Skip to main content

​
Get access

Explore the core concepts to understand how Silmaril approaches protection. Silmaril provisions the API key and endpoints for your deployment. Book a call to arrange access. SDKs use the classify endpoint. LiteLLM uses a separate guardrail endpoint.

​
Prerequisites

Choose the service that owns inference, tool execution, or orchestration for the system you want to protect. That service creates one firewall client and reuses it.

​
Make your first classification

1

Install the SDK

Choose your runtime. Each guide starts with the install command.
2

Configure the environment

Set SILMARIL_API_KEY and SILMARIL_API_URL in the service environment to the values from Get access.
3

Classify a request

Run the first example in your SDK guide. It creates the client and classifies one user input with the USER_INPUT hook. Label tool output with the tool response hook and the tool name.
4

Handle the result

Continue only when the verdict and any applicable governance policy allow the request. Treat a blocked request and a failed call differently. A block is a verdict, while an API or network error means no verdict was returned.

​
Handle blocks and errors

Shadow mode returns results instead of block errors so you can observe would-block decisions before a surface enforces them. Route malicious results with firewall outcomes.

​
Event contract

Each call sends one complete event. The SDK keeps your metadata.conversationId unchanged and adds a metadata.silmaril.request_id per event. The prediction field is the authoritative verdict. Outcome fields add detail for malicious results.

​
Next steps

Use framework integrations when a framework owns model execution, or AI gateway integrations when LiteLLM or Vercel AI Gateway routes your model traffic. For a deployment in your cloud account, see self-hosting.