How to run Claude Code on a schedule

Anthropic documents three built-in ways to repeat work in its scheduling comparison: cloud routines, Desktop scheduled tasks, and /loop inside an open session. It also documents scheduled GitHub Actions workflows. Running claude -p from your own cron is the fourth way, and the most flexible, but Anthropic publishes no recipe for it, so the setup below is ours, built from the documented flags.

The table compares them as documented on October 15, 2026. If the job should report to a team rather than to you, read the last section first: that is where a managed agent platform fits.

Anthropic's headless guide covers print mode. Claude Code exits with code 0 on success and a non-zero code on failure, so a script can branch on the result. With --output-format json, the output includes the result, a session ID, and total_cost_usd, which Anthropic notes is a client-side estimate rather than your bill.

Two flags matter most for scheduled jobs. Set the permission mode explicitly, because a run where nothing sets one takes a built-in starting mode. dontAsk is the mode Anthropic describes for locked-down CI: it denies anything that would prompt, while reads in your working directory and the tools you list still run. And add --bare, which Anthropic recommends for scripts and says will become the default for -p: it skips hooks, plugins, MCP servers, auto memory, and CLAUDE.md, so a run behaves the same on every machine.

Bare mode doesn't use your subscription login. It needs an API key in ANTHROPIC_API_KEY, an apiKeyHelper, or cloud provider credentials. That suits a scheduled job anyway: in -p mode, an API key in the environment is always used and billed per token.

Cron is the most flexible option: any schedule, any machine, any repository, and full access to local files. It is also the one where you own everything. The machine has to stay on, the key has to stay safe, and nobody hears about a failed run unless you build that.

Treat the script below as a starting point. It pulls the latest code, runs one tightly scoped task, refuses to start if the previous run is still going, and writes the JSON result to a dated file. Adapt the prompt and the allowed tools to your job, and give the job only the tools that task needs.

This is a lightly adapted version of the daily report in Anthropic's GitHub Actions guide. To set up, run /install-github-app from Claude Code, which needs admin access to the repository, and store an API key as the ANTHROPIC_API_KEY secret. To use a subscription instead, store a token from claude setup-token and pass it as claude_code_oauth_token.

The summary lands in the workflow run log. To post it somewhere people will read it, grant the tool that does that, such as one that comments on an issue, and add the matching workflow permission. Keep the list short: Claude has no shell or GitHub API access in automation mode until you grant it.

Know GitHub's rules for schedules. The shortest interval is five minutes, schedules use UTC by default, and runs can start late when GitHub is busy. Scheduled workflows run only from the default branch, and in public repositories GitHub disables the schedule after 60 days without repository activity. Each run uses GitHub Actions minutes as well as tokens.

Create a routine with /schedule in the CLI, which requires a claude.ai login. Scheduled routines run at most once an hour and keep working when your laptop is closed. Each run is a full Claude Code cloud session with no permission prompts, so it can use every tool from the connectors you include, writes included.

The limits are about ownership. Anthropic says routines belong to your individual claude.ai account, aren't shared with teammates, count against your usage limits, and act as you in GitHub and connected tools. A green status in the run list means the session started and exited without an infrastructure error, not that the task succeeded, so check the output.

If the job needs local files, use Desktop scheduled tasks instead. They can run every minute with your local tools, but only while the Desktop app is open and the computer is awake, and a task in Manual mode that hits a permission prompt waits for you. For quick polling inside a session, /loop repeats a prompt while the session is open; recurring loops expire after seven days. For work that isn't about code, Claude Cowork has scheduled tasks on paid plans that run remotely.

Start with permissions. Use dontAsk with a short --allowedTools list rather than bypassing permissions. Anthropic's permission modes guide says bypassPermissions belongs only in isolated environments such as containers or VMs without internet access, and that it offers no protection against prompt injection. Claude Code also refuses to start in that mode as root.

Treat what the agent reads as untrusted. Issues, pull requests, web pages, and files can contain text that tries to steer the agent, and Anthropic's security guidance says to avoid piping untrusted content straight to Claude. Before running claude -p in a repository you didn't write, remember that without --bare it runs that project's hooks and connects its MCP servers.

Cap the work. --max-turns ends a run after a set number of agent turns, --max-budget-usd stops at an estimated spend and can overshoot, and a GitHub Actions timeout-minutes ends a stuck job. Keep keys in a secret store or a file with tight permissions, never in the prompt or the repository, and prefer a key created only for this job.

Anthropic's cost guide says Claude Code charges by API token consumption. With a Console API key, usage is billed per token to your organization, and you can set workspace spend limits. On Team and Enterprise plans, usage draws from each member's seat allowance.

Subscriptions work differently. Routines, Desktop tasks, and runs that use a claude setup-token token draw down your plan, and Pro and Max limits are shared across Claude and Claude Code, with a five-hour session limit and a weekly limit. Anthropic says advertised Pro and Max limits assume ordinary, individual usage. A scheduled job that uses up your weekly limit can leave you without Claude when you need it, which is one more reason to give unattended jobs their own API key.

The biggest lever is the size of each run. A narrow prompt, a short tool list, and a turn cap cost less than an open-ended task, and the JSON output's total_cost_usd lets you see what each run cost before the bill arrives.

Claude Code schedules are built for a developer's own work. A routine belongs to one account, a cron job lives on one server, and a GitHub workflow lives in one repository. That is fine for test runs and repository chores. It gets awkward when a sales, support, or operations team needs the result, when the job needs tools like HubSpot or Slack under a team account, or when someone other than its author needs to fix it.

Managed agent platforms take over the scheduler, the runtime, and the history. Developers who are building an agent into a product can use Claude Managed Agents, whose scheduled deployments run sessions on a cron schedule with a time zone, in API beta, at $0.08 per session-hour plus tokens. You still build the app, the users, and the permissions around it.

type.com is built for a team's own recurring work. A scheduled managed agent automation runs a managed Claude or Codex agent with the Space's connections, skills, and memory, on a schedule from every minute to monthly, or on a cron expression in code automations, which are in early access. Each run creates a thread the team can read and continue, a private test shows what a run will do before you enable it, and a schedule pauses after three failed runs in a row, so a broken job stops instead of repeating its mistake. Plans start at $50 a month for two members, including $100 of AI usage at provider rates.