Use Codex for Google Ads reporting and repeatable account reviews
Codex can work with Google Ads through an MCP connection as well as with the files in your project. That makes it useful for teams who want a repeatable campaign review, a written decision record, or a reporting workflow alongside their product work. The advertising access still comes from the Adport project you authorize.
Configure the hosted Adport server
First connect Google and enable the intended accounts in Adport. Then register the hosted MCP server in Codex:
codex mcp add adport --url https://app.adport.dev/mcpIf authorization has not started or the connection needs a fresh sign-in, run:
codex mcp login adportCheck the registered connection with codex mcp list. Complete OAuth for the intended project. OpenAI documents MCP configuration and OAuth sign-in for Codex.
Make account scope part of the workflow
Keep a small working brief with the client project, account IDs, currency, primary conversion, and reporting period. A file can hold that context without holding credentials. Do not put provider refresh tokens or client secrets in a repository.
For an agency, name separate connections clearly if your MCP client supports more than one server entry. A server configuration is not a substitute for project authorization. Each grant should cover the client you intend to inspect, and client-facing access should be limited independently of your own editing access.
If you use the local open-source Adport server, the setup and credential ownership differ from the hosted OAuth flow. Choose the local instructions in Agent access or the repository documentation rather than mixing local stdio configuration with an HTTP endpoint.
Start each review with an account and data check
Use Adport to confirm the Google ad accounts available to this project. For the selected account, read campaign performance for the last seven completed days. Show currency, missing metrics, provider warnings, and any truncated results. Do not write changes.
Inspect the returned account IDs before accepting the analysis. Then compare a small set of campaign rows with the platform's own report. This is especially useful after a new authorization, account-selection change, or conversion-tracking update.
Keep partial results visible in the output. A report that failed for one account is still useful for the accounts that succeeded, but it does not support a complete client-wide total.
Use a consistent review note
A practical review file has five sections: scope, observations, hypotheses, proposed actions, and verification. Observations should contain dates and source rows. Hypotheses should name the evidence still needed. Proposed actions should be specific enough to preview individually.
For example, a campaign's CPA rising from one period to the next is an observation. A change in lead quality is a separate question that may require CRM data. Keep the two separate so the assistant does not present an unverified explanation as a fact.
Write a campaign review with scope, observed changes, hypotheses, proposed actions, and verification steps. Reference the account and campaign IDs for every recommendation. Keep currencies separate. List missing business or attribution data before recommending a budget change.
Use a team-approved destination for client reports. A local report file is not automatically suitable for a public repository or a shared folder with other clients.
Preview, approve, and verify one change
Ask Codex to preview a supported change for one exact account and campaign. Review its values, budget effect, validation result, and policy adjustments. If you approve, the applying call must reuse the matching pending operation and identical arguments.
A general instruction to improve performance should not become standing approval for unrelated changes. When the intended action changes, create a fresh preview. After applying it, read back the platform state and check the audit history.
Codex's own tool permissions and Adport's structural write gate serve different parts of the workflow. The advertising operation still needs valid project access, a suitable cloud plan and role, enabled accounts, and compliant policy limits.
Make the next review easier
Record the action, reason, and follow-up date. On the next run, ask Codex to check the previous decision before proposing new work. This helps avoid repeated changes that never receive a fair observation window.
The audit checklist supplies the investigation order. The agency project guide explains how to keep the same review pattern separate across clients.
Sources & further reading
Provider documentation for the setup and platform behavior described above.
- OpenAI: MCP configurationlearn.chatgpt.com
- Model Context Protocol: server conceptsmodelcontextprotocol.io