Security
What we store about your requests, and what we don't
A plain account of the data Usefully keeps when you call the API: request metadata yes, prompt and response bodies no, and API keys only as salted hashes.
· 3 min read · Usefully team
When you send prompts through any API gateway, it's reasonable to ask what that gateway keeps. Here's the answer for Usefully, in plain terms, so you can make your own decision about what to send through it.
What we keep for each request
Each request produces a log entry with metadata only. That's what powers the Logs page, usage charts and billing:
- Time, API key (by name, never the key itself), and organization.
- Model requested and the provider that served it.
- Input and output token counts, and the cost charged.
- Latency and status (success, failed, partial), plus the error type if it failed.
What we don't keep
Prompt and response bodies are not stored. They pass through Usefully on the way to the provider and back to you, and are not written to our logs or database. That has a practical consequence worth knowing: we can't show you the text of an old request, and support can't look it up either. If you need a record of conversations, keep it in your own system, where you control retention.
API keys
- A key is shown once, when you create it. After that we store only a salted hash, so we can check a key but can't read it back.
- If you lose a key, create a new one; we can't recover it.
- Revoking a key takes effect almost immediately. See rotating API keys without downtime.
Providers
Your request is forwarded to a model provider (Anthropic, OpenAI or Google, depending on the model), and that provider's own data terms apply to what it receives. If a request contains data that must not be retained or used for training, set data_policy so it's only routed to providers that qualify. Details in zero-retention routing.
Web search
Search queries go to our search provider, Exa, to fetch results. As with chat, Logs show the request's metadata and cost.
In transit
Every connection uses HTTPS: from your app to api.usefully.info, and from Usefully to each provider. Keep your own side safe too:
- Call Usefully from your server. Never ship a key in browser or mobile code.
- Use a separate key per environment and service, so one leak means one revocation.
- Set per-key limits (requests and tokens per minute, and a daily spend limit) so a leaked key has a bounded cost.
Payments
Card and payment details are entered on Dodo Payments' checkout, our merchant of record, which also handles tax and invoices.
Why metadata is enough for us
Everything Usefully needs to run the service can be done from metadata. Billing needs token counts and prices. Failover needs to know which provider failed and how. Rate limits and spend limits need counts and costs per key. Debugging a routing problem needs status, latency and the provider involved. None of that requires the text of your prompts, so we don't keep it.
When you contact support about a specific request, include the request_id from the response or error. It lets us find the log entry quickly without you having to share the prompt itself.
The privacy policy is the formal version of this post. Questions about data handling: support@usefully.info.