Trust centre
What we hold, where it goes, and what we are not
A gateway sits between your users and seven providers, which makes it a good place to ask hard questions. This page collects the answers, the documents behind them, and the things we cannot claim.
The short answers
Do you store our prompts?
Not unless you switch it on. By default we keep how a request behaved - model, tokens, latency, cost - and never the text. With storage on, text is kept for a window you choose, up to 30 days, readable only by your account, and turning it off deletes what was stored immediately.
How long is anything kept?
Prompt text: only if you opt in, up to 30 days and shortenable. Request metadata: 24 months. Account data: until you delete the account, which deletes it.
Where does processing happen?
Our platform runs in the EU (Frankfurt). The models do not necessarily: each provider's processing region is recorded per model with its source and the date we checked, and an EU-only guardrail restricts a key to models whose recorded policy processes in the EU.
Do the providers train on our prompts?
Most do not; one does, by its own published policy, and it is off by default and never chosen automatically because of it. Every provider's position is on its page with a link and a date rather than summarised into a badge.
How are credentials stored?
Provider keys you add are AES-256-GCM encrypted under a key held in the runtime, never in the database, and are never returned to a browser. Gateway keys exist in plaintext once, in the dialog that created them; we store a hash.
Who at your end can read our data?
Tenant isolation is enforced by Postgres row-level policies rather than by a WHERE clause somebody might forget. Prompt text we do not store cannot be read by anyone, including us.
Certifications, stated plainly
We are not SOC 2 or ISO 27001 certified. No audit is in progress. Saying so here is cheaper for both of us than saying it in week three of a procurement.
What exists instead is checkable. Every practice on the security page describes behaviour you can verify against the product - that a provider key is never returned to a browser, that a response states its own cost, that turning prompt storage off deletes what was stored. A claim you can test is worth more than one you must take on faith, though it is not a substitute for an audit and we are not pretending it is.
Subprocessors
Requests reach whichever model provider you route them to - today that is OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek, Moonshot AI, xAI - and each one processes the prompt for that request under its own published terms, linked from its page. Beyond the providers, the platform runs on managed hosting and a managed Postgres database in the EU, with payments handled by Stripe.
Every provider, with its data policy and the date we checked it
The documents
Security practices
Key custody, encryption, row-level isolation and the disclosure policy.
Privacy policy
What we collect, what we never collect, and how long each is kept.
Data processing agreement
The processor terms, for customers who need one on file.
Terms of service
The commercial agreement.
Content policy
What the platform screen refuses, and why it is a short list.
Live status
Provider health from real probes, on the same data the router uses.
Found something we have got wrong?
Security reports go to security@finalrouter.com. We will not threaten anyone who reports a vulnerability in good faith. Anything else - a claim on this site you think does not hold - is worth an email too, and will be fixed or explained.