- "Your screening will miss something." For checksum-backed identifiers, card numbers, bank routing numbers, IBANs, a structurally valid value in readable text cannot pass unmatched: that is how the detector is built. For everything else, detection has measured limits, and we publish them instead of promising zero. The evidence pack shows exactly what was checked and what passed. - "I cannot trust a replay months later." Each decision is recorded with the policy version that produced it, so a replay runs that exact version, not whatever is live today. - "The record only covers traffic you can see." True. The record covers traffic through the Gateway. Anything called directly against a provider, bypassing it, falls outside what it can show an auditor. - "Agents will route around your checks." Point an agent's tool calls at the Gateway and each one is checked before the agent runs it, the same as a chat request. People work as before. - "Now your gateway sees all our data." It sees what your providers already see, one hop earlier, where rules can still act. Screening runs before anything leaves, request content capture is yours to switch off, and evidence exports work without it.

Pass the audit with proof.

Today it is hard to show which data reached which model, and why. The Gateway screens each request for regulated data before it leaves. It sends the request only to providers approved for that data and records the rule that decided it.

Join the waitlist See pricing

A screened request, as the Gateway records it. Illustrative values.

request: req_9b04e6a7
found: 2 identifiers, 1 account number
replaced: with stand-ins before sending
provider: approved for this data class
refused: a provider not approved for it
policy: v42, replayable

Answers your auditor accepts

Nothing raw leaves. The Gateway replaces regulated data with stand-ins. It reads every channel of the request: messages, system instructions, tool calls, query parameters, and documents with extractable text. A provider not approved for that data never receives it.
Every tier gets the full shield. Personal, payment card, and health identifiers are screened on every plan, identically. The top tier adds custom detection classes for your own identifier formats.
An unapproved destination is refused, not warned about. Health data, for example, routes only to a provider with a signed BAA on record. A request that has nowhere approved to go is refused, and the refusal is recorded with its reason.
Replay any decision. Pick a past request and replay it against the policy recorded with it. Detection is deterministic, so the replay finds exactly what the original run found.
Evidence in minutes. Scoped, tamper-evident exports answer an audit question, even with request content switched off. Screening counts report by team and person, with no content behind them.
Agents checked too. Point an agent tool at the Gateway. It checks each tool call before the agent runs it, and people work as before.

Objections

"Your screening will miss something." For checksum-backed identifiers, card numbers, bank routing numbers, IBANs, a structurally valid value in readable text cannot pass unmatched: that is how the detector is built. For everything else, detection has measured limits, and we publish them instead of promising zero. The evidence pack shows exactly what was checked and what passed.
"I cannot trust a replay months later." Each decision is recorded with the policy version that produced it, so a replay runs that exact version, not whatever is live today.
"The record only covers traffic you can see." True. The record covers traffic through the Gateway. Anything called directly against a provider, bypassing it, falls outside what it can show an auditor.
"Agents will route around your checks." Point an agent's tool calls at the Gateway and each one is checked before the agent runs it, the same as a chat request. People work as before.
"Now your gateway sees all our data." It sees what your providers already see, one hop earlier, where rules can still act. Screening runs before anything leaves, request content capture is yours to switch off, and evidence exports work without it.

Try to leak a secret

Send a planted set of identifiers through the Gateway.

Inspect what went out. Every identifier is replaced. None is sent raw.
Read the refusal record for each provider that was not approved.
Run the same set twice. The findings match, byte for byte, because detection never guesses.

Bring your hardest audit question.

We admit customers in cohorts, in waitlist order. Don't miss the next one. When your slot comes up, you start on one real workload, on your plan from day one. You keep an evidence report to hand your auditor. If we are not earning our fee, leaving is the same URL change, reversed.

Join the waitlist