TRUST CENTER
Most AI APIs ask you to take their word that your prompts aren't read, logged, or fed into someone's next model. We don't. The links on this page are the artifacts you can use to check, line by line.
Set privacy.mode = "mesh" and these four things are true. The section after this is what makes each one checkable.
Set privacy.mode = "mesh" and your prompt only decrypts inside an enclave whose code hash you can fetch and compare. The network path between you and the model — load balancers, sidecars, log shippers — sees ciphertext.
Cost, model, route, latency, attestation hash at the time of the call — signed by the platform key, retrievable by id. Your prompt body and completion body are never in there.
Prompts and completions hit memory, get used, get freed. No log, no queue, no metric store, no training set. Operators can't pull them back even if they wanted to.
Pull the current measurement registers (PCR0/1/2), check them against the reference values we publish with each release. If they don't match, don't send the prompt.
What turns the claims above from things you read into things you check.
The enclave measures itself at boot — bootloader, kernel, app image — and signs the result. You can fetch that signature, check it against what we say we shipped, and refuse to talk to anything that doesn't match.
A receipt is the proof a call happened. It's what you'd hand to your finance team for reconciliation, your auditor for evidence, or your support team for a dispute — and nothing of your prompt is in it.
Threat models lie when they pretend everything is covered. Here is who you are protected from on this platform, who you are not, and the assumptions that have to hold for the rest of this page to be true.
If the verifier doesn't agree with what we wrote on this page, we're wrong. That's the point. It's open-source, reproducible, and we ship the SHA of every release alongside it.
The usual list. What we have, what is in flight, what we can do for you on request.
| Framework | Status | Notes |
|---|---|---|
| SOC 2 Type II | In progress | Audit window opens once the first quarter of production traffic is in. We'll publish the letter when it's signed. |
| ISO/IEC 27001 | Planned | On the roadmap after SOC 2 ships. |
| GDPR / UK GDPR | Live | DPA and SCCs are ready — email the disclosure address. |
| HIPAA | Available on request | BAA available for in-scope accounts. |
Third parties that handle metadata to keep the service running. None of them see prompt or completion content.
| Provider | Purpose | Region |
|---|---|---|
| Stripe | Fiat payment processing | United States, European Union |
| Cloudflare | Edge network and DDoS protection | Global |
| Sentry | Error tracking. No payload content reaches them. | United States, European Union |
| Vercel | Marketing site hosting and SSR for non-privacy routes | Global |
We give 30 days notice on any material change to this list, sent to anyone on the security feed.
Anything that touches confidentiality, integrity, or availability of customer data lands here within ten business days of being closed.
CLEAN
Nothing reportable yet.
If any of this is wrong, we want to know before our customers do. Send what you have to the address below.
security@meshrouter.app