MCP tools otherwise process their inputs in memory.