Claude Code Cloud Routines
Cloud-hosted scheduled agent feature in Claude Code. Configure a prompt once; it runs on Anthropic's infrastructure on a cron schedule, via API POST, or GitHub event trigger — no local machine required. Announced 2026-04-14.
Key points
- Trigger types: (1) cron schedule (minimum 1-hour interval), (2) API/POST from an external automation, (3) GitHub event webhook (new PR, push, issue, release) [src-010]
- Execution model: each run clones the linked GitHub repo into an ephemeral cloud environment → reads CLAUDE.md and context files → executes the routine prompt → optionally commits changes back → destroys the environment [src-010]
- Resource envelope: 4 vCPUs, 16 GB RAM, 30 GB disk per run [src-010]
- Daily run limits by plan: Pro = 5, Max = 15, Team/Enterprise = 25 [src-010]
- Security: secrets live in cloud environment variables, never in
.envfiles committed to GitHub. Network access has three tiers: Trusted (Anthropic-vetted allowlist), Full (any outbound), Custom [src-010] - Stateless: no shared state between runs — persistent memory must be written to files in the GitHub repo and committed back for the next agent to read [src-010]
- Desktop app: scheduled routines surface in the Claude Code desktop app's calendar view alongside local scheduled tasks [src-010]
- Boris frames routines as the server-side version of loops: the recurring work continues even after the laptop closes, unlike local
/loopsessions [src-054] - Nate's May 2026 deployment comparison treats routines as the strongest "agentic" cloud option for Claude Code users because they preserve the prompt, repo context, tools, skills, and agent loop more directly than plain serverless jobs [src-086].
- Their operational limits still matter: minimum 1-hour interval, plan-based run caps, cloud environment variables, and prompt scoping all become part of deployment design [src-086].
Comparison with related concepts
| Dimension | Cloud Routines | Managed Agents | Local Scheduled Tasks |
|---|---|---|---|
| Infrastructure | Anthropic cloud | Anthropic cloud (SDK) | User's machine |
| Machine required | No | No | Yes |
| Minimum interval | 1 hour | N/A (API-triggered) | 1 minute |
| Context | GitHub repo | Vault + MCP | Local files |
| Target user | Claude Code subscriber | Developers/SDK | Claude Code subscriber |
Related entities
- Claude Code — the product that ships this feature
- Anthropic — infrastructure operator
Related concepts
- Stateless Agent Memory Pattern — required pattern for knowledge persistence across stateless runs
- Claude Code Scheduled Tasks — local-machine alternative; shorter intervals, local file access
- Claude Code Loop Skill — local recurring prompt mechanism that routines generalize to Anthropic-hosted infrastructure
- WAT Framework (Workflows-Agent-Tools) — routines preserve all three layers: Workflows + Agents + Tools
- Routine Network Access Tiers — the security tier model for outbound network access
- AI-Native Organizational Process — routines are one cadence primitive for AI-native work
- Agent Deployment Modes — broader comparison of local loops, scheduled tasks, cloud routines, Modal, and trigger.dev
Source references
- [src-010] Nate Herk — Cloud agents & model releases cluster (2026-04-14 to 2026-04-17)
- [src-013] Nate Herk — "Build & Sell Claude Code Operating Systems (2+ Hour Course)" (2026-05-01)
– Additional detail: cloud routines are accessed via the /routines command; browser-session-based automations (cookies, local state) will fail because the cloud environment has no local browser session.