Use case
An SEO agent that runs every week and cannot overspend
Run Claude Code, Codex or Hermes as a scheduled SEO agent that opens pull requests for fixes, with a dollar cap and approval before anything merges.
An SEO agent is the classic recurring job: check the site, find broken links, thin pages and missing metadata, and propose fixes. It is also a job people hesitate to automate, because an agent with write access to the site and an open-ended model key can do a lot of damage quietly.
What does the setup look like?
One fleet with one agent, on a schedule:
fleet: seo-weekly
runtime: claude-code
goal: "Find and fix SEO issues on example.com"
triggers:
- schedule: "0 6 * * 1"
tools:
github: { allow: [pulls.create], require_approval: [pulls.merge] }
budget: { monthly_usd: 30, per_run_usd: 0.50 }
The agent reads the site and the repository, then opens pull requests. Opening a pull request is reversible, so it runs within policy. Merging needs a person, in Slack, ChatGPT, Claude or the console.
What changes compared with a cron job?
| Cron on a VPS | Heighliner | |
|---|---|---|
| Spend limit | Whatever the model key allows | $30 a month, $0.50 per run, hard stop |
| Write access | The token in .env, every scope | Pull requests only; merge needs approval |
| Between runs | The box stays on | The agent sleeps |
| Record | Whatever the script logs | A receipt for every run |
Where does it stop?
At the cap, the agent pauses and you get a warning first at 50, 75 and 90 percent. It never touches the main branch without approval, and every action lands in the audit trail.
Updated
Questions
- Can the SEO agent push straight to my site?
- Only if you allow it. A common setup lets it open pull requests freely and requires approval to merge, so every change is reviewed before it ships.
- Which agent should I use for SEO?
- Any harness you already trust: Claude Code, Codex, Hermes, OpenCode, a custom container or your own endpoint. Heighliner governs the agent, it does not replace it.
Related
- Nightly reports that are drafted automatically and sent deliberatelyRun an agent every night to pull metrics and draft a report. Sending email to customers is irreversible, so it waits in escrow for a person to approve.
- Why agents need a budget, not a dashboardA dashboard tells you what an agent spent after the fact. A budget stops it before it overspends. How dollar caps and leases make agents safe to leave running.
- FleetA fleet is one or more AI agents running one harness under one manifest, one budget and one policy. A single scheduled agent is a fleet of one.
- Heighliner vs a VPS, a cron job and a .env fileAn honest comparison of running AI agents on your own VPS with cron and running them on Heighliner: cost, control, budgets, approvals and when the VPS is the better choice.