Agentic Marketing for Google Ads and Meta Ads: Stop Budget Leaks

What agentic marketing actually needs

An agentic workflow has a defined responsibility, access to relevant data, business targets, a repeatable deliverable and a trigger. MCP—Model Context Protocol—provides the connection to external tools. Your assistant or agent runner supplies the surrounding workflow.

For example, a search-intent agent can retrieve queries, assess them against your offer and prepare proposed negative keywords. A marketer checks the intent and exclusion risk before implementation.

• Data: the authorised advertising accounts and available reporting tools.
• Context: the offer, geography, conversion definition, targets and approved budget.
• Output: evidence, proposed action, account and entity IDs, confidence and review date.
• Control: enforceable permissions, human approval for changes and an implementation log.
• Cadence: a manual request or a separately configured recurring run.

The official Google Ads and Meta Ads MCPs

A commercial connector can use the official advertising APIs without being a Google or Meta product. The setup below uses the platform-provided options. Google’s MCP documentation and Meta’s MCP documentation describe their respective capabilities.

Connect the official Google Ads MCP

This route runs Google’s server locally and requires an MCP client that can launch a local stdio server. If your assistant only accepts remote connectors, use Google’s Cloud Run deployment guide instead. The local configuration does not create a public server URL.

1. Prepare account access and Google Cloud

Create or select a Google Cloud project, enable the Google Ads API and confirm the appropriate API access level. The authenticated Google identity must also have access to the advertising accounts. Create an OAuth desktop application and download its client credentials JSON. Follow the current Google prerequisites; older guides may use different credential requirements.

2. Install the prerequisites and authenticate

Install Python, pipx and the Google Cloud CLI. Create Application Default Credentials with the required scopes:

gcloud auth application-default login \
  --scopes=https://www.googleapis.com/auth/adwords,https://www.googleapis.com/auth/cloud-platform \
  --client-id-file=/absolute/path/google-oauth-client.json

Replace the path with your downloaded OAuth client file. Sign in with the Google account that has advertising access and note the path where the command saves the resulting credentials. The official repository documents this authentication method.

3. Register the server in your MCP client

For a client accepting mcpServers-style JSON, add:

{
  "mcpServers": {
    "google-ads-mcp": {
      "command": "pipx",
      "args": [
        "run",
        "--spec",
        "git+https://github.com/googleads/google-ads-mcp.git",
        "google-ads-mcp"
      ],
      "env": {
        "GOOGLE_PROJECT_ID": "YOUR_GOOGLE_CLOUD_PROJECT_ID",
        "GOOGLE_APPLICATION_CREDENTIALS": "/absolute/path/adc-credentials.json"
      }
    }
  }
}

Set GOOGLE_APPLICATION_CREDENTIALS to the ADC file produced by authentication, rather than the original OAuth client file. If you access accounts through a manager account, you can also set GOOGLE_ADS_LOGIN_CUSTOMER_ID. The client determines the configuration filename and location.

4. Test the reporting connection

Reload the connection and ask:

Show the Google Ads accounts this connection can access. For account [CUSTOMER_ID], report spend, clicks and primary conversions for the last seven completed days. Include the currency, timezone and exact reporting dates. Make no changes.

Google is read-only: approval does not add write tools

The current Google Ads MCP exposes reporting and discovery tools. It does not expose tools to change budgets, add negative keywords, pause ads or publish campaigns. Giving the authenticated user administrator access does not add these missing capabilities. See Google’s current toolset.

Approved Google changes can be implemented in the Google Ads interface or through a separately built integration with the official Google Ads API. The API’s campaign-management capabilities are separate from this MCP’s read-only tools. An agent must never report a proposed change as completed.

Connect the official Meta Ads MCP

Meta hosts its server at https://mcp.facebook.com/ads. Availability can vary by business. Supported assistants can use Meta’s advertiser connection route without requiring you to own a developer app; follow the client-specific links in Meta’s advertiser setup guide.

Set account rules before enabling changes

Where available, someone with full control of the business portfolio can open Meta Business Suite → Settings → Integrations → Ads MCP server. For an initial reporting workflow, block actions such as budget edits and campaign creation. Meta’s server enforces these rules. If the menu is absent, the feature may not yet be available to that business.

Developer-app connection

1. Create or select a Meta developer app and add the Create & manage ads with ads MCP server use case.
2. Configure your MCP client’s OAuth redirect URL in Facebook Login for Business.
3. Connect to the official endpoint and complete authentication.

Meta documents this example for Claude Code:

claude mcp add --transport http \
  --client-id YOUR_META_APP_ID \
  meta-ads https://mcp.facebook.com/ads

For the developer-app route, Meta lists ads_mcp_management plus ads_read or ads_management as minimum token permissions. Additional tools can require further permissions. Follow the official authentication guide.

Test in analysis mode

Identify the accessible Meta ad accounts. For [AD_ACCOUNT_ID], report campaign spend and results for the last seven completed days. State the result definition, attribution settings where available, currency and timezone. Do not create, edit or activate anything.

Meta’s campaign, ad-set and ad creation tools create paused entities; activation is a separate operation that starts delivery. Existing-entity changes still need review. Combine account rules with the assistant’s tool confirmations. See Meta’s management tools.

Original SEM Lion prompts you can adapt

Start with prompt 01, then add the task prompt you need. Replace every placeholder. Written instructions guide the assistant; tool permissions and platform rules enforce restrictions. Choose a primary conversion that matches the business: a qualified lead is different from a form submission, and revenue is different from profit.

Act as the SEM Lion paid media analyst for [BUSINESS].
Authorised accounts: [PLATFORM AND ACCOUNT IDS].
Business model, offer and locations: [DETAILS].
Primary conversion: [PURCHASE / QUALIFIED LEAD / OTHER].
Targets: [TARGET CPA OR ROAS, WITH DEFINITION].
Approved budgets: [AMOUNT, CURRENCY, PERIOD, ACCOUNT].
Typical conversion delay: [KNOWN DELAY OR UNKNOWN].
Reporting timezone: [TIMEZONE].

Use connected tools and supplied evidence only. Inspect available tools first.
Identify missing inputs before making recommendations.
Use completed days and explain the effect of conversion delay.
Check monetary units before calculating.
Keep accounts, currencies and conversion definitions separate.
Distinguish observed facts from possible explanations.

Work in analysis mode. Do not change account settings, budgets,
keywords, campaigns, ads or delivery status.
Return evidence, recommendations and unanswered questions.

Review the available search-term data for [GOOGLE_ACCOUNT_ID]
over [DATE_RANGE].

Assess queries against our actual products, services, locations
and customer intent. Check existing keywords and negatives where
the tools expose them.

Group findings into relevant demand to retain; potentially irrelevant
intent needing review; promising queries needing a keyword or landing-page
review; and insufficient evidence.

Consider conversion delay and sample size. Zero conversions alone
is not enough to recommend exclusion.
For each proposed negative, show the query, evidence, proposed match type,
intended scope and risk of blocking useful traffic.
State reporting coverage limitations. Produce a proposal only.

Review Meta ads in [ACCOUNT_ID] across two comparable completed periods:
[CURRENT_PERIOD] and [COMPARISON_PERIOD].

Compare ads within similar objectives, audiences and placements.
Use available spend, impressions, frequency, click and conversion metrics.
Identify which CTR definition you use.

Investigate possible creative fatigue alongside audience changes,
placement mix, recent edits and measurement issues.
Propose three creative experiments. For each, specify the observation,
hypothesis, element to change, element to hold constant, success measure
and review point.

If creative assets cannot be inspected, say so. Do not invent details
about their content or launch tests.

Compare month-to-date spend with the approved budget for each
authorised account. Exclude incomplete days.

Show spend to date, remaining approved budget, calendar-based pacing,
a projection using the last seven completed days and assumptions
that could change the projection.

Account for planned promotions, weekday patterns and known outages
when that information is supplied.
Present maintain, reduce and reallocate scenarios where justified.
Keep totals within the approved plan and separate currencies.
Do not propose moving money between different business goals solely
because one platform reports a higher ROAS.

Treat forecasts as estimates. Do not change budgets.

Write a weekly decision brief using [CURRENT_PERIOD] versus
[COMPARISON_PERIOD]. Lead with the most consequential finding.

Separate Google and Meta platform-reported results. Do not add their
attributed conversions and label the total unique customers or
incremental sales.
Select up to three priorities. For each, include evidence, confidence,
recommended action, owner and review date.
Use the supplied change log to investigate explanations. Label unsupported
explanations as hypotheses. Include "no change recommended" where
evidence is insufficient.

End with an action table containing account ID, entity ID, current setting,
proposed setting and execution method. Mark Google changes as requiring
implementation outside its read-only MCP. Execute nothing.

Execute only the Meta changes explicitly approved in [APPROVED_TABLE].

Before each change, verify the account, entity ID and current value
against the approved record. Stop that item if anything differs,
permissions block it or the field is unsupported.
Do not substitute another action. Do not activate anything unless
activation is explicitly approved.

After execution, retrieve the affected entities again and report
confirmed changes, failed changes and unresolved outcomes.
Do not retry an uncertain write until its outcome is checked.

Use the optional execution prompt only after deliberately enabling the required Meta write permissions and approving a concrete action table. It does not apply to the read-only Google MCP.

A weekly verification routine that catches budget leaks

The purpose of a schedule is to reduce the time between a problem appearing and a reviewed response. A practical starting point is a short exception check each working morning, followed by one focused task per day.

The example trigger time is Europe/Bucharest; reporting comparisons must still respect each advertising account’s timezone. Configure a scheduler or agent runner with the authorised connections, stored instructions, output destination and failure alerts. Writing “run every Monday” in a prompt does not itself create a schedule.

What to record after every review

• Evidence: account and entity IDs, reporting dates, data freshness, currency and monetary units.
• Measurement: primary conversion, attribution window where available, conversion delay and relevant limitations.
• Decision: proposed action, business rationale, approval, responsible person and implementation method.
• Verification: actual setting after implementation, failures or uncertainty, and the next performance review date.

Keep platform-reported results separate unless you have a defensible reconciliation method. A higher reported ROAS on one platform is not proof that moving budget there will create incremental sales. When a metric or breakdown is unavailable, the agent should say so.

Stop budget leaks before they compound

Official MCPs make account data accessible. Scheduled verification makes reviewing it a routine. Human approval and a change log connect the findings to accountable action.

That combination matters because wasted advertising spend does not wait for the next monthly report. Irrelevant clicks, deteriorating creatives, unnoticed budget drift and measurement faults can keep consuming money until someone investigates and responds. A useful agentic workflow shortens that delay.

The goal is to stop avoidable money leaks from your advertising accounts—not to produce more dashboards. Start with one account, one recurring check and one clear owner. Measure what changes after implementation, then expand the workflow using evidence.