MCP for Agencies: Audit 20 Client Accounts in an Afternoon

MCP for agencies works as a way to run the same audit checklist across many client accounts quickly, as long as access is prepared in advance and a person reviews the output. MCP (Model Context Protocol) is an open protocol that lets an AI assistant such as Claude read data from tools like Google Ads, GA4 and Meta Ads through connectors called MCP servers. It does not replace the audit judgment. It removes the part where someone logs into 20 interfaces and copies numbers into a spreadsheet.

This article is a practical workflow: what to prepare, how to run the afternoon, what to check by hand, and where this approach breaks.

What does an MCP audit workflow for agencies look like?

The workflow has four stages:

1. Prepare access (days before). Read-only credentials for every client, on every platform you audit.
2. Define the checklist (one hour). A fixed list of questions, identical for every client.
3. Run the audits (the afternoon). One prompt template, run once per client, results saved in the same format.
4. Review and roll up (the end of the afternoon). A person checks the findings against the platform interfaces and builds a cross-client summary.

A planning budget helps: if you give yourself four hours for 20 accounts, that is roughly 12 minutes per client including review. That is a target to test on two or three pilot clients, not a benchmark. If the pilot takes 30 minutes per client, you have a one-day job, not an afternoon job.

What do you need before you start?

• An MCP client. Claude Desktop or Claude Code are the usual choices for agency work. Anthropic's documentation says custom connectors using remote MCP are available on Free, Pro, Max, Team and Enterprise plans, with Free limited to one custom connector. On Team, only Owners and Primary Owners can add a connector, and members then connect individually. Check Anthropic's current help pages for your plan.
• Platform access for each client. Covered in the next section.
• A person who can read the results. The assistant produces a draft. Someone who knows paid media and analytics must judge it.
• A written client data policy. Check your contracts and data processing terms before pushing client data through any AI service. This is a legal and trust question, not a technical one.

How do you connect the ad and analytics accounts safely?

Start read-only. For an audit you do not need the ability to change anything, and a mistake across 20 accounts is expensive.

Source	Official server	Hosting	Access pattern for many clients
Google Ads	Yes, open source, read-only	Self-hosted (no Google-hosted endpoint)	Manager account (MCC) plus login customer ID is the standard pattern. Verify in Google's docs.
GA4	Yes, labelled experimental, read-only	Runs locally	Service account added as Viewer on each property
Meta Ads	Yes, hosted, open beta	Hosted by Meta, OAuth sign-in	Access through Business Manager permissions. Verify details in Meta's Business Help Center.

Google Ads

Google publishes an open-source MCP server that is read-only: it can report on accounts and run queries, but it cannot change bids, pause campaigns or create assets. You host it yourself, which needs a Google Cloud project, a developer token and OAuth or service-account credentials. Sources disagree on exactly how many tools it exposes, so check the official repository. Developer token access level and API quotas matter when you query many accounts, so confirm those in Google's documentation before the audit day. Our guide Connect Claude to Google Ads with the Google Ads MCP Server covers the setup in more detail.

GA4

Google's Analytics server is experimental and runs locally. It wraps the Admin and Data APIs and covers account summaries, property details, custom dimensions and metrics, standard reports, funnel reports and realtime reports. The practical agency step: create one service account and add its email as a Viewer on every client property. That is a per-client onboarding task, so do it early. Because the server is local, it fits Claude Desktop and Claude Code, not web chat tools, unless you add a wrapper.

Meta Ads

Meta launched official AI connectors on 2026-04-29: a hosted MCP server and a command-line tool, in open beta, with OAuth sign-in. Reports describe around 29 tools, including write access. Do not copy an endpoint URL from a blog post. Sources disagreed, so confirm the current address in Meta's Business Help Center. For an audit, restrict permissions to reporting where Meta allows it, and test on one client first. See Meta Ads MCP: Connect Claude or ChatGPT to Your Ad Account for the connection steps.

Third-party multi-account layers

Several vendors sell hosted MCP layers that cover many platforms and claim features like staged writes or approvals. Treat their safety claims as vendor marketing until you have tested them, and ask where client data is stored and who can see it.

How do you run the afternoon, step by step?

1. Write the checklist. Pick 8 to 12 questions you ask of every account. Examples: which campaigns spent money with no conversions in the last 30 days, which campaigns changed most in spend versus the previous period, are conversion events firing in GA4, do GA4 sessions by channel roughly match what the ad platforms report, are there obvious naming or structure problems.
2. Turn it into one prompt template. State the date range, the client name and the account IDs as variables. Ask for a fixed output: a short table of findings, each with the metric, the number, the source account and a confidence note. Ask the assistant to say "not available" when a tool cannot return the data, rather than estimating.
3. Pilot on two clients. Time it. Compare the output to what you see in the interfaces. Fix the prompt, not the results.
4. Run each client in a fresh conversation. This keeps one client's data out of another's context and avoids the assistant blending accounts. Name the output files consistently.
5. Save the raw queries. If the server shows the queries it ran, keep them with the report so a colleague can reproduce a number.
6. Spot check by hand. For every client, verify at least the headline numbers (spend, conversions, the top flagged campaign) in the platform interface.
7. Roll up. Once all reports exist, ask the assistant to compare them: which issues repeat across clients, which accounts are the highest priority. Review this summary too.
8. Convert findings to actions. Findings become tasks in your normal ticketing flow. The audit itself changes nothing in any account.

Our article Ten MCP Use Cases for E-commerce Marketing Teams has more ideas for what to ask once the connections work.

What can go wrong?

• Wrong or invented numbers. The assistant can misread a result or summarize loosely. This is why the manual spot check is a step, not an option.
• Numbers that differ between platforms. Google Ads, Meta and GA4 use different attribution and time zones. An audit that treats them as one truth will produce false alarms. Tell the prompt to report each platform's numbers separately.
• Quota and access limits. Querying many accounts in a short time can hit API limits. Developer token access level and Meta beta limits may change, so confirm them before the day.
• Access gaps. The most common failure is a client where the service account or Business Manager permission was never granted. Check access for all 20 clients the day before.
• Accidental writes. Meta's server reportedly supports write actions. Use read-only permissions and never run an audit prompt in a session where write tools are enabled without a deliberate reason.
• Context overload. Very large accounts can return more data than a conversation handles well. Ask for aggregated queries (by campaign, by week) rather than row-level exports.
• Client confidentiality. Using AI tools on client data may need client consent or a contract change. Settle this before the audit.
• Version differences. The official MCP blog lists 2026-07-28 as the current spec version, but clients and servers negotiate versions and vendors adopt them at their own pace. Do not assume every server behaves the same way, and retest after updates.

Is Claude a good fit for agencies?

Claude for agencies works well for this kind of task because it can connect to several MCP servers in one session and follow a structured prompt. The limits are practical. Local servers need to be installed on each person's machine, remote connectors depend on plan and role permissions, and the output still needs review. Other MCP-capable clients such as ChatGPT or Cursor can also connect to servers, but support differs by server and plan. Check each vendor's documentation. For what Claude can and cannot do inside Google Ads specifically, see Claude for Google Ads: What It Can and Cannot Do.

What should you not automate?

Automated account audits are good at finding candidates for attention: spend with no return, tracking gaps, odd changes. They are poor at deciding what a client should do. Business context (margins, stock, seasonality, contract terms) is not in the ad account. Keep recommendations, budget changes and client communication with people.

FAQ

Use the questions below as a quick check before you schedule the afternoon.