Gateway · request-time control
Request-time control in your network: each AI request decided under the signed policy and recorded.
The gateway runs in your network, between your applications and the model providers. It decides each request under the signed policy revision from the Control Plane and records the decision. A request served without a record is counted in a gap record.
Request path
Capabilities
- A decision for each requestEach request is allowed, redacted, routed, redacted and routed, or refused under one signed policy revision.Policy
- Record-only mode firstThe modes are Record only, Advisory and Enforce. The mode is part of the signed policy revision, so a change needs approval by a second person.Policy
- Redaction of personal dataIn Advisory and Enforce modes, the personal data the gateway detects is redacted before the request is forwarded. The record keeps its position and type, never the text.Ledger
- A record of every decisionEach record is chained to the one before it. If a record is changed, verification names it.Ledger
- Emergency stopsStop a run, an agent your application labels (X-Saible-Unit), a workload or an application at every gateway. Stops apply in every mode, including record-only mode.Emergency stops
- An application key for each applicationEach request is attributed to the application that sent it. A request with an unknown key is not forwarded.Registry
- Pulled, never pushedEach gateway pulls the signed policy revision and the emergency stops. Nothing is pushed into your network.Gateways
- Replay of recorded requestsReplay a window of records step by step, from records that hold no request text. The replay opens on the most recent emergency stop.Gateway → Overview
- Routing and cache estimatesRouting compared with the usual model is an estimate at list price, over requests to auto or an alias only. A named model is never swapped. The cache runs in shadow mode: it counts the requests it would have answered, prices them at list price, and serves nothing.Gateway → Overview
- Installation by your teamSaible provides a Terraform module for ECS Fargate in your AWS account, and Kubernetes manifests that have not yet been run on a cluster.Gateways
Three modes
Start in record-only mode. Nothing changes for your applications, and you see what the policy would have done.
Record onlyrecord_only
Every request is forwarded unchanged to its model's default endpoint, and everything the policy would have done is recorded. Requests under an emergency stop, and requests with an unknown model or key, are refused in every mode.
- Forwarded unchanged, except for stops
- Personal data redaction applies
- Rules route or refuse
- Emergency stops refuse
- Each decision recorded
Advisoryadvisory
The rules only record what they would have done. Redaction of personal data still applies before forwarding.
- Forwarded unchanged, except for stops
- Personal data redaction applies
- Rules route or refuse
- Emergency stops refuse
- Each decision recorded
Enforceblocking
The rules are applied. A request is redacted, routed or refused as the policy states, and recorded.
- Forwarded unchanged, except for stops
- Personal data redaction applies
- Rules route or refuse
- Emergency stops refuse
- Each decision recorded
Steps
- 1On Gateways, click Add a gateway and issue its credential.
- 2Install the gateway in your network, in record-only.
- 3Point one application's SDK at the gateway, with that application's application key.
- 4Send one request and open its record in the Ledger.
Gateway topics
- Gateway01A request, through the gatewayEvery request takes the same path. The policy decides the outcome. A record is written on every path, or a gap record states that it was not.
- Applications02Claude Code, through a gatewayOne session at a time, with the application's key. Each turn has a decision record, or a gap record if one could not be written.
- Ledger03Contents of a decision recordWhere personal data was found and what was decided, never the text itself.
- Control04Emergency stopsStop a run, a workload or an application in every mode, including record-only mode.
- Access05Keys and their holdersFive keys, each with one purpose.