Growth Mind Ads MCP

Manage supported Google Ads changes from an AI agent through a local MCP server. Built from Growth Mind's operating workflow: read the account, prepare a specific change, validate it, get the account owner's approval, apply the unchanged batch, and read the result back.

This is the management bridge, not a complete marketing platform. Pair it with a reporting connector such as Google's Google Ads MCP. Your brand workbench supplies the customer, offer, voice and goals. Separate tools provide your website, CRM, analytics and keyword data.

Start here

• Installing with an agent? Give it docs/AGENT_SETUP.md.

• Installing yourself? Follow installation, Google access, then Codex/Claude configuration.

• Using the tools? Read the exact change protocol and workflow recipes.

• Writing application documents? Use the privacy and app-description worksheet.

• Checking the trust boundary? Read SECURITY.md before allowing writes.

Related MCP server: Google Ads MCP Worker

Quick local check (no Google credentials needed)

Requires Git, uv and Python 3.11+. The repository selects Python 3.12.

git clone https://github.com/bryce-liggins/growth-mind-ads-mcp.git
cd growth-mind-ads-mcp
uv sync --frozen
uv run --frozen python -m unittest discover -s tests -v
uv run --frozen python scripts/smoke_mcp.py

The tests mock Google calls. The smoke test starts a real local MCP process, discovers its five tools and requests capabilities using a fictional customer ID. It neither authenticates to Google nor changes an account.

After Google setup, check configuration without exposing values:

uv run --frozen --env-file /absolute/path/to/private/ads.env growth-mind-ads-doctor

Then register the server in your chosen agent client. The server speaks stdio, not HTTP; it does not open a public listener.

Supported operations

Allowlisted create, update and remove operations cover budgets, campaigns, ad groups, ads, criteria, assets/associations and shared negative sets. New campaigns, ad groups, ads and positive ad-group criteria are forced paused. Negatives and some other resources can affect live delivery immediately after application.

Conversion actions are deliberately narrow: create a secondary completed-booking website action, or change an existing action's primary/secondary role. No offline customer-data uploads, conversion-action removal, billing or account-user management. Full operation reference.

Approval is a workflow, not an independent human signature

Validation produces a single-use, eight-hour ticket bound to the normalized customer and payload. Application consumes that ticket and validates with Google again before writing.

The server cannot prove that a human approved. An agent with both tools can request and consume a ticket. Configure your agent/host to require the account owner's approval before either apply tool, and don't give untrusted agents unrestricted credentials. Paused-create defaults do not protect against every live effect. See SECURITY.md.

What is verified

This release includes mocked operation/approval tests, configuration/header/audit tests, and a real local stdio discovery test. See VALIDATION.md for recorded results and limits. Automated tests do not establish live Google access, a successful account mutation, marketing performance, or successful sign-in inside every agent product.

The package does not install a scheduler, deploy a website, connect a CRM, or provision your Google application. The documentation walks through those boundaries and the evidence to collect.

License and attribution

MIT, copyright 2026 Bryce Liggins. Dependencies retain their own licenses; see THIRD_PARTY.md. This is an independent Growth Mind project, not an official Google, OpenAI or Anthropic product.