Scheduled Astro maintenance

Let desktop agents maintain simple Astro websites safely.

Simple informational sites are good candidates for recurring agent maintenance when repository scope, tools, secrets, stop conditions, staging, and public verification are all explicit.

View the source toolkit Explore the guide

What the process produces

Durable repository instructions
Scoped filesystem and Git access
Package and browser testing tools
Cloudflare and PageSpeed API access
Staging-first deployment policy
Human-readable run report

Choose a scheduler that can reach the repository

Use Codex automations when they are available in your Codex workspace. ChatGPT Scheduled Tasks are created and managed on ChatGPT web or mobile, not the desktop Scheduled tab, and tasks cannot access files attached to a project. For repository maintenance, use Codex automation or an API or CI scheduler that starts an agent with the checkout and tools in scope.

Keep each site in version control
Schedule no more often than the maintenance need justifies
Run in one dedicated task or thread per site
Require a report even when no change is made

Connect only the tools the workflow needs

A local agent needs scoped file and terminal access. Add GitHub, Cloudflare, PageSpeed, browser testing, and secret-manager tools only when the site uses them. Claude Desktop supports local desktop extensions and remote MCP connectors, but Anthropic does not document a general recurring scheduler in Claude Desktop. Pair Claude Desktop with a trusted operating-system, CI, API, Zapier, or Workato trigger when unattended recurrence is required.

Filesystem and Git checkout
Node package registry and current-version lookup
Chromium and WebKit browser runner
Cloudflare Pages or Workers API
PageSpeed Insights API
Secret manager with least-privilege credentials

Encode the complete maintenance contract

Store the prompt and hard rules beside the site. The agent should inspect instructions, check repository state, resolve the latest compatible dependency set, apply targeted changes, build, test, deploy the exact candidate to staging, and stop before production unless every gate passes.

Never overwrite unrelated human changes
Never force an incompatible dependency tree
Never print or commit secrets
Never promote a changed candidate without repeating the gate
Never call production complete without canonical-host verification

Keep the unattended scope intentionally small

This pattern is intended for static or frontend-only informational sites with deterministic builds and documented deployment paths. Database migrations, commerce, authentication, destructive content changes, legal publishing, and unclear infrastructure require a human decision before execution.

Good fit: brochures, documentation, associations, portfolios
Review first: forms, CMS sync, search, redirects
Human required: databases, payments, auth, destructive actions

Current product references

OpenAI: Scheduled Tasks in ChatGPT OpenAI: Codex automation use cases Anthropic: Local MCP servers on Claude Desktop Anthropic: Desktop and web connectors