Claude agents can now run on a schedule without a server

Anthropic launched scheduled deployments for Claude Managed Agents on June 9. Set a cron expression, and the agent starts a fresh session at every scheduled moment, completes its task, and shuts down. No server to manage, no scheduler to build, no persistent connection. Think of it as hiring someone who shows up precisely when you need them, does exactly their job, and clocks out, every time, without reminders.

The feature is in public beta and available to Claude for Work and Enterprise subscribers, Anthropic confirmed in its platform blog. Alongside it, Anthropic launched vault environment variables: a secure way to inject API keys and credentials into an agent's session without the model ever seeing the actual secret.

What are scheduled deployments, exactly?

A scheduled deployment gives an agent a cron schedule. Every time the schedule fires, a completely new session starts from scratch. The agent runs its instructions and closes when done. No shared memory between sessions, no accumulated state, no errors that compound over time.

You keep full control. You can pause, resume, or archive a deployment at any time, and trigger additional runs on demand. Starting with a few manual test runs before enabling the schedule is strongly recommended: confirm the output looks right before letting it run unattended.

One thing scheduled deployments cannot do: react to events. An incoming email, an uploaded file, or a CRM webhook does not start a session. This is purely time-based. For event-driven automation, you still need tools like n8n or Make paired with the Claude API directly.

What can you actually automate with this?

The strongest use cases are recurring tasks that currently require manual effort or a custom scheduler. A few concrete examples:

• Nightly data sync. The agent pulls data from three sources at 2 AM, combines it, and writes the result to a shared folder.
• Weekly compliance scan. Every Monday, the agent checks whether your documentation, privacy policy, and data processing agreements are still current.
• Daily customer report. The agent summarizes yesterday's support tickets and posts an overview to your Slack channel.
• Monthly invoice reconciliation. The agent compares incoming invoices against purchase orders and flags discrepancies.

For context: Zapier handles simple if-this-then-that flows well. A Claude agent can reason, analyze unstructured text, and interpret documents. That weekly compliance scan requires reading legal text and making judgment calls. That is not something Zapier can do.

How does the vault work?

The vault is the most underrated part of this update. Automated agents typically need to connect to external services: accounting software, CRM platforms, project management tools. That requires API keys. Until now, those keys had to live somewhere in the prompt or configuration, where the model could read them.

Vault environment variables change that. You store credentials in the platform's secure vault. When a session starts, the keys are injected at the network boundary, outside the model's context entirely. The agent can use them to make authenticated requests, but the model never sees the actual secret. Even if a prompt injection attack manipulates the agent, the key stays protected.

Currently, the vault supports five services: Browserbase, KERNEL, Notion, Ramp, and Sentry. Anthropic is expanding the list; for services not yet supported, OAuth or manual configuration are the fallback options.

How does this compare to the competition?

Anthropic is not alone in scheduling AI agents. GitHub Copilot Cloud Agent launched a comparable feature in early June, aimed at developers: automated code reviews, issue triage, and repository maintenance on a fixed schedule or triggered by repository events.

The key difference: Claude Managed Agents is built for tasks where you need a model that can reason. Zapier is better for straightforward data flows. Copilot targets code. The right choice depends on what you want to automate.

What does this cost per month?

Scheduled deployments carry no separate line item. You pay the standard Claude platform rate: $0.08 per session-hour while the agent is running, plus standard API token costs for the model you use.

A typical daily reporting workflow, ten minutes per weekday, runs roughly 2.9 session-hours per month. At $0.08, that is about $0.23 in session fees. Token costs for Sonnet-class models add another $2 to $11 depending on how much text the agent processes. Total: $3 to $12 per month for a workflow that runs five days a week.

Worth noting: Anthropic paused its Agent SDK billing split on June 15, the very day it was due to take effect, so current subscription pricing applies without the separate credit pool that was announced. See our full coverage for context on where pricing stands.

How do you get started?

If you already have a Claude for Work or Enterprise subscription, you can set up a first scheduled deployment today through the Anthropic platform documentation. Start small: a daily inbox summary, a weekly project status digest, or a monthly uptime check. Run it manually a few times first to confirm the output is right, then enable the schedule.

No Claude for Work subscription? The Claude API paired with an open-source scheduler like n8n gets you the same core capability at lower cost. You miss the vault security and managed hosting, but the agent itself behaves identically.

The broader trend is clear. According to McKinsey's 2026 State of AI report, 72% of companies now use AI in at least one business function, up from 55% in 2024. Recurring, scheduled AI tasks represent the next wave after chat-based tools. You can track adoption rates and deployment patterns in our generative AI statistics page. Scheduled deployments are a concrete step toward AI that runs as part of your operations, not just when you ask it to.