Data policy
The gateway logs metadata, never content. This page states exactly what is recorded, for how long, and what happens to your prompts — the developer-facing summary of our Privacy Policy.
Metadata only
Prompts and responses are not stored. They are processed in memory to serve the request and then discarded. What we keep per request is metadata: model, token counts (input / cached / output / reasoning), cost, latency, status and error code.
Not even on errors
A failed request logs an error code and counts — never the content that triggered it. There is no debug mode on our side that quietly captures your prompt.
What is logged, and for how long
| Data | What exactly | Retention |
|---|---|---|
| Usage metadata | Model, token counts, cost, latency, status, error code — no content. | Detailed: 6 months · aggregates: long-term |
| Technical logs | IP address and user-agent at the edge/server layer, for abuse prevention. | 7 days |
| Cache entries | Responses cached for the caching feature, keyed within your organization by default. | Up to 24 hours |
| Prompt / response content | Not stored. | — |
Account and billing data (email, top-up history) follow the retention rules in the Privacy Policy — the authoritative document this page summarizes. We do not sell data, do not train models on your content, and do not run ads on your data.
About the cache
The exact-match cache stores a response only so an identical request can be served again — scoped to your organization by default, expiring automatically within 24 hours. Caching is controllable per request via the cache extension field; a cross-organization scope exists only as an explicit opt-in.
Upstream transparency
Inference runs on Cloudflare Workers AI. We publish this deliberately rather than hide it: every model lists its upstream id (@cf/...) on the models page, and every API response names the provider:
x-gw-provider: cloudflare-workers-ai
Your prompts transit Cloudflare's infrastructure for processing under their data usage terms — not used to train models. Performance numbers we publish always carry their measurement conditions, and live service state is on the status page.
Optional content logging: not available yet
There is currently no way to make us store your content, even if you want it for debugging. A per-organization, explicitly opt-in debug log with a 7-day retention is planned as a future feature; until it ships, nothing you send is retained anywhere.
Questions about data handling: [email protected].