MCP server

Your Google Ads findings, inside Claude and ChatGPT.

Connect Claude, ChatGPT or Cursor to Goldbeater’s MCP server, and your assistant reads what the checks found on your accounts, with the evidence and the drafted fix.

Start the 7-day trialSee how to connect

Or connect one account for a free first look, no card needed: how many problems it finds, and what they cost.

The server's address
https://api.goldbeater.ai/mcp
One question
Sample data
You ask your assistant

What is Cedar & Pine Interiors wasting money on?

It calls
list_findings({"customer_id":"882-014-4517"})
Goldbeater answers
{"account": "Cedar & Pine Interiors","findings": [{"check": "Paying for searches that never convert","severity": "high","title": "\"sofa bed free\" has taken $412 and converted nothing","evidence": "Its 64 clicks in the last 30 days brought no conversion, which happens 1% of the time by chance where this campaign's other searches convert 6.7% of clicks.","impact_per_month": 412,"proposes": ["create campaign_criterion.negative_keyword on \"Negatives of Sofas — Phrase\" (customers/8820144517/campaigns/21904417002): [sofa bed free] exact negative"]}]}
The answer, shortened to one finding. Its fix waits in Goldbeater for your approval.
Read-only
The tools read what the checks found, and can’t change your account.
166 checks
They run on your accounts every 24 hours, and whenever you ask.
One sign-in
You sign in with Goldbeater, with no Google Cloud project to make.
No queries
A call reads stored findings, and sends nothing to Google Ads.

What your assistant gets

Your assistant reads the finished analysis, not raw rows.

Google’s own MCP server has your assistant write each query; Goldbeater’s hands it what 166 checks for 274 problems already found.

Ask it

What did Goldbeater find on my account, most costly first?

Every open finding, the most severe and costly first.

What’s the evidence that this conversion stopped recording?

The finding’s own sentence, with the figures and tables behind it.

Which checks ran last time, and did any part of my account go unread?

Every check that ran, and any part the pass couldn’t read.
Connect Google’s own server too

The three tools

Three tools, and all three only read.

Your assistant picks the tool your question needs.

ToolWhat it returns
list_accountsEvery account you can open in Goldbeater, with its currency, when its checks last ran and how many findings are open.
list_findingsWhat the checks found on one account, most severe and costly first, each with its evidence, what it’s worth and its drafted fix.
what_ranEvery check that ran on an account’s latest pass, how each judges, and whether it found anything.

Access

What your assistant can see, and what it can’t.

A connection sees what you see in Goldbeater, and nothing more.

Only your accounts

Every call checks what you can open in Goldbeater at that moment.

Nothing it can start

It can’t start checks or apply a change: those stay with you in Goldbeater.

A consent screen first

Before you allow it, a consent screen names the app that’s asking and the accounts it will read.

A first look sees counts

On a free first look, it sees how many findings there are and what they’re worth.

Yours to withdraw

Disconnect it under Connected apps in your Goldbeater Profile, or remove it in the assistant.

Under your assistant’s terms

What the tools return goes into your conversation, under your terms with the assistant’s maker.
Read the security page

Connect your assistant

Connect Claude or ChatGPT to your Google Ads findings.

Each assistant takes this address and your Goldbeater sign-in; Claude calls it a custom connector.

The server's address
https://api.goldbeater.ai/mcp

Add it from Cursor Directory (opens in a new tab), or find it in the official MCP Registry (opens in a new tab). An AI agent installs its skill from skills.sh (opens in a new tab) or GitHub (opens in a new tab).

Connect Claude on the web and in the desktop app

To connect your Google Ads findings to Claude, add Goldbeater as a custom connector. It works on every plan, and the Free plan allows one custom connector. On Pro and Max:

  1. Go to Customize › Connectors.
  2. Select the + button, then Add custom connector.
  3. Name it Goldbeater, paste the server’s address, and select Continue.
  4. Keep what Claude detects, Sign in now and Use Claude’s published identity, and select Add.
  5. Select Connect, sign in to Goldbeater if it asks, check the accounts the consent screen lists, and select Allow.
  6. Ask about an account. Claude asks before it uses each tool; all three only read, so Always allow saves the question. If a conversation doesn’t use Goldbeater, turn it on from the + button, under Connectors.

Questions

What to know before you connect an assistant.

How is this different from Google’s Google Ads MCP server?

Google’s own Google Ads MCP server lets an assistant query your account: it writes each query itself, so it has to know what to look for. Goldbeater’s hands it the finished analysis instead: what the checks found, each finding with its evidence, its figures and tables, what it’s worth and the fix Goldbeater drafted. It’s the same reading Goldbeater’s AI analyst works from. The two work side by side.

How do I connect Claude on a Team or Enterprise plan?

An Owner adds it first, under Organization settings › Connectors: Add, then Custom, then Web, and the server’s address. Each member then finds Goldbeater under Customize › Connectors and selects Connect.

Can my assistant change my Google Ads account?

No. The tools only read. Changes are made in Goldbeater, where a person approves each one and presses Apply, and an applied change can be undone.

How a change is approved and applied

Does a call count against my Google Ads API limits?

No. A call reads the findings Goldbeater stored from the latest pass, and sends nothing to Google Ads.

How fresh are the findings?

As fresh as the latest pass. The checks run every 24 hours and whenever someone presses Run checks in Goldbeater, and each account says when its checks last ran.

How does my assistant read the money in a finding?

Money is in each account’s own currency, and only wasted spend is money lost. The server tells your assistant never to add different kinds of money together, and to say an estimate is an estimate.

What does a workspace on its free first look see?

What the app shows it: how many findings the checks raised and what they’re worth. The findings themselves open with the free trial.

Which accounts can my assistant read?

The accounts you can open in Goldbeater, and no others. Every call checks what you can open at that moment, through the same workspace rules as the app, so an account you lose access to, or a workspace you leave, is out of its reach at once.

Where does what my assistant reads go?

Into your conversation with the assistant, under the terms you have with its maker, as anything you paste into it does.

What the security page says about connected assistants

Sources

Every step was read from its maker’s own documentation.

We name these pages rather than link them.

  1. [1]Claude Help Center: Get started with custom connectors using remote MCP

    • Custom connectors work in Claude on the web, in the desktop app and in Cowork, on the Free, Pro, Max, Team and Enterprise plans, and Free allows one.
    • On Pro and Max, a connector is added under Customize › Connectors: the + button, Add custom connector, the server’s address, then Add.
    • On Team and Enterprise, an Owner adds it first, under Organization settings › Connectors (Add, then Custom, then Web), and each member then connects to it under Customize › Connectors.
    • Claude reaches a connector’s server from Anthropic’s cloud, whichever Claude app you use, and a conversation turns connectors on from its + button, under Connectors.

  2. [2]Goldbeater's own connection from Claude on the web

    • After the name and address, Claude detects that the server signs you in and reads Claude’s published identity, and a new connector’s tools ask for approval each time until you choose Always allow.

  3. [3]Claude Code documentation: Connect Claude Code to tools via MCP

    • Claude Code’s mcp add command adds a remote server over the http transport, and the user scope makes it available in every project.
    • Inside Claude Code, /mcp runs a server’s sign-in in your browser, and claude mcp login does the same from your shell.

  4. [4]OpenAI developer documentation: ChatGPT developer mode

    • Developer mode is available to Plus, Pro, Business, Enterprise and Education accounts on ChatGPT’s website, and is turned on under Settings › Security and login.
    • An app for a remote server is made from the + button on ChatGPT’s Plugins page, and a conversation uses it from Developer mode in its + menu.
    • ChatGPT signs in with OAuth, by a client ID metadata document or dynamic client registration where the server supports them.
    • ChatGPT treats a tool marked read-only as a read, and asks you to confirm a tool that writes.

  5. [5]OpenAI developer documentation: Connect and test your plugin (ChatGPT)

    • An app takes a name, a description and, under Connection, the server’s address with its /mcp path, and whether developer mode is available can depend on a workspace’s policy.

  6. [6]Cursor documentation: Model Context Protocol (MCP)

    • A remote server is an entry with a url in mcp.json: ~/.cursor/mcp.json for every project, or .cursor/mcp.json in one.
    • Cursor signs in to a remote server with OAuth, registering itself unless it’s given a client’s credentials, and asks before it uses an MCP tool unless told otherwise.
    • Servers are turned on and off from Customize, in Cursor’s sidebar.

7-day trial · nothing charged for 7 days

Try it for a week. See the money it saves. It pays for itself.

Start the 7-day trial