/n8n-cli-from-mcp — MCP / Skill → n8n Workflow

Take an MCP server config or a Claude Code skill and generate an n8n workflow that does the same thing without needing an agent at runtime.

Procedure

1. Identify the source:

   • MCP server: ask for the server name, command, and tools it exposes (or read from ~/.config/claude/mcp.json style config)
   • Claude Code skill: ask for the skill name or path to its SKILL.md

2. Read the source's actual capabilities:

   • For an MCP server: list the tools, their input schemas, what they return
   • For a skill: read the procedure section, identify what commands/APIs it calls

3. Map MCP tools / skill steps to n8n nodes:

   • HTTP API call → HTTP Request node
   • Slack post → Slack node
   • Notion read/write → Notion node
   • File read/write → Read/Write Files node
   • Code execution → Code node (JS or Python)
   • LLM call → OpenAI / Anthropic node
   • Branching logic → IF / Switch node

4. Identify what the workflow's trigger should be:

   • If the original MCP/skill was invoked by an agent on demand → Manual Trigger or Webhook
   • If it was meant to run on a schedule → Schedule Trigger
   • If it was reactive to events → Webhook Trigger or specific event trigger node

5. Build the workflow JSON:

   • Start with the trigger
   • Chain the mapped nodes
   • Add credentials where needed (use existing instance credentials, not new ones)
   • Add error handling for any external API calls

6. Show the proposed workflow before importing:

   • Print the node-by-node plan
   • Highlight any places where the mapping is approximate ("MCP tool X has no direct n8n equivalent — using Code node with the same logic")

7. Import (only after user approval):

   • n8n-cli wf import /tmp/from-mcp.json (don't activate)
   • Walk the user through testing it manually first
   • Then activate

Output format

• Show the source's tools/steps as a numbered list
• Show the mapped n8n nodes as a parallel numbered list
• Highlight any mappings that need user judgment
• Confirm before importing

Tips

• The most natural conversions are MCP servers that just call HTTP APIs or run scripts — those map 1:1 to HTTP Request or Code nodes.
• MCP servers with deep state (like memory MCPs) don't map well — n8n is stateless between runs unless you wire up a database. Surface this limitation up front.
• Some Claude Code skills are pure prompts (e.g. "summarize this") — those need an LLM node in the n8n workflow, which means a credential and per-run cost.
• For complex MCPs (browser automation, etc.), suggest keeping them as MCPs and having n8n call out to them via HTTP if exposed, rather than rebuilding them.
• Document the resulting n8n workflow with /n8n-cli-document so the team can maintain it without referring back to the original MCP.

Source: 8Dvibes/n8n-cli — distributed by TomeVault.