Claude Routines: Easy AI Tasks but Capped at 5/Day on Pro
Anthropic's Routines run Claude prompts on schedules, GitHub events, or API calls via cloud infra, but Pro users get only 5 runs/day, making cheaper self-hosted agents like Hermes preferable for heavy use.
Triggers Unlock Reliable AI Automations
Routines execute Claude prompts autonomously on Anthropic's cloud, using three triggers: schedules (recurring intervals, specific dates/times, local/remote), GitHub events (e.g., PR opened), and API POST requests. To build a daily 9:00 a.m. RSS fetcher from JS Weekly, React Status, and Node Weekly, use /schedule in Claude Code CLI with a prompt like: "create a daily 9:00 a.m. trigger that fetches RSS... picks 10 good articles for YouTube videos to send me via Slack." Claude auto-generates a draft prompt, sets timezone/environment, and creates a remote routine viewable in the app. It fetches feeds via bash curl (with custom environment allowing domains) or web fetch tool, selects top 10 articles, and posts to Slack without HR dividers to avoid block errors. For PR reviews, create via desktop app: select repo, GitHub event "PR opened," add custom skills from a repo's .claude folder (e.g., settings.json hooks to copy skills). The routine clones repos, runs the skill, uses GitHub MCP tool for tokenless access, and adds inline comments like "automated review complete, no issues found."
Test runs don't count toward daily limits, but live triggers do—proven by manual re-runs showing 2 executions but only 1 counted.
Setup Demands Prepped Connectors and Workarounds
Pre-install connectors (e.g., Slack) before routine creation, as routines run hands-off without permission prompts. Use custom environments to bypass bash outbound restrictions: switch to custom, add allowed domains (RSS URLs blocked otherwise). Web fetch avoids this via Anthropic's secure infra. For cloud skills/settings, bundle in a GitHub repo's .claude folder—routine clones it, triggering hooks to copy to cloud instance. Prompts need guardrails, e.g., verify skill loading. CLI limits to schedules; desktop enables GitHub/API. Avoid auto-linked repos if unneeded; edit prompts for output compatibility (no HR in Slack).
Pro Limits Make It Impractical for Scale
Research preview requires Pro/Max/Team/Enterprise; counts against subscription token limits like sessions, plus daily caps: 5 routines/24h on Pro, 15 on 20x Max. Each live run (e.g., PR trigger) deducts one, halting after cap despite test-free reruns. No overload on infra, but enforces moderation.
Trade-offs Favor Cheaper Alternatives
Routines simplify setup—one prompt handles fetch/parse/post or review—faster than scratch-building with LLMs. Chain via repo skills for complexity. But high cost deters: prefer webhooks + Hermes agent or Multica on GLM-4.1/GPT coders for volume. Not universal—one-prompt limits nuance without repo hacks. Hints at bigger Anthropic cloud plays (e.g., managed agents), but won't replace N8N/self-hosted for cost-sensitive builders.