Ask an LLM to build an n8n workflow and it will happily emit JSON — which then breaks, because n8n ships 2,000+ nodes whose property names and connection shapes the model is mostly guessing at. The fix here isn't a smarter model; it's giving the agent a lookup table. This MCP server exposes every node's real schema, operations, and a library of working templates so the agent grounds its output in fact rather than memory.
What Sets It Apart
- Catalog over recall: 2,063 nodes (816 core + 1,247 community) with ~99% property coverage and 2,352 templates — the agent queries the live catalog instead of leaning on stale or partial training data.
- Validation in the loop: beyond docs lookup, it ships tools that check a workflow before it ever reaches your instance, catching bad node references and broken connections early.
- Optional write path: with n8n API credentials, 13 management tools let the agent create, update, run, and inspect workflows end to end — not just describe them.
- Agent-node awareness: 265 AI-capable tool variants are flagged, so building agentic workflows inside n8n is a first-class case rather than an afterthought.
Who It's For
Great fit if you drive n8n through Claude Desktop, Claude Code, Cursor, or Windsurf and you're tired of hand-fixing hallucinated node configs. It's an independent project (by Romuald Członkowski / AiAdvisors), not an official n8n release, so treat it as community plumbing rather than vendor support. Look elsewhere if you don't run n8n or want a visual no-code builder — this serves AI clients, not a UI. Community-node coverage is broad but trails core-node accuracy, so double-check niche nodes.
