01
Verify prerequisites first
Before using this guide, read the Developer Assistant overview. You need normal Google Ads API access first: a developer token, OAuth configuration, the customer or manager IDs you intend to query, Git and Python 3.10 or newer for the default Python execution path.
The full v4 Assistant is currently documented for Antigravity and Claude Code. Google Ads agent skills have broader host support, but that is a separate layer. Fix authentication with a normal API call before adding the AI tooling.
02
Install the current v4 plugin
Clone Google's current repository and use the platform-specific installation script documented for your host. The official guide currently shows ./install.sh agy for Antigravity or ./install.sh claude for Claude Code on Linux/macOS, with PowerShell equivalents on Windows.
Version 4 is a major plugin-architecture change and is not backward compatible with older project-bound releases. If upgrading, follow Google's current migration guidance rather than carrying legacy settings into the new plugin tree.
03
Configure credentials and customer context
Keep the client-library configuration in the expected home-directory location, such as google-ads.yaml for Python. Keep developer tokens and OAuth secrets out of prompts, screenshots and repositories.
If you repeatedly use one customer, configure the intended customer context. For manager-account workflows, resolve the MCC hierarchy before trusting results so the agent does not report against the wrong client account.
04
Start with a validated read-only report
Use a low-risk first request such as enabled campaigns with impressions, clicks, cost and conversions for the last seven days. Inspect the generated GAQL and let the Assistant validate fields and compatibility against the active schema before execution.
When a field is ambiguous, use the schema/object inspection workflow instead of guessing. Compare the first returned report with another trusted account view to make sure authentication, customer context and metric interpretation are correct.
05
Keep mutation operations behind human review
The Assistant can generate mutate code, but the documented safety model does not automatically execute those account changes. Review generated campaign-changing code separately before running it.
A practical division of labor is: agent skills for Google Ads instructions, MCP for controlled read-only data access and the Developer Assistant for the larger coding, validation and diagnostic workflow. Keep credentials isolated and preserve a human approval gate for consequential changes.
Sources
Primary and supporting sources
Facts were rechecked against the linked sources immediately before publication. Pricing, product availability and rollout status can change.