Cloudflare Agents

Cloudflare's Agents SDK runs its LLM calls through the Vercel AI SDK (streamText / generateText), and its tools are ordinary AI SDK tools (tool({ description, inputSchema, execute })). So the AxioRank Vercel AI SDK adapter guards them directly: there is no Cloudflare-specific package to install.

Install

npm install @axiorank/sdk

Guard your agent's tools

Wrap the tools before you hand them to streamText, inside your AIChatAgent's onChatMessage. An Agent exposes this.env, so build the client there:

import { AIChatAgent } from "agents/ai-chat-agent";
import { streamText } from "ai";
import { AxioRank } from "@axiorank/sdk";
import { guardTools } from "@axiorank/sdk/vercel";

export class ChatAgent extends AIChatAgent<Env> {
  async onChatMessage(onFinish) {
    const axio = new AxioRank({ apiKey: this.env.AXIORANK_KEY }); // axr_live_...
    return streamText({
      model,
      messages: this.messages,
      tools: guardTools(this.tools, axio, { onDeny: "return" }),
      onFinish,
    });
  }
}

Every tool call the agent proposes is now scored for leaked secrets, PII, destructive operations, prompt injection, and egress, checked against your policy, and written to the audit log before it runs. On a deny, onDeny: "return" hands the model a short, readable refusal it can recover from; onDeny: "throw" fails the tool step instead.

Next steps

• Vercel AI SDK adapter: the adapter this reuses.
• Content-inspection engine: what each call is scored against.
• Gateway API: the HTTP contract behind every adapter.