Claude Code Finally Gave Us Scheduled Automations
After months of running automations tethered to open laptops, Claude Code has shipped web-based routines that run on Anthropic's infrastructure — but the migration isn't plug-and-play. Early adopters are discovering that API keys don't transfer automatically, network permissions default to restrictive settings, and stateless execution breaks cookie-dependent workflows. The promise is clear: set-and-forget automations that survive restarts and run hourly without local hardware. The question is whether the new constraints — GitHub-only file access, 15-run daily caps, and one-shot prompt design — make these routines genuinely more powerful than the desktop tasks they replace.
Key Takeaways
Routines run on Anthropic's cloud infrastructure and persist without an open laptop, but require a linked GitHub repository and store API keys in cloud environment variables rather than local .env files.
Network access defaults to «trusted» mode (vetted domains only); many integrations like ClickUp require switching to «full» access, which introduces theoretical risk of data exfiltration if malicious content is read during a run.
Each routine execution is stateless — the GitHub clone is destroyed after each run — so workflows relying on local cookies, session persistence, or browser automation will break unless redesigned around API endpoints.
Daily run limits are tight: 5 on Pro, 15 on Max, 25 on Team/Enterprise, with a minimum interval of one hour, making high-frequency or resource-intensive automations impractical.
Prompts must be specific enough for one-shot execution; vague instructions that work in interactive chat will fail in autonomous routines, requiring upfront testing with the «run now» feature.
In a Nutshell
Claude Code's web-based routines deliver the first truly autonomous scheduling for AI agents, but success depends on mastering environment variables, network permissions, and stateless execution — desktop automations won't migrate without deliberate redesign.
What Are Routines and How Do They Work?
Routines are scheduled, API-triggered, or event-driven prompts running on Anthropic's cloud.
Claude Code routines represent a fundamental shift from desktop-bound automations to cloud-native execution. A routine is essentially a pre-configured prompt that can fire on a schedule, via API call, or in response to GitHub events like new pull requests or pushes. Unlike desktop scheduled tasks, routines run on Anthropic's web infrastructure, meaning your laptop can stay closed and the automation will still execute.
Each routine run clones your linked GitHub repository, reads your claw.md context file, executes the prompt, and then destroys the ephemeral environment. If the automation modifies code or performs a review, Claude creates a new branch in your repo; otherwise, the session history is saved but the runtime is discarded. This stateless architecture is powerful but introduces constraints: local files outside the repo are invisible, cookies don't persist, and every run starts fresh.
You can create routines from three interfaces: the desktop app's scheduled tasks panel, the web interface at claudeai/code, or directly in the terminal. All three methods require a GitHub repository as the source of truth and support the same set of triggers, connectors, and cloud environment configurations.
The API Key Problem: Environment Variables vs .env Files
Secrets in .gitignore won't transfer; use cloud environment variables instead.
The API Key Problem: Environment Variables vs .env Files
The most common migration failure is forgetting that .env files are git-ignored and won't be cloned into the cloud runtime. API keys must be manually entered into the cloud environment's «Environment Variables» section in settings. Even if your local automation reads from .env perfectly, the routine will fail silently unless you explicitly configure environment variables and — crucially — tell Claude in the prompt to «use the API key directly from the environment» rather than looking for a .env file.
Setting Up Your First Routine
Five configuration steps from prompt to permissions.
Name and Prompt Create a specific, one-shot prompt that doesn't require user interaction. This is the instruction Claude will execute autonomously.
Link GitHub Repository Connect the repo containing your claw.md, scripts, and skills. This becomes the file system Claude can access during the run.
Configure Cloud Environment Add all necessary API keys to the «Environment Variables» section. Set network access to trusted (vetted domains) or full (unrestricted outbound).
Set Schedule or Trigger Choose hourly, daily, or weekday cadence (minimum one hour), or configure API/GitHub event triggers instead of a schedule.
Add Connectors and Permissions Enable MCP services like Slack or Gmail if needed. Set permission level to fully autonomous to avoid blocking prompts.
Network Access: Trusted vs Full
Default trusted mode blocks many APIs; full mode works but increases risk.
What Breaks When You Migrate
Routines vs Desktop Tasks vs /loop
Three automation modes compared across seven key dimensions.
Usage Limits and Resource Constraints
Daily run caps and per-routine compute allocations vary by plan.
Why This Beats Traditional Automation
Agentic reasoning survives the transition to cloud execution.
“We are actually keeping the agentic framework. When we push those automations to the cloud and it's just a Python script, we're losing the agentic piece. We're only sending off really the tools and the workflow. But in this case, we're keeping the WA and the T all running together because the agent is looking at the claw.md, it's looking at its scripts and it's figuring out what to do. And if it runs into errors midrun, it will self-correct.”
Best Practices for Writing Routine Prompts
People
Glossary
Disclaimer: This is an AI-generated summary of a YouTube video for educational and reference purposes. It does not constitute investment, financial, or legal advice. Always verify information with original sources before making any decisions. TubeReads is not affiliated with the content creator.