AI agents · OpenClaw · self-hosting · automation

Quick Answer

What is Claude Code Routines? (Anthropic, May 2026)

Published:

What is Claude Code Routines? (Anthropic, May 2026)

Claude Code Routines are scheduled, headless Claude Code agents that run on Anthropic’s cloud. Triggered by cron, API, or GitHub events. Announced at Code with Claude 2026 — here’s what they are and how they fit in May 2026.

Last verified: May 18, 2026

The one-paragraph answer

Claude Code Routines are saved Claude Code automations that run on Anthropic’s cloud infrastructure, triggered by a schedule, API call, or GitHub event. Announced as a Research Preview at Code with Claude 2026 (San Francisco, May 6, 2026), Routines are the “autonomous Claude Code” tier — instead of a developer running Claude Code interactively in a terminal, you save a Claude Code workflow once and have it execute on a schedule (e.g., every weekday at 9am) or in response to a GitHub event (e.g., when a PR opens, when a dependency advisory fires). They use the same Claude Opus 4.7 model, MCP connectors, and skill ecosystem as interactive Claude Code, but run headlessly on Anthropic’s compute so they keep going when your laptop is closed.

What Routines actually do

Trigger (cron / API / GitHub event)
        │
        ▼
Anthropic Cloud (Claude Code Routine runner)
        │
        ├─► Claude Opus 4.7 model
        ├─► Your MCP connectors
        ├─► Your skills / project files (git checkout)
        └─► Tool sandbox (read/write/execute as scoped)
        │
        ▼
Output: PR opened / Slack message / Linear ticket / report

A Routine is just a saved Claude Code prompt + config + trigger. You author it once, register a trigger, and Claude Code runs it for you in the cloud.

Triggers

TriggerUse case
Cron schedule”Every weekday at 9am, summarize new GitHub issues”
API callWebhook → Routine. “Slack /standup command runs the standup-digest Routine”
GitHub push”On every push to main, run security review”
GitHub PR opened”Auto-review every new PR with the team’s style guide”
GitHub issue created”Label new issues, suggest reproducer, ping owner”
GitHub release”Generate marketing copy on every release tag”

What you can build

  • Daily dependency upgrade Routine — Claude Code picks up Dependabot alerts, runs upgrades, runs tests, opens a PR.
  • Nightly security scan — runs npm audit + Claude Code analysis, posts a Slack report.
  • Weekly cost optimization audit — Claude Code reads your Cloudflare, AWS, OpenAI usage, ranks expensive items, proposes fixes.
  • PR auto-review — every PR gets a Claude Code review with the team’s style guide and security checklist applied.
  • Documentation refresh — on every release tag, Claude Code regenerates the changelog and reviews docs for staleness.
  • Test-flake triage — every morning, Claude Code reads last night’s CI failures and groups them by likely root cause.
  • Monday status digest — Claude Code reads Linear/Notion/GitHub and emails a “what shipped, what’s blocked” digest to the team.

Routines vs interactive Claude Code

Interactive Claude CodeClaude Code Routines
Where it runsYour terminal, your machineAnthropic cloud
Who’s watchingYouNo one (autonomous)
TriggerYou typeCron / API / GitHub
Best forBuilding features, debugging, explorationRepeatable, scheduled, async work
ModelClaude Opus 4.7Claude Opus 4.7
MCP connectors✅✅
Skills✅✅
Cost modelClaude Code subscriptionTBD (preview)

Routines vs Claude Managed Agents (Outcomes, Dreams, Multi-agent)

These all came out of Code with Claude 2026, and people mix them up. Quick map:

  • Claude Managed Agents — the umbrella platform for autonomous Claude agents.
  • Outcomes (Public Beta) — rubric-based eval system; producer agent iterates until grader agent passes.
  • Dreams (Research Preview) — agents learn from past sessions, extract patterns, reorganize memory.
  • Multi-agent Orchestration (Public Beta) — lead agent decomposes a task and delegates to up to 20 specialist agents.
  • Code Routines (Research Preview) — scheduled/triggered headless Claude Code workflows.

Routines is the trigger + scheduling tier. Outcomes and Dreams are the quality + memory tier. Multi-agent is the scale tier.

Routines vs GitHub Actions

GitHub Actions runs scripts you wrote. Claude Code Routines runs an agent that thinks. Concretely:

GitHub ActionsClaude Code Routines
What runsBash / YAML steps you authoredClaude Code agent with a prompt
DecisionsFixedAgent reasons each run
Tool useWhatever’s in the runner imageMCP + skills + sandbox
Cost modelPer minute computeTBD (per run / per minute Claude)
Best forDeterministic CI/CDOpen-ended autonomous tasks

Use GitHub Actions for deterministic work (build, test, deploy). Use Routines for judgment work (review, triage, summarize, propose).

Strengths

  • Autonomous, not interactive — runs while you sleep.
  • Same skills + MCP as Claude Code — no rewrites.
  • GitHub-native triggers — slot into the workflow you already have.
  • Anthropic cloud compute — no self-hosting Claude Code agents.
  • Backed by Code with Claude 2026 ecosystem (Outcomes evals, Dreams memory, etc.).

Weaknesses

  • Research Preview — not GA, no SLA, gated access.
  • No pricing finalized in May 2026.
  • Anthropic-only — no Claude-Code-Routines-via-OpenRouter style multi-vendor.
  • Compute on Anthropic cloud — some regulated industries may want self-hosted or BYO compute.
  • Black-box runtime at preview — observability story still being built.

How to think about using it

  1. Start with one Routine — pick the most painful weekly chore.
  2. Author the Claude Code prompt locally first to verify it works.
  3. Add a trigger — cron is easiest; GitHub events are second easiest.
  4. Pipe output to Slack / Linear / GitHub PR / email.
  5. Add Outcomes evals once you have a few Routines to measure them.
  6. Move to Multi-agent Orchestration if your Routine needs to decompose across specialists.

What’s next

  • GA after Research Preview.
  • Pricing announcement.
  • Tighter integration with the Claude skill marketplace.
  • Better observability dashboard.
  • Cross-cloud compute options.

TL;DR

Claude Code Routines = headless, scheduled, triggered Claude Code on Anthropic’s cloud. The autonomous tier of Anthropic’s developer experience. Announced at Code with Claude 2026, currently in Research Preview, expected to GA later in 2026. If you already love interactive Claude Code, Routines is the natural next step.

Sources: Anthropic “Code with Claude 2026” recap (anthropic.com), Simon Willison conference notes (simonwillison.net/2026/May/6/code-w-claude-2026/), VentureBeat coverage of Dreaming + Code with Claude announcements, dotzlaw conference notes — May 2026.