n8n-mcp is an MCP (Model Context Protocol) server that gives Claude, Cursor, and other MCP-compatible agents 12 tools for working with n8n: generate workflow JSON, lint with 20+ rules, diff workflow versions, scaffold custom nodes, explain failed executions per-node, replay single nodes with synthetic input, render execution timelines, and drive a live n8n instance via REST.

How is n8n-mcp different from other n8n MCP servers?

Other n8n MCP servers compete on breadth (20+ tools, full node catalog). n8n-mcp is focused on debugging and first-run correctness: execution_explain diagnoses failed runs per-node, execution_replay lets you test a single node with synthetic input, workflow_lint catches silent failures (credential drift, rate-limit gaps, expression staleness, deprecated nodes), and workflow_diff tracks what changed between versions.

Does n8n-mcp need a live n8n instance?

No. Seven tools are fully stateless (workflow_generate, workflow_lint, workflow_diff, execution_explain, execution_replay, execution_timeline, node_scaffold) and work offline against JSON pasted into chat. The five REST tools (workflow_list, workflow_get, workflow_create, workflow_activate, execution_list) require N8N_API_URL and N8N_API_KEY env vars.

How do I install n8n-mcp?

Run npx @automatelab/n8n-mcp (no install needed) or npm install -g @automatelab/n8n-mcp (requires Node 20+). Add the server to your MCP host config (~/.cursor/mcp.json for Cursor, claude_desktop_config.json for Claude Desktop), then restart the host. All 12 tools appear in the MCP panel.

What does n8n-mcp cost?

The server is MIT-licensed and free. You pay only for your AI agent's model usage and your n8n instance.

Free · MIT licensed

n8n-mcp

An MCP server that gives Claude, Cursor, and any MCP-compatible agent twelve tools for n8n. Generate workflow JSON, lint for the silent failures, and - the wedge - diagnose failed executions node by node so the agent stops guessing.

Get it on GitHub npm package Read the launch post

What it gives your agent

Twelve tools, split between stateless (work on pasted JSON, no n8n instance needed) and live (drive a real n8n via REST, gated on N8N_API_URL + N8N_API_KEY).

Tool	Type	Purpose
`node_scaffold`	offline	Generate a node scaffold.
`workflow_generate`	offline	Generate importable workflow JSON.
`workflow_lint`	offline	Find structural errors in a workflow.
`workflow_diff`	offline	Compare two workflow versions.
`execution_explain`	offline	Explain node outputs in an execution.
`execution_replay`	offline	Inspect a recorded execution without running it again.
`execution_timeline`	offline	Show the recorded execution sequence.
`workflow_list`	live	List workflows on a configured n8n instance.
`workflow_get`	live	Read a workflow from a configured instance.
`workflow_create`	live	Create a workflow after approval.
`workflow_activate`	live	Activate a workflow after approval.
`execution_list`	live	Read execution history from a configured instance.

What makes it different

Debugging-first, not breadth-first

Other n8n MCP servers index the entire node catalog (20+ tools). This one is focused on the failure modes - the diagnoses an LLM cannot do alone from a workflow dump.

Per-node execution diagnosis

execution_explain calls out which node returned zero items, which expressions failed to resolve, which branch the IF/Switch took. Solves the #1 n8n debugging pain point: silent data loss between nodes.

Opinionated AI-Agent topology

Generator emits proper LangChain clusters - ai_languageModel / ai_memory / ai_tool connections wired upward to the agent, not via main. Imports cleanly on n8n 1.x.

Lint catches the silent killers

Function vs Code, spreadsheetFile vs convertToFile, IF v1 schema, AI Agent without a chat model, Webhook missing webhookId. The errors n8n accepts on import and bites you on first run.

Stateless tools work offline

Generate, lint, and explain don't need an n8n instance. Paste JSON in chat, get answers. REST tools turn on when you set the API key.

Paired Agent Skill

Ships with a SKILL.md that teaches the model when to use which tool, plus references split so they don't bloat the prompt. The skill is portable across MCP-capable harnesses.

Install in 3 steps

Install the npm package (requires Node 20+) npm install -g @automatelab/n8n-mcp
Add the server to your MCP host config (~/.cursor/mcp.json for Cursor, claude_desktop_config.json for Claude Desktop)
```
{
  "mcpServers": {
    "n8n": {
      "command": "npx",
      "args": ["-y", "@automatelab/n8n-mcp"],
      "env": {
        "N8N_API_URL": "https://your-n8n.example.com",
        "N8N_API_KEY": "n8n_..."
      }
    }
  }
}
```
The env block is optional - the 4 stateless tools work without it. Get an API key in n8n at Settings → API → Create API key.
Restart your MCP host. The nine n8n_* tools appear in the MCP panel.

Works with Claude Desktop, Claude Code, Cursor, Cline, and any MCP-compatible agent harness. Full reference: the GitHub readme.

FAQ

What does it cost?

The server is MIT-licensed and free. You pay only for your agent's model usage and your n8n instance.

Do I need a live n8n instance?

No. The four stateless tools (generate, scaffold, lint, explain_execution) work offline against JSON you paste into chat. The five REST tools are gated on N8N_API_URL + N8N_API_KEY - set those when you want the agent to drive a real instance.

Which agents does it work with?

Anything that speaks MCP: Claude Desktop, Claude Code, Cursor, Cline, Continue, and most MCP-capable harnesses. Configuration is the same JSON block in each host's MCP config file.

How is this different from czlonkowski/n8n-mcp?

That server competes on breadth - 20+ tools and an indexed corpus of every n8n node. This one is the debugging-and-first-run-correctness MCP: explain_execution is the wedge, the generator is opinionated about AI-Agent topology, and lint encodes the silent failure modes. Different niches; many users will run both.

What does it not do?

It doesn't index the full n8n node catalog or expose every node's option surface. It doesn't generate workflows from screenshots. It doesn't run executions for you - it diagnoses the ones you already ran.

Migrating from leonardsellem/n8n-mcp-server

The @leonardsellem/n8n-mcp-server npm package was officially deprecated in March 2025 and has had no commits since July 2025. If you found this page from a stale config or a Google result pointing there, you're in the right place.

Drop-in replacement — same environment variables, same MCP host config shape:

Old package	This package	Notes
@leonardsellem/n8n-mcp-server	@automatelab/n8n-mcp	npx @automatelab/n8n-mcp
list_workflows, get_workflow	workflow_list, workflow_get	Same n8n API, new naming
create_workflow, execute_workflow	workflow_create, workflow_activate	Create + activate are separate ops
list_executions, get_execution	execution_list + execution_explain	`execution_explain` adds per-node diagnosis
— (none)	workflow_lint, workflow_diff, workflow_generate, execution_replay, execution_timeline, node_scaffold	New debugging tools, no n8n instance needed

The credential tools (list_credentials, get_credential) are not included — the n8n API does not expose credential secrets, so a credential management MCP is inherently read-only and adds limited value over the n8n UI.

Want it wired into your n8n setup?

We use n8n-mcp daily to build, lint, and debug our own workflows. If you want it set up and tuned for your stack - or a custom workflow built end to end - we can do that.

Get in touch

How to use this result

Start with workflow_lint on exported JSON. Use execution_explain for a failed execution. Live tools need an n8n API URL and key; review workflow_create and workflow_activate before allowing changes.

Current installation and examples: n8n MCP source guide.

Current MCP tool names

node_scaffold
workflow_generate
workflow_lint
workflow_diff
execution_explain
execution_replay
execution_timeline
workflow_list
workflow_get
workflow_create
workflow_activate
execution_list

Tested usage walkthrough

Publish a concrete usage walkthrough with inputs, verification and result limits.

Read the usage walkthrough