Claude Code remote routines run automations inside customers' clouds to operate on GitHub repos

Claude Code is turning from a terminal helper into something closer to an always-on software worker. That is the real news here. Anthropic’s new routines feature lets you save a coding job, point it at a GitHub repo, and have it run in the cloud on a schedule, from an API call, or off a GitHub event while your laptop is closed. Basically, Claude Code is moving from “help me code right now” to “go handle this class of work whenever it comes up.” (pasqualepillitteri.it) ### What is a routine? A routine is a saved automation package for Claude Code — prompt, repo access, and optional connectors bundled into one reusable job. Instead of starting a fresh interactive session every time, you define the task once and let Anthropic’s cloud rerun it when the trigger fires. The trigger can be a schedule, (pasqualepillitteri.it)per infrastructure. (pasqualepillitteri.it) ### Where does the code actually run? Not on your machine. That is the point. The job runs inside Anthropic-managed infrastructure against a cloned copy of the GitHub repository, then the environment gets torn down afterward. The most concrete spec floating around right now is 4 vCPUs, 16 GB of RAM, and 30 GB of disk per run. Ant(pasqualepillitteri.it)e, even if the hardware numbers showed up first in third-party coverage. (mindstudio.ai) ### Why does that matter? Because interactive coding agents have had an annoying limitation — they usually depend on an open laptop, a live terminal, and a human hanging around to babysit the session. Routines remove that constraint for a certain class of work: repo scans, dependency checks, issue triage, repetitive remediation, tes(mindstudio.ai)t code-related work to happen because a condition was met, not because somebody remembered to keep a tab open. (pasqualepillitteri.it) ### Is this the same as GitHub Actions? Not quite. GitHub Actions is a general automation runner. Claude Code routines are an agent-shaped layer on top of that kind of workflow. You are not just executing a fixed script — you are giving a coding model a standing job with repo context and triggers. Anthropic already pushed Claude (pasqualepillitteri.it)lt pipelines, more persistent delegated work. (anthropic.com) ### What changed versus earlier Claude Code? Earlier Claude Code was mostly about live, supervised sessions in the terminal or IDE. Anthropic had already added checkpoints, subagents, hooks, and background-task support, which made the product more autonomous. But routines go one step further — they survive terminal closures and restarts because the execution lives in the cloud, not (anthropic.com) not a minor convenience feature. (anthropic.com) ### What is the catch? Control and trust. The more you let an agent operate unattended, the more you care about scope, permissions, review boundaries, and failure modes. Anthropic has spent months talking publicly about permission systems, auto mode, and harness design for long-running agents, which tells you the company already sees this as the hard part. The compute box is easy to describe. The governance layer is the real product. (anthropic.com) ### Why is this bigger than one feature? Because it points to where dev tools are heading. Developers are starting to treat coding agents less like autocomplete and more like coworkers with queues. If that model sticks, the important interface stops being the prompt box and starts being the trigger, the repo boundary, and the approval policy. Claude Code routines fit that shift exactly. (allthings.how) ### Bottom line? Anthropic did not just add another Claude Code convenience. It added a way to operationalize the agent — saved, triggered, cloud-run coding work against GitHub repos. That nudges Claude Code into the same mental bucket as CI, bots, and internal automation, which is a much bigger category than “AI pair programmer.” (pasqua([allthings.how)))