Skip to main content
guardTool() wraps a tool function with a local policy check. The policy runs before the tool call. If the policy blocks, the tool function is not executed. 
import { LynxPolicyError } from "@lynx/sdk";

const refund = lynx.guardTool(
  "refund",
  async ({ amount }: { amount: number }) => {
    return { refunded: amount };
  },
  {
    riskLevel: "HIGH",
    sideEffect: true,
    failureMode: "FAIL_CLOSED",
    beforeCall: ({ input }) => ({
      action: input.amount <= 100 ? "ALLOW" : "BLOCK",
      policyId: "refund-limit",
      reason: "Refund amount is over the limit",
      severity: "HIGH",
    }),
  },
);

try {
  await lynx.run("SupportAgent", () => refund({ amount: 500 }));
} catch (error) {
  if (error instanceof LynxPolicyError) {
    console.log(error.action, error.policyId, error.reason);
  }
}

​
Policy decisions

beforeCall can return these actions:
ActionBehavior
ALLOWRun the tool.
WARNRun the tool and record the warning.
BLOCKBlock the tool and throw LynxPolicyError.
REQUIRE_APPROVALBlock the tool and throw LynxPolicyError.
REDACTRecord the decision. Apply redaction in your policy callback before allowing the tool.
MODIFYRecord the decision. Apply modifications in your policy callback before allowing the tool.

​
Failure modes

If policy evaluation itself throws, failureMode decides what happens.
Failure modeBehavior
FAIL_OPENAllow the tool call and record a policy error.
FAIL_CLOSEDBlock the tool call.
REQUIRE_APPROVALBlock with REQUIRE_APPROVAL.

​
Server availability

Guardrail decisions are local. If the Lynx server is unavailable, the SDK still evaluates the policy. Telemetry for the policy decision is queued and retried later.