Getting started

This guide takes an agent action from "calls the tool directly" to "passes through AxioRank first." The shape is the same whether the action is a tool call, an outbound card, or a protocol message.

1. Point the agent at the gateway

Instead of calling a tool's API directly, the agent calls the AxioRank gateway endpoint. The gateway authenticates the agent with its API key and forwards the request only after it clears inspection.

import { AxioRank } from "@axiorank/sdk";

const axio = new AxioRank({ apiKey: process.env.AXIORANK_API_KEY });

// The agent's intended action, described to the gateway.
const verdict = await axio.inspect({
  agent: "support-bot",
  tool: "billing.refund",
  payload: { amount: 4200, account: "acct_19f..." },
});

if (verdict.decision === "block") {
  throw new Error(`Blocked by AxioRank: ${verdict.reason}`);
}

2. Read the verdict

Every inspection returns a decision and the signals behind it:

• allow — no material risk; proceed.
• redact — the payload was rewritten to strip sensitive content; use the returned payload.
• block — the action is denied; surface reason to the operator.

3. Set a policy

Defaults are conservative. Tighten or relax enforcement per agent and tool from the dashboard — for example, always block destructive database operations, or redact-and-allow when PII is detected in an outbound message.

Next steps

• Content-inspection engine — what the scanner looks for and how scores are produced.
• Protocol adapters — apply the same gate to MCP/A2A traffic and outbound cards.