n8n-as-code

Operator's take

The pain n8n-as-code is targeting is the one any n8n-fluent operator notices the moment they ask Claude Code or Cursor to "write me an n8n workflow that does X" — generic models don't know n8n's node schema in any meaningful depth, so they invent property names that don't exist, mis-version trigger nodes, or hallucinate connector resources. The fix here is bluntly mechanical: ship the schema as the agent's knowledge, not as something the agent has to retrieve from docs at runtime. 537 nodes, 10,209 properties, 17,155 option values, 7,702 community templates — all installed locally, searchable in ~5ms, and version-pinned to a real n8n release.

The second-order claim is the more interesting one for anyone running n8n in production: the package now closes the runtime loop, not just the authoring loop — the agent can detect which credentials a workflow needs after a push, prompt the user for the secret values it can't infer, provision the credential without bouncing to the n8n UI, activate the workflow, execute it, and inspect failed executions inline. That's the difference between a code-completion convenience and an actual operating surface — once the agent can drive a workflow end-to-end, the n8n UI becomes optional for a large share of day-to-day workflow ops.

What's load-bearing for whether this is worth adopting: how much of your n8n work is genuinely amenable to TypeScript-first authoring vs. clicking through the canvas. The TypeScript decorators are clean and Git-diffable, but n8n's value to many teams is the canvas — visual flow + click-to-configure. n8n-as-code's bet is that AI agents read TypeScript better than they read n8n JSON, and the people most likely to adopt are the ones already comfortable in a coding-agent loop.

What it's good at

• Grounding AI agents in the real n8n schema — 537 nodes (433 core + 104 AI/LangChain), 100% schema coverage, version-pinned to the latest stable n8n release.
• Closing the runtime loop — agent can detect missing credentials, provision them, activate the workflow, execute webhook/chat/form workflows, and inspect failed executions.
• Bidirectional JSON ↔ TypeScript conversion — workflows become Git-diffable and easier for AI agents to read and edit; convert back to JSON anytime.
• GitOps for n8n workflows — pull, edit, push, 3-way merge conflict detection, multi-instance support.
• Multiple agent surfaces from one ontology — VS Code/Cursor extension, Claude Code plugin, OpenClaw plugin, MCP server, raw CLI; same n8nac underneath.
• Open-source (MIT) — no vendor lock-in on the tooling layer.

What it's not

• Not a replacement for n8n — requires a running n8n instance to push to; this is a developer surface over n8n, not a substitute for it.
• Not vendor-blessed — third-party project; outdated n8n instances render generated workflows as broken nodes because the schema bundle is pinned to the latest stable n8n release. The README flags this explicitly.
• Not for non-coders — TypeScript decorators, CLI, and IDE extension are developer-shaped surfaces; the canvas-first n8n user gets less out of this than the agent-first user.
• Not a managed service — self-installed, MIT-licensed; you maintain the toolchain and update cycles yourself.