Hey n8n Community,
Follow-up to my previous post where we shared what happened using Claude as a co-engineer on a production n8n project. A lot of people asked about the practical setup, so here’s the breakdown.
What this actually is
Claude connects to your n8n instance through MCP. It can search workflows, read node configs, and execute workflows. Not generating JSON you paste in – it’s working with your live instance. Think pair programming where your partner has system access.
What you need
Self-hosted n8n with a publicly reachable endpoint (for cloud MCP clients). Toggle MCP access in Settings → Instance-level MCP, then enable it per workflow. Nothing is exposed by default.
Auth scoped correctly. OAuth or access tokens. Your token is only shown in full once – copy it immediately. Take time here. You don’t want Claude triggering production workflows while experimenting.
Strong model. We ran Opus. The difference vs earlier models is significant for n8n-specific stuff – node types, expression syntax, error handling branches.
Stack your connectors. n8n alone is useful. Adding Google Drive, Chrome, and Slack lets Claude pull files, verify live data, and message your team – all in one conversation. That’s where the real leverage is.
If using community MCP nodes, set N8N_COMMUNITY_PACKAGES_ALLOW_TOOL_USAGE=true or they won’t work as tools in AI Agent nodes.
How to prompt
Doesn’t work: vague requests like “build me a CRM automation.”
Works: describe the what and why with real context. Reference actual files, actual workflows, actual systems.
Real example from our project:
“Look at the upload workflow. I need a sub-workflow that takes raw supplier files with 13 columns and maps them into our 47-column BSMS_ITEM_UPLOAD format. The January file is the reference.”
Claude returned file ingestion, column normalization, lookup nodes, output formatter. ~80% correct. The 20% was edge cases. Treat it like a senior engineer on their first week – technically excellent, needs context about your setup.
Where it’s strong
Workflow architecture, Code node JavaScript, reading and diagnosing existing workflows through MCP (honestly the killer feature), data analysis, and everything around the workflow – emails, presentations, stakeholder comms.
Where it struggles
Can’t inspect live execution logs (you check those and describe what you see). Canvas layout is always a mess visually. Advanced n8n expressions sometimes need manual fixes. Cross-workflow dependencies are tricky. Domain-specific platform behaviors need teaching – but once you explain a pattern, it sticks.
Tips that saved us time
- Start with read, not write. Ask Claude to analyze an existing workflow first. Verifies the connection and shows you how well it understands your setup.
- Enable workflows one at a time. Any connected MCP client sees all enabled workflows – no per-client scoping.
- Keep production separate. Read access to prod, execute only on dev/staging.
- Use it for debugging. Claude diagnosing a stale Google Sheets schema by reading the node config through MCP saved us hours. If something breaks, let Claude inspect it before you dig manually.
- Teach it your patterns. When you fix something Claude got wrong, explain why. It applies the pattern going forward.
Is it worth it?
Simple automations: yes, without hesitation. Setup is an afternoon.
Complex production systems: also yes, but you’re still the architect. The productivity multiplier is real – entire features in our system (built with EasyBits AI pipelines) exist only because Claude made the cost of building them negligible.
Give it a try and let me know how it goes. Happy to answer questions.
Best,
Felix
