Vercel AI SDK helper

guardAiSdkTools takes the tool map you pass to generateText or streamText and returns a copy where every execute calls authorize first. It has no runtime dependency on ai: the tool type is structural. A refused call does not throw. It resolves to a ToolDenial with the decision code, so the model reads the refusal as the tool result and stops. For the throwing variant built on the policy engine, see Vercel AI SDK.

npm install @glinr/theauth

Runnable snippet

import { generateText, tool } from "ai";
import { z } from "zod";
import { createTheAuth } from "@glinr/theauth";
import { guardAiSdkTools, isToolDenial } from "@glinr/theauth/integrations";

const auth = await createTheAuth({
  database: { provider: "sqlite", url: ":memory:" },
  agents: { enabled: true },
});

const agent = await auth.agent.create({
  ownerId: "user-alice",
  name: "support-agent",
  type: "autonomous",
  permissions: [{ resource: "tool:lookupInvoice", actions: ["execute"] }],
});

const tools = guardAiSdkTools(
  {
    lookupInvoice: tool({
      description: "Read an invoice",
      inputSchema: z.object({ id: z.string() }),
      execute: async ({ id }) => ({ id, status: "paid" }),
    }),
    refund: tool({
      description: "Refund an invoice",
      inputSchema: z.object({ id: z.string() }),
      execute: async ({ id }) => ({ id, refunded: true }),
    }),
  },
  { authorize: auth.authorize, agentId: agent.id },
);

// lookupInvoice runs. refund resolves to:
// { ok: false, denied: true, toolName: "refund", code: "no_permission", reason, approvalRequired: false, auditId }
const result = await tools.refund.execute({ id: "inv_1" }, {} as never);
console.log(isToolDenial(result) ? result.code : "ran");

Pass tools to generateText({ model, tools, prompt }) as usual.

Options

Option	Default	Meaning
`authorize`	required	`theauth.authorize`, or any function with the same shape
`agentId`	required	Agent id, or a function from the per-call context to an id
`action`	`execute`	Action checked
`resource`	`tool:<name>`	Resource checked, from the tool name and arguments
`requestApproval`	none	Called on `approval_required`. Return an approved request id to retry once
`contextOf`	the AI SDK call options	Pick the per-call context passed to `agentId`
`onDecision`	none	Log each outcome. Errors thrown here are ignored

If authorize throws, the call is refused with code guard_error. The guard fails closed.

Approval

const tools = guardAiSdkTools(rawTools, {
  authorize: auth.authorize,
  agentId: agent.id,
  requestApproval: async ({ agentId, action, resource }) => {
    // Open a request, wait for a human, return the approved id.
    return await waitForApprovedRequestId({ agentId, action, resource });
  },
});

The approved id is spent on a single retry. If the function returns undefined, the model gets a denial with approvalRequired: true.

Last modified on October 19, 2026

Identity

Authorization

Observability

Controls

Tutorials

Vercel AI SDK helper

Runnable snippet

Options

Approval

​Runnable snippet

​Options

​Approval

Runnable snippet

Options

Approval