If you've used Claude Code for a while, you've probably hit the same wall I did: Claude is great at a task, but you have to be there to kick it off. Nightly doc audits, PR triage every morning, weekly dependency checks — all useful, all requiring you to open a terminal and type something.
Anthropic opened the research preview for Claude Code Routines on April 14, 2026, and it's the answer to that problem. Routines let Claude run autonomously on a schedule, on API call, or in response to GitHub events — on Anthropic-managed cloud, not on your machine.
What a Routine Actually Is
A routine is a saved configuration with four components:
- Prompt — what you want Claude to do
- Repositories — which repos to give Claude access to
- Connectors — external services Claude can read from or write to (Slack, Linear, Google Drive, GitHub)
- Trigger — what starts the routine
You set it up once. Claude runs it autonomously, with no manual approvals during execution, on Anthropic's infrastructure. Your laptop doesn't need to be on.
Three Trigger Types
1. Schedule
The simplest trigger. Set a cadence — hourly, nightly, or weekly — and Claude runs your prompt at that interval.
Trigger: schedule
Cadence: nightly (02:00 UTC)
Prompt: "Scan all open GitHub issues created in the last 24 hours.
Categorize them by type (bug, feature, question) and post
a summary to #triage in Slack."
This replaces the cron job + shell script + Slack webhook setup most teams have cobbled together for this kind of task.
2. API
Routines expose a unique HTTP endpoint. POST to it — from your CI pipeline, your internal tooling, anywhere — and the routine runs.
# Trigger a routine via API
curl -X POST https://api.claude.ai/code/routines/{routine-id}/trigger \
-H "Authorization: Bearer $CLAUDE_API_KEY" \
-H "Content-Type: application/json" \
-d '{"context": "deploy completed on production"}'
This is useful when you want to tie Claude's judgment to an existing workflow event — deploy finished, test suite passed, ticket status changed — without rebuilding everything inside Claude Code.
3. GitHub Webhook
The most interesting trigger for most developers. Connect a routine to a GitHub repository, and it fires on repository events: pull_request, push, issue, check_run, workflow_run, discussion, release, merge_queue.
Trigger: GitHub webhook
Event: pull_request.opened
Prompt: "Review this PR for security issues, API contract changes,
and missing tests. Post a structured comment on the PR with
findings organized by severity."
This is where Routines gets genuinely powerful. Instead of a static linting check, you get Claude's actual judgment on your PR — running automatically, without a reviewer having to remember to look.
Connectors: What Claude Can Read and Write
Beyond repositories, routines can connect to:
- Slack — post messages, read channel history
- Linear — read and update issues
- Google Drive — read documents, sheets
- GitHub — beyond the triggering repo, Claude can access additional repos and write comments, create issues, update PRs
For most team automation use cases, Slack + GitHub covers the majority of what you need.
Plan Limits
| Plan | Daily Runs |
|---|---|
| Pro | 5 |
| Max | 15 |
| Team / Enterprise | 25 |
Routines require Claude Code on the web to be enabled. You manage them at claude.ai/code/routines or via the CLI with /schedule.
Practical Use Cases
Nightly backlog triage
Trigger: schedule (nightly)
Prompt: "Review all open issues in the repo. For each issue older
than 14 days with no activity, add a 'needs-triage' label
and post a summary to #engineering-triage in Slack."
PR security review
Trigger: GitHub webhook (pull_request.opened)
Prompt: "Review this PR for OWASP Top 10 vulnerabilities, hardcoded
secrets, and unsafe dependencies. Post findings as a PR
review comment with severity ratings."
Deploy verification
Trigger: API (called from CI after deploy)
Prompt: "Check the production health endpoints, scan the latest
deploy logs for error patterns, and post a status summary
to #deployments. Flag anything that needs immediate attention."
Weekly documentation drift
Trigger: schedule (weekly, Monday 09:00)
Prompt: "Compare the API documentation in /docs against the actual
routes in /src/routes. List any endpoints that are documented
but removed, or implemented but undocumented. Open a GitHub
issue with the findings."
What's Different from GitHub Actions or Cron
The honest answer: Routines don't replace GitHub Actions for deterministic tasks. If your automation is "run tests, check coverage, fail if below threshold" — GitHub Actions is better. It's faster, more auditable, and handles that class of problem well.
Routines are better when the task requires judgment, not just execution. "Triage these issues" is hard to encode in a YAML workflow. "Review this PR for security issues" can't be a bash script. That's where Claude's comprehension actually matters, and that's where Routines adds something GitHub Actions can't provide.
The practical question is: would you normally have a person do this task? If yes, Routines is a reasonable fit. If no — if it's pure automation logic — stick with Actions.
Research Preview Limitations
This is a research preview, which means a few things:
- Run limits are per-plan (5–25/day) and may change
- Not all connector types are available in all regions yet
- The GitHub webhook event list may expand; the current supported set is:
pull_request,push,issue,check_run,workflow_run,discussion,release,merge_queue - Audit logs for routine runs are available in the web UI but not yet exportable via API
Setup in 5 Minutes
- Go to
claude.ai/code/routines - Click New Routine
- Write your prompt (be specific about output format)
- Add repositories
- Add connectors (authorize Slack, Linear, etc. if needed)
- Set your trigger
- Save and enable
Or via CLI:
# Start the routine setup wizard
claude /schedule
The CLI wizard walks through the same steps with prompts.
The most interesting thing about Routines isn't the scheduling — it's what it signals about Anthropic's direction for Claude Code. Moving from "terminal tool you invoke" to "cloud automation platform you configure once" is a significant shift. Whether Routines at 5–25 runs/day is enough for serious team workflows depends on your use case, but the foundation is solid.
Worth setting up one routine this week, even a simple nightly summary, to get a feel for how autonomous Claude behaves when you're not watching.
Related:
- Claude Code April 2026 Update: /powerup, MCP 500K, Session Stability (Earlier April updates)
- OpenAI Codex CLI April 2026 Update (Comparing approaches to CLI AI automation)