Skip to content

Connect AdPilot to ChatGPT, and why ChatGPT says “high risk”

AI and toolsUpdated 13 Nov 20263 min read

Cover image: a simplified app-creation dialog with a server URL field, an OAuth field, a warning label and a checkbox, next to a shield with three ticks
Contents
  1. Before you connect
  2. Six steps in ChatGPT
  3. Why ChatGPT says “high risk”
  4. Connected, but still read-only
  5. AdPilot added a tool ChatGPT can’t see yet

AdPilot connects to ChatGPT with an MCP address and ALFA’s sign-in screen; there’s no token to hunt for. This article walks through the six steps, explains the high-risk warning ChatGPT shows near the end, and says when you need to reconnect.

1. Before you connect

  • Pick an address. AdPilot for Google Ads has three addresses on the same server: Full, View + propose and View only. Not sure? Pick the first one (source).
  • Your ChatGPT plan. You need an AI app that lets you add custom connectors — some apps only allow this on a paid plan. Don’t have one? Use Studio on the ALFA web app — ALFA pays for the tokens (source).
  • No token. Everyone pastes the same address; after you paste it, ChatGPT opens ALFA’s sign-in screen so you can sign in and pick your organisation (source).

2. Six steps in ChatGPT

ALFA tested ChatGPT with its Vietnamese interface, so the button names below describe what each one does; the exact English labels may differ slightly (source).

  1. Open the plugins page and click Add in the top right corner. A dialog for creating an MCP app opens.
  2. Name: you type it yourself, for example ALFA AdPilot Google. The icon and description are optional, and ChatGPT doesn’t take them from the server; if you want an icon, upload a PNG yourself, up to 10 KB.
  3. Connection: choose server URL, not Tunnel. Paste one of the three addresses.
  4. Authentication: choose OAuth, so ChatGPT opens ALFA’s sign-in screen.
  5. Read the warning, then tick the confirmation box. The dialog carries a red high-risk label and makes you tick an “I understand and want to continue” box before you can create it.
  6. Click Create, then sign in to ALFA and pick your organisation. Done — try a question such as “What needs doing today?” (source).

3. Why ChatGPT says “high risk”

The warning is right. ChatGPT says an untrusted MCP server could take information you share, or trick the AI into using tools you didn’t intend, including changing or deleting data. That’s true of MCP in general, so do read it (source).

AdPilot answers the warning with a mechanism (source):

MechanismWhat it means
No tool writes straight to Google AdsThe AI can only create proposals; a proposal hasn’t touched the account yet.
A person agrees before anything is appliedA proposal nobody approves expires on its own after 24 hours.
High-risk changes must be clicked on the ALFA web appNo AI can approve them on your behalf, ChatGPT included.
Changes can be undoneAdPilot says clearly which parts can’t be restored.
The View only addressStill uneasy? Paste this one: the AI can look, but can’t propose or change anything.

How proposals are approved and applied: see Approve AdPilot proposals.

4. Connected, but still read-only

That isn’t a bug. The real level for each call is the lowest of four things: the permission you gave the AI app, the address you’re using, your organisation’s maximum level in Guardrails, and the plan you’re on. The most common case: your organisation set a Guardrails maximum below the address you pasted — then the Guardrails level wins (source).

5. AdPilot added a tool ChatGPT can’t see yet

The tool list is fetched when you connect, so new tools only show up after you reconnect ALFA AdPilot (source).