Vercel AI SDK

Wrap any AI SDK agent so every tool call is verified.

npm install @detecte/sdk ai

import { Detecte } from "@detecte/sdk";
import { detecteWrap } from "@detecte/sdk/integrations/vercel-ai";
import { generateText } from "ai";
import { openai } from "@ai-sdk/openai";
 
const detecte = new Detecte({ apiKey: process.env.DETECTE_API_KEY! });
 
const guarded = detecteWrap(
  {
    model: openai("gpt-4o"),
    tools: {
      refund_order: {
        description: "Refund an order",
        parameters: z.object({ orderId: z.string(), amount: z.number() }),
        execute: async ({ orderId, amount }) => doRefund(orderId, amount),
      },
    },
  },
  { detecte, agentId: "agent_support_bot" }
);
 
const { text } = await generateText({
  model: guarded.model,
  tools: guarded.tools,
  prompt: "Refund order ord_123 for $42.",
});

When the model calls refund_order, detecteWrap calls verify() first. If blocked, the tool returns an error result back to the model loop with the policy reason — the model then continues, often gracefully informing the user it can't perform the action.

What you get

Streaming-aware — works with both generateText and streamText.
Trace IDs — every Detecte decision carries the AI SDK toolCallId in context, so you can correlate from either side.
OpenTelemetry — spans are tagged with ai.tool.name and detecte.decision.id when OTel is configured.