Vercel AI SDK

The @axiorank/sdk/vercel adapter wraps the tools object you already hand to generateText / streamText, so every tool call is scored by AxioRank before it runs.

Install

npm install @axiorank/sdk

Guard your tools

guardTools wraps each tool in a tools record; the record key is sent to AxioRank as the tool name, so guarded calls line up with your audit log.

import { generateText, tool } from "ai";
import { openai } from "@ai-sdk/openai";
import { AxioRank } from "@axiorank/sdk";
import { guardTools } from "@axiorank/sdk/vercel";
import { z } from "zod";

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

const tools = guardTools(
  {
    deployToProd: tool({
      description: "Deploy the current build to production",
      parameters: z.object({ service: z.string() }),
      execute: async ({ service }) => deploy(service),
    }),
  },
  axio,
  { onDeny: "return" }, // let the model read the refusal and adapt
);

await generateText({ model: openai("gpt-4.1"), tools, prompt });

guardTool(name, tool, axio, opts) wraps a single tool when you'd rather not wrap the whole record. A tool with no execute (a client-side or provider-executed tool) is returned unchanged.

Govern the whole model call

guardTools guards the tools an agent calls. To also inspect the prompt and completion of every model call, wrap the model itself with the @axiorank/ai-sdk middleware. One wrapLanguageModel call governs generateText, streamText, and generateObject: the prompt is inspected before it is sent, the completion is inspected and optionally redacted before your app sees it, and proposed tool calls are inspected before the SDK runs them.

npm install @axiorank/ai-sdk ai

import { wrapLanguageModel } from "ai";
import { openai } from "@ai-sdk/openai";
import { createAxioRankMiddleware } from "@axiorank/ai-sdk";

const model = wrapLanguageModel({
  model: openai("gpt-4o"),
  middleware: createAxioRankMiddleware({ apiKey: process.env.AXIORANK_API_KEY! }),
});

Leave the API key off to run the detectors in-process as an advisory guard, with no signup and no network call.

Next steps

• Framework integrations: the wrap-vs-callback model and onDeny.
• Content-inspection engine: what gets flagged.