Anthropic is turning Claude Code into something closer to a 24/7 junior engineer that quietly works in the background, and the new “routines” feature is the clearest step in that direction so far. Instead of just chatting with an AI about code, you can now set up automations that watch your repos, react to alerts, and file PRs while you sleep.

At a basic level, a routine is an automation you define once with three main ingredients: a prompt (what you want Claude to do), a code repo for context, and any connectors like GitHub, Linear, Slack, or Datadog. You then decide how it should run: on a fixed schedule, via a simple API call, or in response to a GitHub event such as a new pull request. The important part is that all of this runs on Anthropic’s cloud infrastructure, not your laptop, so your workflows keep chugging along even if your machine is closed.

Scheduled routines are the easiest on-ramp, and they’ll feel familiar to anyone who has ever set up a cron job or a CI task. You give Claude Code a natural language instruction plus a cadence like hourly, nightly, or weekly, and it executes like clockwork from the cloud. Anthropic’s own examples are very “real dev team” chores: nightly backlog triage that labels and assigns new Linear issues, or weekly scans of merged PRs to catch documentation that went stale and open PRs to fix it. There is no extra server to babysit or cron configuration to remember, and if you have been using the /schedule command in the Claude Code CLI, those tasks now show up as routines under this new system.

The second flavor is API routines, which effectively let you bolt Claude Code onto any system that can hit an HTTP endpoint. Every routine gets its own endpoint and auth token, so a POST request with some payload is enough to spin up a fresh session and get back a URL where you can inspect what the AI is doing. The intended use cases are classic “on-call pain” scenarios: your CD pipeline pings the routine after every deploy so Claude can run smoke checks, scan logs for regressions, and post a structured go/no-go update into a release channel before a human even clicks through. Similarly, teams can point Datadog or Sentry at the endpoint so when an alert fires, Claude pulls traces, correlates them with recent commits, and drafts a fix or at least a high‑signal triage note for the on-call engineer.

The third piece is webhook‑style routines, starting with GitHub integration that turns Claude into a PR-aware agent. You subscribe to a routine for certain repository events, such as PRs that touch specific paths like an /auth-provider module, and Claude spins up one dedicated session per PR. It can then run your own checklist for security, performance, or style, leaving inline comments and summaries before human reviewers arrive, and it keeps following that same PR as comments come in or CI fails. Anthropic is signaling that more event sources will plug into this webhook system over time, but GitHub is clearly the first big one because it sits at the heart of so many engineering workflows.

Early patterns from users show how quickly this can evolve from “cute automation” to meaningful offload of engineering drudge work. Teams are wiring up nightly backlog management that sorts issues, assigns owners based on CODEOWNERS files, and posts digest summaries directly into Slack channels used for standups. Others are leaning into library portability: when a Python SDK PR is merged, a routine automatically ports the change into a Go SDK and opens a matching PR, turning a historically tedious multi‑language maintenance task into something close to one-click. There’s also a growing trend around bespoke code review, where routines run your house rules on every PR, flagging common security or performance pitfalls before senior reviewers even take a look.

If you zoom out, routines are Anthropic’s answer to a problem that has dogged “AI coding agents” from the start: the cost of keeping them running reliably around real production systems. Earlier setups often involved local agents or bespoke scripts that depended on a developer’s laptop being awake, plus brittle glue code to connect different tools. By moving routines into Claude Code’s web infrastructure, giving each one a stable endpoint, and bundling repo and connector access into the core feature, Anthropic is trying to make this feel more like a managed service and less like a side project someone hacked together after hours. It also fits neatly into a broader push toward more “agentic” workflows in the Claude ecosystem, alongside recent moves like the advisor strategy that pairs cheaper executor models with a powerful advisor model for complex decisions.

On the practical side, this is not a free-for-all; Anthropic has put some clear guardrails and quotas around usage. Routines are available in research preview for Claude Code users on Pro, Max, Team, and Enterprise plans who have Claude Code on the web enabled. Pro users can run up to 5 routines per day, Max users get 15, and Team or Enterprise can run up to 25 per day, with the option to pay for extra runs on top of those caps. Behind the scenes, routines draw down the same subscription usage limits as regular interactive sessions, which is a reminder that every automated run still costs tokens and compute.

Of course, the freedom to let an autonomous agent touch your codebase and infrastructure comes with a responsibility to write careful prompts, and the early ecosystem around routines is already emphasizing that. Guides circulating in the community and in the docs recommend being explicit about what “success” looks like, tightly scoping tasks to avoid runaway changes, and clearly stating when a routine should stop. That might mean telling Claude to only touch issues created in the last 24 hours, or to only modify files in a specific directory, or to skip anything that already has both a label and an owner. It’s a very human reminder that as these tools get more powerful, prompt design starts to look less like casual chat and more like writing a spec for a very literal junior engineer who never gets tired.

If you are already living inside the Claude ecosystem, routines are easy to bump into rather than some separate product you have to learn from scratch. You can reach them from the Claude Code web experience at claude.ai/code, from the CLI using the /schedule command, or from the redesigned desktop app that Anthropic shipped around the same time with parallel sessions, integrated terminal, file editing, and faster diffs. Combine that with Claude Cowork’s growing enterprise feature set and the advisor-style agent patterns on the API side, and you can see Anthropic sketching a future where Claude is less of a chatbot and more of a full stack of always-on collaborators woven into the software development lifecycle.