IMPLICATOR
Anthropic has shipped native scheduled task support for Claude Code, letting the AI coding agent run prompts on recurring cron schedules without anyone at the keyboard. The feature, available now in the Claude Desktop app, allows users to configure tasks that fire hourly, daily, or weekly, with Claude Code operating against any project files and MCP tools in its environment. Only desktop app users can access the capability for now. Terminal and IDE extensions don't support it yet.
The Breakdown
- Claude Code now runs scheduled tasks via the desktop app, firing hourly, daily, or weekly without human input
- Tasks are agentic and self-correcting, unlike deterministic cron jobs that stop on errors
- Each run is stateless by default, but a single state file per task provides continuity between executions
- Desktop app must stay open for cron to fire; terminal and IDE support not yet available
How scheduled tasks work
Setup takes about a minute. Users open the schedule tab in the desktop app, name the task, write a prompt, select a model, choose a project folder, and set a frequency. Typing /schedule inside any existing Claude Code session works too.
Once a cron fires, Claude Code opens a fresh session, reads the prompt, and works through the project on its own. It reads files, calls tools, edits code, executes scripts. Same capabilities as an interactive session, minus the human watching.
The difference from a standard shell script matters here. Claude Code tasks are agentic, not deterministic. A bash cron job hits an error and dies. Claude Code hits an error, tries alternative approaches, evaluates which one worked, and updates its own logic to avoid the same failure on the next run. That self-correcting loop turns what would be a brittle automation into something closer to a junior developer left alone with a checklist.
Stateless by design, with a workaround
Every scheduled run starts clean. No memory of previous executions carries forward. For a daily standup summary or a weekly dependency audit, that works fine. The agent reads current project state and reports.
But some workflows need continuity. The recommended pattern: maintain a single state file per task. The prompt tells Claude Code to read the file before starting and overwrite it when finished, logging what happened, what broke, and what the next run should know. Crude. Functional.
A growing append log scales poorly. Run a task a thousand times, and the next agent has to parse a thousand entries before it can start working. One overwritten file keeps context lean enough to be useful.
Stay ahead of the curve
Strategic AI news from San Francisco. No hype, no "AI will change everything" throat clearing. Just what moved, who won, and why it matters. Daily at 6am PST.
No spam. Unsubscribe anytime.
The laptop problem
Scheduled tasks require the desktop app running on your machine. Close your laptop and the cron stops.
Anthropic built in a catch-up mechanism. When the app reopens, it scans back seven days for missed executions and runs them. That doesn't save time-sensitive workflows. A morning briefing that fires at 3pm because you finally opened your lid defeats the purpose.
All scheduling logic and cron metadata live inside the desktop app. Terminal Claude Code can see task configuration files on disk, since they exist as regular files, and can even edit them. Creating new tasks or triggering the scheduler from terminal is not possible. Anthropic locked that logic inside the desktop app.
Keep the desktop app open in the background. Work in VS Code or terminal the way you normally would.
Permissions deserve scrutiny
An unsupervised agent with write access to your codebase should make any developer nervous. Scheduled tasks inherit whatever permissions the user has configured, and default settings were not designed for autonomous operation.
Lock down destructive bash commands before enabling any recurring task. Restrict file operations to specific directories. Review MCP tool access. Claude Code can help configure these guardrails through its local project settings, but the responsibility sits with the user, not the agent.
Notifications need manual setup too. The desktop app organizes completed runs in a task view, but built-in alerts are easy to miss. Adding a hook that plays a sound or fires a message on session completion gives scheduled tasks the visibility they need before you trust them running unsupervised.
What actually changes
Scheduled tasks don't expand Claude Code's toolbox. The skills and MCP integrations were already there in interactive sessions. What changes is the trigger. A clock instead of a human typing a prompt.
The agent's abilities haven't changed. But removing the human from the loop turns Claude Code from an assistant you summon into infrastructure that runs on its own. Morning briefings and code reviews become background processes. Each failed run leaves clues. The next agent reads them and adjusts.
People keep saying "AI employee." Anthropic just gave the phrase a scheduler. Your laptop has to stay awake.
Frequently Asked Questions
Can I run scheduled tasks from the terminal or VS Code?
Not yet. Scheduled tasks only work in the Claude Desktop app. Terminal Claude Code can see and edit task configuration files on disk, but it cannot create new tasks or trigger the scheduler. Anthropic is expected to add terminal and IDE support in a future update.
What happens if my computer is off when a task is supposed to run?
The task won't fire. Anthropic built a catch-up mechanism that scans back seven days for missed executions when the desktop app reopens. Missed tasks run at that point. Time-sensitive workflows like morning briefings lose their value if they execute hours late.
Do scheduled tasks remember what happened in previous runs?
No. Every run starts a fresh session with no memory of prior executions. The workaround is a single state file per task. The prompt instructs Claude Code to read the file before starting and overwrite it when finished, creating a lightweight memory layer between runs.
How do I prevent a scheduled task from making destructive changes?
Configure Claude Code's local project settings to deny dangerous bash commands and restrict file operations to specific directories. Review MCP tool access before enabling any recurring task. Run each new task manually first to confirm it completes without requesting additional permissions.
How do I get notified when a scheduled task finishes?
The desktop app shows completed runs in a task view, but its built-in notifications are easy to miss. Set up a hook that plays a sound or sends a message through a chat tool every time a Claude Code session ends. Ask Claude Code to configure this through the hooks system.
Marcus Schuler
Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: [email protected]
