OpenAI Agents SDK helper

guardFunctionTool wraps a function tool definition (a name plus execute(input, context)) so authorize runs before execute. It does not import @openai/agents. tool() returns an object that exposes invoke, not execute, so guard the plain definition first and pass the result to tool(). For the throwing variant built on the policy engine, see OpenAI Agents SDK.

npm install @glinr/theauth

Runnable snippet

import { Agent, run, tool } from "@openai/agents";
import { z } from "zod";
import { createTheAuth } from "@glinr/theauth";
import { guardFunctionTool } from "@glinr/theauth/integrations";

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

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

const lookupInvoice = guardFunctionTool(
  {
    name: "lookupInvoice",
    description: "Read an invoice",
    parameters: z.object({ id: z.string() }),
    execute: async ({ id }: { id: string }) => ({ id, status: "paid" }),
  },
  { authorize: auth.authorize, agentId: theauthAgent.id },
);

const agent = new Agent({
  name: "Billing",
  instructions: "Answer billing questions.",
  tools: [tool(lookupInvoice)],
});

const result = await run(agent, "Is invoice inv_1 paid?");
console.log(result.finalOutput);

Denials

A refused call returns a ToolDenial object as the tool result, which the SDK serializes for the model:

{
  "ok": false,
  "denied": true,
  "toolName": "refund",
  "code": "approval_required",
  "reason": "approval required for tool:refund",
  "approvalRequired": true,
  "auditId": "..."
}

Use formatDenial to return text instead:

guardFunctionTool(definition, {
  authorize: auth.authorize,
  agentId: theauthAgent.id,
  formatDenial: (d) => `Refused (${d.code}). ${d.approvalRequired ? "Ask the user." : "Do not retry."}`,
});

agentId may be a function of the run context, the second argument of execute. requestApproval, action, resource and onDecision work as in the Vercel AI SDK helper. Use guardFunctionTools(definitions, options) for a list.

Python

The Python Agents SDK has the same guard, guard_function_tool from theauth.integrations. It wraps the FunctionTool that @function_tool returns. See LangGraph and CrewAI for the setup and a snippet.

Last modified on October 19, 2026

Identity

Authorization

Observability

Controls

Tutorials

OpenAI Agents SDK helper

Runnable snippet

Denials

Python

​Runnable snippet

​Denials

​Python

Runnable snippet

Denials

Python