External Triggers for Claude Code

🚀 Main message

No first-party mechanism gives the Claude Desktop Code tab a stable inbound endpoint that also buffers events. What works is a stable ingress that stores events (cloud or SaaS) plus an outbound poller or runner on the PC. Tunnels alone fix only half the problem: every tunnel returns an error while the PC is off.

Source: a deep-research run on 2026-10-15 (30 items × 37 fields, one agent each). The full 750 KB report is archived at _raw/processed/2026-11-01_2026-10-15-research-claude-code-external-triggers.md; claims come from the research agents and are not independently verified, apart from Svix’s 7-day Free retention, which was checked on its pricing page.

📒 The two sub-problems

  1. The sender needs a stable endpoint (my local URL changes).
  2. The event must survive while the PC or app is off.

🧩 Designs that solve both

#DesignGood forLimits
1Cloud-first: Claude Code Routines API triggerWork that needs no local files; stable URL on Anthropic’s side, runs with the PC off30 fires/hour per routine, 100/hour per account; no idempotency key (retries create duplicate sessions); text up to 65,536 characters; no local files or locally added MCP servers; Local (Desktop) routines have no API trigger
2GitHub repository_dispatch + self-hosted runner running claude -pRepo work on my PCJob waits up to 24 h for an offline runner; run the runner as my own user, private repo only; runs do not appear in the Desktop Code tab
3Hosted queue + outbound poller on the PC — a Cloudflare Worker (verifies signature, enqueues) with Cloudflare Queues, or Svix IngestDesktop work on local filesRetention: Cloudflare Queues 24 h Free / 4–14 days Paid; Svix Free up to 7 days
4AgentsRoom (packaged)Easy setupVendor server queues events up to 7 days and sees every payload; closed source, single maintainer; runs in its own app, not the Code tab
5Always-on host running claude -pRepo or service work, both problems solvedLoses local files; unattended runs need --permission-mode dontAsk with an allowlist; claude setup-token gives a one-year subscription token

Poller options for design 3:

  • a mod ($.clock.every + $.prompt.submit) — runs in the Code tab but only while a session is open, needs de-duplication across sessions, and the loop has not been tested end to end (see Claude Code Mods (RoboNuggets))
  • a Desktop Scheduled Task — minimum 1 minute, app open and PC awake, one catch-up run for the latest miss within 7 days; every poll costs a model session unless a script checks the queue first
  • claude -p from Task Scheduler — invisible in the Desktop app, and two open Windows scheduler bugs were reported

⚠️ What does not help here

  • Tunnels alone (Cloudflare Tunnel, Tailscale Funnel, ngrok, Dev Tunnels, zrok, Azure Relay) and Hookdeck without a live listener
  • Channels (CLI only), cross-session messaging (no public API), Remote Control and Projects (nothing external can start them), Dispatch, deep links and Slack (a human is in the loop), background sessions (terminal only)
  • Chat bridges, OpenClaw and Hermes Agent gateways, GitHub/Linear bridges (no durability, or abandoned), Power Automate, Temporal and Inngest (licensing or over-engineered), Managed Agents (not Claude Code, API billing, WSL)

📖 Full item list (fit for my scenario)

  • High: Routines, mods as in-session poller, GitHub self-hosted runner, always-on host with headless Claude Code, Svix Ingest + polling, outbound pull queues (Cloudflare Queues, SQS, Upstash Redis), AgentsRoom (flagged uncertain)
  • Medium: self-hosted environments for routines, Channels, hooks as event intake, background sessions, claude -p as own receiver, Desktop Scheduled Tasks / /loop / Cron / Monitor, claude-code-action on GitHub Actions and GitLab, n8n / Make / Zapier glue, tunnels, Hookdeck, chat/email transports
  • Low: cloud-session CLI (--cloud, --teleport), cross-session inbox, Remote Control, Dispatch/deep links/Slack, PR/CI monitor and Auto-fix delivery, Managed Agents with a self-hosted worker, workflow engines, OpenClaw/Hermes gateways, event bridges, chat bridges

✅ To verify before building

  • an end-to-end test of the mod poller in the Desktop Code tab on Windows
  • whether polling endpoints are enabled on Svix’s Free plan
  • whether a cloud routine can message a local Remote Control session (would be a native cloud-to-local path)
  • AgentsRoom’s trust model, since the vendor sees all payloads

Resources