Set up AdRails for a client with an AI agentTwo commands · adrails.dev

What the agent runs

1. Install and scaffold one campaign

   init reads the campaign's name, status, end date and budget, and writes three files in a new git repo. It refuses to overwrite anything. Use one directory per campaign so each client's caps and history stay separate. --json gives machine-readable output.

   shell

   pipx install adrails
   adsops-mcp init --dir ~/adrails/acme --customer-id 123-456-7890 \
     --campaign-id 2233445566 --sa ~/.adsops/sa.json --json

2. Check everything with doctor

   doctor prints one row per check: the key file, whether the developer token is set (never its value), live access, the files, clean git and the guard heartbeat. Each failure carries the exact fix; relay it to the human as written. Exit code 0 means every check passed.

   shell

   adsops-mcp --config ~/adrails/acme/config.json doctor
   PASS  service_account_file  service account: ads-bot@acme.iam.gserviceaccount.com
   FAIL  developer_token  missing
         fix: Google Ads > Tools > API Center: copy your developer token, then export ...

3. Register the server and read a plan

   Register the local server in the client, then call adsops_status and adsops_plan. Because the state file matches the live campaign, the first plan has zero operations.

   shell

   claude mcp add adrails -- adsops-mcp --config ~/adrails/acme/config.json

4. Hand the caps to the human

   The policy file starts conservative: the budget cap equals today's budget, enabling is off, a plan ID is required and commits are off. The human edits the caps they want, sets commit_policy.allow_commit and allow_set to true in config.json, and commits. The agent must not edit the policy file, and no AdRails tool writes it.

5. Schedule the spend guard

   Run adsops-mcp --config ~/adrails/acme/config.json guard-all every 15 minutes from launchd, systemd or cron (templates in the repo's deploy/). Until the guard has a fresh heartbeat, budget raises and re-enables are refused.

AdRails reduces risk; it does not remove it. The guard is a threshold with automatic pause attempts, not a hard spend ceiling, and Google can deliver up to 2x the average daily budget on a given day. Keep a billing limit in Google Ads as the real ceiling. See what the guardrails do and don't prevent.

Agent setup questions

Can an AI agent sign a client up for AdRails?

There is nothing to sign up for. AdRails is free, open-source software with no account. The agent installs it and runs two commands; the human provides the three Google items once.

Can the agent raise the budget after setup?

No. The budget cap starts at today's budget and lives in the policy file, which only a human edits. A raise past the cap is refused with outside guard range.

Does it work for an agency with many clients?

Yes. Add the service account under your manager account, pass --login-customer-id, and run init once per campaign in its own directory.

Does it work with ChatGPT?

Not directly. AdRails runs as a local process and has no hosted URL by design, so it needs a client that can launch one: Claude Code, Claude Desktop, Cursor, the Grok CLI and others. See MCP clients.

Related docs

• Install AdRails with Claude Code — the same setup by hand.
• What AdRails guardrails do, and don't, prevent — the policy file, the guard and the limits.
• How to let an AI agent manage Google Ads safely — the rails and the checklist.
• Use AdRails with any MCP client — one local stdio command.
• AdRails docs index — every guide on one page.