Anthropic launched Claude Code routines on April 14 2026 as a research preview. You define a prompt once along with repositories and connectors. The system then executes that prompt on their servers according to your chosen schedule API call or GitHub event. The laptop can stay shut.
This setup directly tackles the overhead that has limited agentic coding deployments. Earlier approaches demanded constant local monitoring or brittle custom scripts. Routines move everything to Anthropic infrastructure and assign each one a unique API endpoint. That shift lets teams route real operational signals straight to a coding model without extra layers.
Trigger Types and How They Combine
Scheduled routines operate on hourly nightly or weekly cycles. One example pulls the top bug from Linear at 2am generates a fix and submits a draft PR. Previous CLI schedule commands now convert to these routines automatically with no manual migration steps.
API based routines expose an individual endpoint. Teams post an alert payload to it. The model determines the responsible service assembles a triage report and sends it to the oncall Slack channel. The response returns a session URL so reviewers can inspect the full reasoning trace.
GitHub webhook routines activate on pull requests commits or new issues. A routine can scan for modifications to an auth provider folder then publish a summary to a dedicated channel that covers security impact and test recommendations. Anthropic plans to expand event sources. One routine can listen for multiple trigger kinds at once.
Every execution begins with a clean session. The prompt must contain all necessary instructions and context. Generated PRs and session logs need close human examination before merge. That discipline stays essential even as the infrastructure burden decreases.
Plan Limits Shown Visually
The chart above displays the daily caps. Pro accounts get five executions. Max users reach fifteen. Team and Enterprise plans receive twenty five. Additional runs become available through usage based billing. Because this remains a preview those figures could change based on observed usage patterns.
Five daily runs support a few critical scheduled jobs. Fifteen opens space for mixed scheduled and reactive automations. Twenty five fits larger groups that direct several alert streams through the model. The caps push teams to identify which repetitive tasks produce the clearest time savings.
Setup Options on Web and CLI
Through the browser head to claude.ai/code/routines then select new routine. Provide a clear name. Craft a fully self contained prompt. Choose the model. Link the relevant codebases and connectors. Creation makes the routine live and callable immediately.
From the command line the slash schedule command accepts a description and registers the routine. The same command lists current items or handles modifications. Any prior schedule tasks transferred over without user effort. That continuity matters for developers already embedded in the Claude Code workflow.
Writing Prompts That Stand Alone
Because each execution is stateless the prompt carries the entire job specification. Include explicit criteria for success failure modes what information to extract and how to format outputs. Test prompts in normal Claude Code sessions first to verify they produce reliable results before converting to a routine. This step saves time. For coding tasks reference specific style guides or test requirements inside the prompt. The better the prompt the less review work per run. Vague language leads to outputs that demand heavy editing. Specific instructions produce outputs closer to what an experienced engineer would submit.
Connecting to Current Coding Model Performance
This feature gains strength when paired with strong coding models. Recent benchmarks show certain models excel at real repository tasks. Using a top performer for routines that generate PRs reduces the volume of fixes needed during review. For example a routine that attempts Linear bug fixes benefits from a model that scores well on relevant benchmarks. The output quality directly impacts how much time you save. Lower tier plans restricted to weaker models will see more manual correction. That reality makes the plan limits more than just rate limits. They shape what you can realistically automate. See my post on Claude Mythos Preview for details on its coding strengths that complement these routines well at https://adam.holter.com/claude-mythos-preview-first-ai-to-complete-aisi-32-step-cyber-range-end-to-end/.
Expanded Use Cases
Beyond the official examples consider nightly code quality scans that post summaries to Slack. Or automated labeling of incoming GitHub issues based on content analysis. Deploy hooks can trigger post deployment verification routines via the API endpoint. The webhook support for PRs allows automatic security reviews on changes to sensitive modules. These patterns turn one off analyses into persistent background processes. The cloud execution means they continue even when your machine is off. Teams gain consistency without assigning extra headcount to watch queues overnight.
Preview Considerations and Best Practices
Treat this research preview as an experiment. Monitor output consistency across runs. Prompt behavior may shift with backend updates. Start small with one scheduled routine and one webhook. Measure the time saved against review effort. Adjust prompts until the triage or fix quality meets your standards. Combine triggers thoughtfully. A routine that activates on both schedule and specific GitHub events can cover more ground without exceeding daily caps. Documentation at code.claude.com/docs/en/routines covers prompt engineering trigger setup and current billing details in depth.
I have followed the back and forth in coding tools for months. Features like this one stand out because they address concrete operational problems instead of promising vague future gains. The dedicated endpoints particularly reduce glue code between your systems and the model. Cloud execution without local dependencies changes the calculus for always available agents. It is not a complete autonomous engineer but it does handle the repetitive slices of work that used to require constant oversight. For teams measuring their coding model performance on benchmarks like SWE bench the routines feature lets those models act on a cadence or in response to events. The numbers become more meaningful when model outputs flow directly into your repositories and communication channels.
Implementation requires little time. A first routine focused on one nightly task or one webhook for a critical path can go live inside an hour. Begin narrow. Monitor the generated sessions and PRs closely. Refine the prompt language based on what you observe. That iteration cycle mirrors the one used for model selection itself. As the outputs stabilize the routines become reliable extensions of the team instead of experiments.
This addition does not transform software development overnight. It does eliminate a category of overhead that sat between the model and the work. Earlier approaches demanded constant local monitoring or brittle custom scripts. Routines move everything to Anthropic infrastructure and assign each one a unique API endpoint. That shift lets teams route real operational signals straight to a coding model without extra layers.