Claude Code Officially Launches Routines: Run Mac in the Cloud, Work Uninterrupted Even When Closed

AIbase Report | April 15, 2026

No longer need to manually handle cron jobs and servers while you sleep—let AI fix bugs, review PRs, and clear backlogs for you.

Yesterday (April 14), Anthropic launched a new feature in Claude Code—Routines, currently available as a research preview. In one sentence: package prompts, code repositories, and connectors into a single bundle, save it, and let Claude run automatically based on a schedule, API calls, or GitHub events, all executed on Anthropic's cloud, allowing your local computer to be turned off directly.

What exactly is it?

Previously, developers who wanted Claude Code to work on a schedule had to set up their own system: write cron jobs, rent servers, configure MCP server, and link everything together. Anthropic admitted in their official blog that this was a high barrier that discouraged many people. Developers were already using Claude Code to automate software development processes, but before this, they needed to manage cron tasks, infrastructure, and additional tools like MCP servers themselves.

Routines packages all of this. Configure once, and you get a "Claude session template with repository permissions and connectors," which knows where to go, which tools to use, and what the goal is.

Three triggering methods cover most common automation scenarios:

Run by schedule. Hourly, nightly, weekly, and even custom cron jobs in CLI (minimum interval of 1 hour). The official example is straightforward: pull the highest-priority bug from Linear at 2 AM every night, try to fix it, and open a draft PR. You can just review it when you come to work the next day.

Triggered by API call. Each Routine has its own endpoint and authentication token. Send a message via POST, and get a session URL in return. Datadog alerts, CI/CD hooks, internal dashboards, or any place that can send an HTTP request can directly integrate Claude.

Triggered by GitHub events. After installing the Claude GitHub App, events such as pull_request, push, issue, and check run can all act as triggers, and can be filtered by author, tag, branch, or draft status. A practical use is to monitor a specific sensitive directory—flag all PRs involving the /auth-provider module. Any changes to this module should be summarized and posted to #auth-changes. Each matching PR gets a dedicated session, and it will continue to follow up on subsequent comments and CI failures.

Who Can Use It, How Often?

The research preview is only available for Pro, Max, Team, and Enterprise subscriptions, with a hard limit on daily runs: 5 per day for Pro, 15 for Max, and 25 for Team and Enterprise. Additional usage beyond the limit will incur extra charges. All Routine runs share the same subscription quota as regular sessions.

GitHub-triggered routines also have hourly rate limits; events exceeding the window will be discarded directly. So if your repository is very active, it’s best to set up filters in advance to avoid wasting runs on irrelevant PRs.

One default security setting is worth noting: By default, Claude can only push to branches with the claude/ prefix and does not directly modify main. This is a prudent safeguard; unless your downstream review process is well-established, it’s recommended to keep it enabled.

The entry point is claude.ai/code/routines, or you can type /schedule in the Claude Code CLI.

What Early Users Are Doing With It

Anthropic's official blog and early user feedback have already demonstrated several established patterns:

• Nighttime backlog cleanup: Schedule to pull new issues, tag them, assign responsibilities, and push a summary to Slack in the morning so the team can view the queue right away.
• Alert triage: Monitor the API endpoint of the Routine, and let Claude pull stack traces and associate recent commits to pre-open a candidate fix draft PR, so the on-call person can see something ready when they open the page.
• Customized code reviews: Trigger when a PR opens, leave inline comments according to the team's checklist, and perform mechanical checks for security, performance, and style first, letting human reviewers focus on design.
• Deployment verification: Run a Routine after CD pipeline deployment to run smoke tests, scan logs for regressions, and send the go/no-go decision directly to the release channel.
• Documentation drift correction: Scan recently merged PRs once a week to find documentation referencing changed APIs and automatically open update PRs.
• Cross-language SDK synchronization: When a PR is merged in Python SDK, the Routine automatically applies the change to Go SDK and opens a corresponding PR, keeping both libraries aligned.

The last category especially illustrates the problem—teams that have used multiple SDKs know how time-consuming this kind of synchronization can be.

Community Reaction

After the official tweet, the developer community reacted consistently: Finally, no need to keep Macs running all night. The discussion thread quickly reached the front page on Hacker News, with several key issues emerging: how to control token consumption, ensure idempotency (whether repeated event triggers cause chaos), and the positioning differences with multi-agent frameworks like OpenClaw.

Some developers have already scheduled their first set of Routines: one for PR reviews, one for issue triage, and one for automatic smoke testing after deployment. Some non-engineering users said this was the first time they felt that "letting AI handle daily chores" was truly feasible.

Some Observations

Routines itself doesn't bring any groundbreaking new model capabilities, but it fills another gap: transforming Claude Code from a tool that requires opening a terminal into a continuous online background worker. This aligns with Anthropic's recent actions (redesign of the Claude Code desktop app, parallel session sidebar, Cowork GA): pushing from real-time collaboration towards scheduling and orchestration.

Of course, there are uncertainties typical of the research preview phase. The /fire endpoint is published under the experimental-cc-routine-2026-04-01 beta header, and request and response formats, rate limits, and token semantics may change during the research preview. The official also mentioned that webhook trigger sources will expand from GitHub to more event sources in the future, but without a timeline.

For teams already using Claude Code, this feature is worth trying today. For those still watching, at least it brings the idea of "AI as a continuously running team member" much closer to reality.

Source: Anthropic's official blog, Claude Code documentation (code.claude.com/docs/en/routines), @claudeai tweets, and reports from The Decoder, 9to5Mac, and other media. Details are subject to the latest official documentation.