Claude Code Plugins Worth Scheduling — And Which Aren't

Every "best Claude Code plugins" list you have read was written for someone sitting at the keyboard. That is a different filter than the one you need for Claude Code plugins worth scheduling. A plugin that is brilliant when you are there to answer its questions can be worthless at 2am, because at 2am nobody answers. The plugin sits at a prompt until the timeout kills it, and you wake up to an empty log.

I run agents unattended on my own Mac most nights — code passes, blog drafts, social discovery, video renders. Here is the test I use to decide what earns a slot, which plugins and skills pass it, and the short list of things that should never run while you are asleep.

What makes a plugin schedulable

Being useful is not the same as being schedulable. Three questions sort them, and a plugin has to pass all three.

1. Does it end on its own?

An unattended run needs a terminal state. The agent has to reach a point where it says "done" and stops, without a human deciding what "done" means. Interactive plugins that open a menu, ask which file you meant, or wait for you to pick an option will hang. Run it once with your hands off the keyboard: if it ever pauses for you, it fails.

This is why Anthropic's own guidance points at verification loops packaged as skills — the check runs every time, automatically, instead of relying on someone remembering to run it.

2. Does it leave something behind?

A scheduled run you cannot inspect in the morning did not happen. The output has to be durable and outside the session: a commit on a branch, a markdown file, a database row, a draft in your CMS. "It printed a nice summary to the terminal" is not an artifact. The best scheduled agents write their findings to a file with a predictable name — ARCHITECTURE-REVIEW.md, SECURITY-REVIEW.md — so tomorrow's run can read yesterday's.

3. Is it safe to run twice?

Schedules fire whether or not there was anything to do. A good scheduled agent is idempotent: run it on an unchanged repo and it should do nothing rather than invent work. Agents that always produce something because they were asked to will pad. An agent told to "improve the code" every night at 2am will eventually rewrite things that were fine.

Claude Code plugins and skills worth scheduling

These are the categories that pass all three questions. Note how few of them are about writing code.

The examples come from PlugMyPlugin, a directory of Claude Code plugins I built. Every listing gets a capability card parsed from the plugin's own files rather than guessed.

Repo maintenance you would never do by hand

This is the sweet spot: valuable, boring, and nobody ever gets to it. Dead code removal, architecture drift review, dependency hygiene. A one-paragraph prompt is usually enough — "find dead code in this repo, unused functions, unreachable branches, stale feature flags, commented-out blocks, delete it, then run the tests." It ends, it commits, and it does nothing on a repo that has not changed.

Off the shelf, wshobson's codebase-cleanup bundles tech-debt reduction, dependency updates and refactoring into three commands and two agents with no external services. Pair it with dependency-management for the outdated-packages and vulnerability pass. And update-claudemd may be the purest scheduled agent in the directory: it rewrites CLAUDE.md to match recent code changes, a job that only makes sense behind the "only if the repo changed" gate below.

Review and security passes

Code review and security audit plugins are built to produce a verdict, which makes them natural scheduled agents. Point one at the last day of commits, have it fix what is trivially safe and write the rest to a file. You read the file with coffee instead of reading the diff.

Anthropic's own code-review plugin runs several specialised review agents in parallel, filters findings by confidence, and can write the result locally instead of posting to GitHub. The local mode is the one you schedule. For the security pass, netresearch's security-audit scans against OWASP Top 10 and CWE Top 25 with CVSS scoring and needs nothing but python3. Both produce a verdict and stop.

Test coverage, aimed narrowly

"Add tests" is a bad nightly prompt — you get coverage padding. "Find the most important untested code paths and add focused tests for error handling and edge cases" is a good one, because it has a natural stopping point and the output is reviewable as a normal PR. wshobson's tdd-workflows has the agents for this; give them the narrow prompt, not the whole methodology.

Content agents that end in a draft

This is the category most people miss. An agent that researches a topic, writes an article and creates it as a draft is perfectly schedulable, because the draft is the terminal state. It publishes nothing. You review it in the morning and flip the good ones live. Same shape for changelogs, release notes and release summaries.

This is the one I built for myself. SEO Blog Writer is a free Claude Code plugin: /seoblog:write picks a topic from your keyword queue or a gap in your niche, runs live SERP and competitor research, writes the article, generates hero, inline and OG images through fal.ai, looks at every image with vision to catch the AI tells, then creates a draft on WordPress, Shopify or an Articles API. It asks zero questions. I removed the outline-approval, image-spend and publish gates specifically so it could run on a schedule, and draft-first is hardcoded, not a convention. Image spend is fixed by article length (a medium article is three images at $0.15 each, plus at most two fix rounds per image) and lands in the run report. The post you are reading started as one of its drafts. Install with /plugin marketplace add mrieck/claude-plugins then /plugin install seoblog@productive-mark.

The one-off you set for tonight

Not everything worth scheduling repeats. "Render the demo video," "do the design pass on the landing page," "summarise this week's commits into a changelog" are single fires you set before bed. A once-only agent that disables itself after it runs is the tidiest version of this.

My one-off of choice is DemoDay. /demoday:create-demo-video reads the codebase, writes a script, drives the app on screen, narrates it with AI voiceover and renders a finished MP4 with captions through Remotion, plus a vertical Short if you want one. Set it for tonight after a release and wake up to the video.

What to keep off the schedule

The list of things that should never run unattended:

• Anything that spends money without a ceiling. No purchases, no metered API calls you have not capped, no infrastructure that scales on its own.

• Anything needing 2FA or an account signup. There is no one there to read the code out of your inbox, Google login is fine, but creating accounts could get you banned.

• Permanent deletes and force pushes. A branch is recoverable; a force-push at 4am is a bad morning.

• Open-ended prompts. "Make the product better" has no terminal state. It will find work until the timeout.

• Anything interactive by design. If the plugin's whole value is the back-and-forth, keep it at the keyboard.

The pattern behind all five: an unattended agent should be allowed to propose anything and commit to very little.

Two plugins in the directory enforce that pattern for you. wshobson's review-agent-governance gates PR comments, merges and CI workflow edits behind an explicit human-approval flag, so an unattended run cannot merge at 4am even if it wants to. netresearch's github-release intercepts gh release create, delete and edit and routes releases through a PR instead. Install either and the blast radius shrinks to what you can recover from.

Where these actually run

Overboard for Mac turns your Mac into a 24/7 agent orchestrator with no cloud in the path: slots fire via launchd, with a global queue capped at 1–8 concurrent runs, a per-slot timeout, an "only when the repo changed" switch, and a usage cap so a maintenance pass does not eat your plan.

The part that matters most for trust: runs are not hidden. Each one launches as a real Claude Code session inside Herdr, so you can watch a 3am run live or take it over mid-flight. There is no API key and no token bill — scheduled runs use the Claude Max or Pro plan you already pay for. It costs $49 once with a 14-day trial.

If you are still choosing between a scheduler on your Mac, Claude's desktop scheduled tasks and cloud routines, I compared all of them in Claude Code scheduled tasks: cloud, desktop, or your Mac?

The short version

Pick plugins that finish on their own, leave a file behind and do nothing when there is nothing to do. Combine it with a scheduler like one in Overboard, and you can token-max your subscription and accomplish more than you ever dreamed.