AI governance on the request path.

Applications use the branch gateway.

Configure an application or agent to send supported model API requests to its provisioned Caged gateway URL using a virtual API key. Manage provider connections and policies from the portal; the gateway applies those controls to the requests routed through it.

1. 01 / APPLICATIONApp or agentSends a supported request with a virtual key.
2. 02 / CAGED GATEWAYApply policyCheck access, usage limits, and configured privacy rules.
3. 03 / MODEL ENDPOINTCloud or local modelReceives the approved request; results return to the caller.

Cloud and local models share a policy plane.

Use supported cloud providers or configured local model endpoints. Select the connection and model for each application, then apply its access, budget, and privacy policy. For local deployments, include the model's hosting location, API format, and hardware requirements in your plan.

Caged supports OpenAI-compatible Chat Completions and Responses when enabled, Anthropic Messages, and model discovery. Streaming and tool calling depend on the selected model and request format. The application remains responsible for executing its tools.

Use the branch gateway endpoint supplied for your deployment. The product compatibility summary lists provider and format coverage.

Choose how detected sensitive data is handled.

Configure the data types, detection rules, and policy for supported request content. Detection coverage depends on the payload, selected detector, and configured entities. Choose the mode that matches the application's requirements:

Monitor

Detect and report configured sensitive-data matches. Monitoring does not redact the request.

Mask

Replace detected sensitive values in supported request content before forwarding to the model.

Block

Reject a request when it matches the configured blocking policy.

Validate the policy with representative data, including fields that must be detected and cases that should remain unchanged. Include detection failures and unavailable services in the deployment's failure-handling tests.

Give applications controlled model access.

Use virtual API keys for application access, with model allow-lists, target branches, privacy policy, and configured usage limits. Each provider-owned key belongs to its provider connection; manage multiple connections and keys from the same portal.

Set spend or request budgets and request/token rate limits for the intended workload. Exercise the limits in the pilot and review the resulting decisions and usage records. Keep provider credentials under administrative control rather than distributing them to every application.

Review decisions through metadata audit.

Metadata audit and usage records help operators review which keys and models were used and how policy handled the requests. Available usage, cost, timing, and verdict information depends on the selected integration and request.

Prompt and response bodies are excluded from these metadata audit records. Configure retention and access for your deployment, and review the model provider's own logging and retention settings as part of the data-handling plan.

Start with a defined application and site.

1. 01Identify the applications and agents that will use the gateway, their providers, models, and API formats.
2. 02Choose the appliance and gateway deployment for the site, and enable the required software and license features.
3. 03Configure provider connections, model access, virtual keys, privacy rules, budgets, and the branch gateway endpoint in the portal.
4. 04Point the applications at that endpoint. Validate allowed, masked, blocked, and budget-limited requests with representative test data.
5. 05Review audit records and any paths that bypass the gateway, then expand the deployment after the pilot meets your requirements.

Compare the Small / Portable office, Medium office, and Large office hardware against the site's interfaces and workload. Record the software release, supported formats, policy settings, and test results used to approve the deployment.

Review the privacy inspection lab results for measured component timings, test conditions, and raw samples.

Use the branch AI gateway pilot guide to plan an office team using Claude Cowork and Claude Code with virtual keys, privacy checks, and usage controls.

Put every intended request on the governed path.

The API controls apply to supported requests routed through the gateway. Consumer AI websites, off-network devices, and applications that bypass that path need separate controls. Plan the network and application configuration accordingly.

Provider compatibility does not establish support for every model, endpoint, file type, or payload. PII detection is bounded by its rules and supported content; an allowed request is not proof that every sensitive field was found. Validate the request formats and detection coverage needed for your applications during the pilot.