Anthropic Ships Cloud Routines for Claude Code

Anthropic on Tuesday introduced Routines for Claude Code, automated tasks that run on the company's web infrastructure without requiring a developer's machine to stay on. The feature, shipping as a research preview, lets users package a prompt, repository, and connectors into a job that fires on a schedule, an API call, or a GitHub webhook, according to Anthropic's announcement. Pro subscribers get five Routines per day, Max gets 15, and Team or Enterprise gets 25, the company's product documentation states.

The launch closes a gap that has bothered power users since Anthropic first added scheduled tasks to Claude Code in March. That earlier release ran jobs on the user's own desktop. Close the laptop lid, kill the cron.

A Routine bundles three things. A prompt describing the task. A repository the prompt operates on. Any external connectors the work needs, like Slack or Asana. Once configured, it fires on a cadence, an inbound HTTP call, or a GitHub event such as a pull request opening. Each Routine gets its own endpoint and authentication token, so deployment hooks and alerting tools can POST a message and receive a session URL back.

Anthropic offered specific examples in its blog post. A nightly job pulls the top bug from Linear, attempts a fix, and opens a draft PR. A webhook listens for new pull requests and runs a team-specific review checklist, posting inline comments. An API endpoint kicks off a smoke test after every deployment. The Decoder reported the feature also handles backlog triage with Slack summaries and SDK porting between languages.

You configure a Routine at claude.ai/code or by typing inside the CLI.

Until March, Claude Code had no native scheduling at all. Boris Cherny, the engineer who built the product, was telling users to wire up cron jobs and manage MCP servers themselves. Then came Desktop Scheduled Tasks, which ran on the local machine. Useful. Tied to whether the lid was open.

Routines moves all of that to Anthropic's cloud. The internal documentation now lays out three options side by side: Cloud Routines, Desktop tasks, and the in-session command. Cloud is the only one that survives a power-off. Desktop is the only one that touches local files. dies when the terminal does, though that's almost beside the point for any serious automation.

The split matters. It also explains why Cherny said in a recent X thread that he runs multiple loops in parallel for code review comments, Slack-driven PRs, and stale-PR cleanup. He told followers to turn workflows into skills and loops, then schedule them.

Five Routines a day for a $20 Pro subscription is not generous. That works out to one Routine every five hours, which will bite anyone trying to use the feature for incident response or always-on monitoring. Extra runs require buying additional usage on top of the base plan.

The timing is also rough. Developers spent the past two weeks accusing Anthropic of quietly degrading Claude Code's reasoning depth to ration compute, with one viral GitHub analysis from an AMD AI director arguing the model had become "unusable for complex engineering tasks," according to VentureBeat. Cherny disputed that conclusion and pointed to product changes, a March 3 default shift to medium effort and a UI tweak that hides the model's full reasoning trace, but conceded the changes were real. Some users feel cornered.

Anthropic told Fortune it will start defaulting Team and Enterprise users to high effort. Pro subscribers do not get that escape hatch. Meanwhile, The Register reported users are burning through quotas faster after Anthropic shortened the prompt cache TTL from one hour to five minutes for many requests. Cherny said larger contexts are now common because users are "pulling in a large number of skills, or running many agents or background automations."

Routines are exactly that kind of background automation. The cache pressure will get worse before it gets better.

Anthropic says webhook support beyond GitHub is on the roadmap. Desktop work is not done either. TestingCatalog reported the company is preparing a major Claude Code desktop overhaul, codenamed Epitaxy, with multi-repo panels and a Coordinator Mode that orchestrates parallel sub-agents.

The move from local to cloud puts Anthropic on the same axis as OpenAI's Codex, which is racing toward its own Magic TODO task system. Both teams are building toward the same idea. Developer workflows that survive logout. The question now is whether Anthropic has enough GPUs to run them.