n8n-MCP

n8n-MCP is an open-source MCP server that lets Claude, ChatGPT, Cursor, and other AI assistants operate n8n: read failed executions, edit workflows in place, validate and test them, manage credentials, and build new workflows. It works with any n8n instance that has the Public API enabled, and can use n8n's built-in MCP server as a second backend.

What You Can Do With It

• Debug failing workflows - n8n_executions reads an execution and, with mode: "error", returns the error, the failing node, and sample items from the node before it.
• Edit workflows in place - n8n_update_partial_workflow applies targeted operations to the nodes and connections named in the request, so the assistant does not resend the whole workflow. This includes find-and-replace inside one large field such as a Code node (patchNodeField). By default n8n-MCP saves a snapshot before it changes a workflow's nodes or connections, and n8n_workflow_versions can diff and roll back.
• Validate and fix before going live - validate_node, validate_workflow, and n8n_validate_workflow report unknown nodes, invalid node versions, missing required fields, and broken connections or expressions. n8n_autofix_workflow repairs the common ones.
• Test - n8n_test_workflow runs a workflow through its webhook, form, or chat trigger, or, through n8n's built-in MCP server, with pinned data when it has no such trigger.
• Manage the instance - create, update, and delete credentials, work with data tables and folders, and run a security audit (n8n_audit_instance).
• Build new workflows - bundled knowledge of 3,501 nodes and 2,352 templates, so the assistant configures real nodes instead of guessing.
• Work across several n8n instances - on the hosted service, register production, staging, and client instances once and name the target on each call. Self-hosted HTTP mode can serve several instances with per-request instance headers.

n8n-MCP and n8n's Built-In MCP Server

n8n ships its own instance-level MCP server, and it covers a lot: building and editing workflows, executions, version history, data tables, and Agents. n8n-MCP does not replace it. The two differ in how they are built:

	n8n's built-in MCP server	n8n-MCP
Runs	Inside one n8n instance	As a separate server that talks to n8n
Instances per connection	One	Several (hosted service, or multi-tenant HTTP mode)
Credentials	Lists credentials and edits their descriptions	Creates, updates, and deletes credentials
n8n versions	Present since late n8n 1.x; most tools need n8n 2.12 or later, and each tool has its own minimum version	Workflow and execution tools work through the Public API on older versions too, including the 1.x line. Credential actions depend on what that version's API exposes, and newer tools state the version they need
Workflow access	Most operations need the workflow's "Available in MCP" setting, which is off by default	Every workflow the API key can reach

The built-in server is enough when you work with a single instance on a recent n8n 2.x, build and edit workflows on that instance, and set up credentials yourself in the n8n UI.

Add n8n-MCP when you run more than one instance, want the assistant to create and update credentials, are on an older n8n version that predates the built-in tools you need, or want autofix, execution error analysis, templates, and a security audit.

Or use both through one connection. Set N8N_MCP_ACCESS_TOKEN and n8n-MCP routes the operations the Public API cannot perform to n8n's own server: test runs with pinned data, n8n's native version history, live dropdown values, n8n Agents, and data-table column changes. The assistant still sees one MCP server. See Official MCP Setup.

Naming note: n8n's documentation uses n8n-mcp as the example connection name for its built-in server in some client configurations (for example claude mcp add --transport http n8n-mcp https://<your-n8n-domain>/mcp-server/http). That name points at n8n's own server. This project is a separate open-source server: github.com/czlonkowski/n8n-mcp, published on npm as n8n-mcp.

Node and Template Coverage

The documentation and validation tools work without an n8n connection, against a bundled database:

This is the beginning of the README. Read the rest at the original. Read the full README ↗