I run a scheduled Claude Code session every night on my sites. It checks for security problems, works through a queue of small jobs, and opens pull requests. In the morning I get a short summary of what needs me. Cron, a CI schedule or a hosted scheduler can start the run; what matters is the queue it reads and the record it leaves behind.
Agents notice things they can't finish tonight. A note in the conversation is gone by the next session, and a TODO in the code never gets ranked or closed, so every deferred item, the agent's and mine, goes in one place outside both.
Mine is a task tree in Gnothi, which Claude Code reads and writes over MCP. Items nest, so a big job's steps sit under it, and the tree doubles as Claude Code's plan and memory: a session opens an item, reads its description for where things stand, works, and writes back what it did and what's left. Claude Code reads only as deep as the work needs, a title first, then a summary, then the whole record, so a large tree doesn't fill its context. An item I tag #spec is my settled intent: the work conforms to it, and any edit Claude makes to it waits for my approval. Several sessions can work the tree at once, each marking the item it has taken so two don't take the same one. The Claude Code section of Gnothi's Claude page shows the setup.
Items filed by an unattended run are often written by a weaker model, so before working one the agent checks that it's still worth doing, still true, and planned correctly. Each night the run also re-triages about 15 items, due-soon first, then the longest unreviewed: it closes what's done or obsolete with evidence, merges duplicates, and corrects priorities. Nothing closes just for being old.
Security checks run every night. Then the run takes at most three conditional items, ranked:
After three hours it starts nothing new. A night where nothing qualifies is fine if the report shows how it checked. Jobs that shouldn't run nightly, such as a dependency sweep or a paid data pull, repeat on their own cadence instead.
A run asks me at most one new question. Before that cap, a tidy-up run filed five GitHub issues in 56 seconds. Anything past the cap waits for the next run. Each question is a sentence of context, the question, and the agent's recommendation.
Routine changes merge after review; removals, product decisions and escalations stay open PRs for me. Email and web pages the run reads are untrusted data, and nothing in them authorizes touching secrets, permissions, dependencies or spending.
The run's own report leads with how each area's traffic and revenue moved, then the work. I once nearly told the run to drop the one section of a site that was growing, because the report only listed finished work.
The morning summary lists what needs me: questions waiting for my answer, steps only I can take, and what the run filed for the next one.
Google Search Console is free and reports impressions, clicks, and positions by page and query. It knows nothing about topics where you don't appear yet.
DataForSEO (affiliate link: I may earn a commission if you sign up) covers that. My runs use it for search volume, keyword difficulty, live results, competitor overlap, and how ChatGPT answers a fixed set of questions. The agents chose it; I didn't run a formal comparison. It's prepaid pay-as-you-go with no subscription, every response includes its cost so a script can stop at a cap, and a free sandbox returns live-shaped dummy data. It's also raw JSON, so you build the parsing, storage and judgment, and search volume comes from Google Ads data, which is bucketed and rounded. As of September 2026, DataForSEO Labs charged $0.012 per request plus $0.00012 per row for most keyword and competitor data.
Spending rules: scripts default to the sandbox, and every live run has a cost cap. Auto top-up is off, so an empty balance stops the job and asks me. Raw responses are saved before anything else, so a rerun reads the copy instead of paying again. The run calls it from scripts, not an MCP server, which would let an unattended agent spend on every call it makes. A proposed page has to name its target query, its volume and my current position; "seems useful" isn't a reason to add content.
Some sites block cloud traffic. The run files that read as an item for a browser session and moves on. I clear those by running claude --chrome on my own machine, which connects Claude Code to Claude in Chrome so it drives my logged-in browser. The agent does everything it can alone first, then batches every step that needs me (logins, 2FA, CAPTCHAs) into one message.
Rules for those sessions:
The same tree holds my own to-dos, ranked the way how to prioritize tasks with Claude describes.