Claude n8n Bridge — DevOps skill for Claude Code

n8n — self-hosted, Claude-Code-driven

Open-source n8n running locally in Docker, wired so that **Claude Code authors the workflows as files** and **n8n runs them on triggers you don't have to remember**.

• Editor: http://localhost:5678
• n8n 2.35.4 (community), Postgres 17
• Everything in this folder is a file in git except .env and shared/.

---

Why this shape

n8n and Claude Code are good at opposite things, and the split is the whole design:

	Good at	Bad at
n8n	schedules, inbound webhooks, retries, fan-out over rows, API glue, "did this run?"	judgment, reading a repo, anything needing local Python or a signed-in browser
Claude Code	judgment, repo context, writing/verifying content, multi-step reasoning	being awake at 09:00, catching a webhook, remembering to run weekly

So: **n8n owns triggering. Claude Code and the repo's engines own the work.**

n8n lives in a Linux container and cannot run `claude` or `python -m engine.run`. `runner/runner.py` is the bridge — a small HTTP server on Windows that n8n calls at `http://host.docker.internal:8787/run`.

Schedule / webhook          host bridge                     the actual work
┌──────────────┐   HTTP    ┌────────────────────┐  argv   ┌────────────────────┐
│ n8n (docker) │──────────▶│ runner.py (Windows)│────────▶│ claude -p  …       │
│              │  Bearer   │  • token required  │         │ python -m engine…  │
│              │  + task   │  • task allowlist  │         │ (cwd = repo root)  │
└──────────────┘   name    └────────────────────┘         └────────────────────┘

The two rules that keep the bridge safe

The runner executes code, so:

1. Bearer token on every request (RUNNER_TOKEN, shared via n8n/.env).
2. n8n sends a task name, never a prompt and never a shell command. Every runnable thing is declared in runner/tasks.json, which lives in git and which you control. There is deliberately no endpoint that acce