How to connect Claude Desktop to your n8n workflows using either instance-level MCP access or the MCP Server Trigger node.
An n8n MCP server lets an AI assistant like Claude reach into your n8n workflows and use them as tools — instead of you manually triggering automations, the assistant can call them directly during a conversation. This tutorial covers both ways n8n officially supports this: instance-level MCP access, and the per-workflow MCP Server Trigger node.
If you’re new to the underlying standard: MCP (Model Context Protocol) is an open protocol that lets AI assistants discover and call external tools in a consistent way, regardless of which application built those tools. An n8n MCP server is simply n8n acting as one of those tool providers.
Two Different Ways to Run an n8n MCP Server
n8n’s own documentation draws a clear line between two setups, and picking the right one matters before you start connecting anything.
Instance-level MCP access creates one connection for your entire n8n instance, enabled via Settings > Instance-level MCP. Once enabled, you control which workflows (and, on instances with the agents feature, which agents) are exposed to MCP clients. A connected client — Claude Desktop, Claude Code, Cursor, VS Code, and others are all supported — can then use the exposed tools for workflow management, workflow building, and agent management, all through a single instance-wide connection rather than a separate setup per workflow.
The MCP Server Trigger node works differently — you add it inside a single specific workflow, and it exposes only the tools connected to that one workflow. This is the better fit when you want to hand-craft a narrow, purpose-built MCP server rather than opening up your whole instance.
Setting Up Instance-Level MCP Access
This is the simpler path if you want Claude Desktop to reach multiple workflows without building a dedicated trigger workflow for each one.
Step 1: Enable Instance-Level MCP Access
In your n8n instance, go to Settings > Instance-level MCP and select Enable MCP access (this requires instance owner or admin permissions). Once enabled, use the Access section to choose which workflows are exposed to MCP clients.
Important: only published workflows containing a Webhook, Form, Schedule, or Chat Trigger are eligible to be enabled for instance-level MCP access — a workflow without one of these trigger types won’t appear as an option.
Step 2: Connect Claude Desktop via OAuth2
In Claude Desktop, go to Settings > Connectors, click “Add custom connector,” and enter a name (for example, “n8n MCP”). For the Remote MCP Server URL, go back to n8n’s Settings > Instance-level MCP > Connect a client dialog and copy the exact Server URL shown there — it ends in /mcp-server/http, not your normal n8n editor or base URL. Paste that Server URL into Claude Desktop, save the connector, then complete the authorization flow when n8n prompts you.
Step 3: Verify the Connection
Once connected, ask Claude to list your available n8n workflows. If the OAuth2 setup worked, it should return the workflows you enabled in Step 1.
This instance-level path tends to suit teams who already have several stable, tested workflows and simply want an AI assistant able to trigger or query them on demand — you’re not building anything new, just opening a door to what already exists.
Setting Up the MCP Server Trigger Node
Use this approach when you want a single, purpose-built workflow exposed as its own MCP server.
Step 1: Add the MCP Server Trigger Node
Create a new workflow and add the MCP Server Trigger node. On its setup page, copy the Production URL — you’ll need this to connect a client later.
Step 2: Connect a Tool to the Trigger
Add a tool node to the MCP Server Trigger — this defines what the exposed MCP server can actually do. The specific tool depends on what you’re building: a CRM lookup, a database query, a specific app action, and so on.
Step 3: Configure Authentication
Double-click the MCP Server Trigger node, and under Authentication, select Bearer Auth. Create a new credential and paste in a token you’ve generated, then save it.
Step 4: Connect Claude Desktop to This Specific Trigger
Because Claude Desktop expects a stdio-based connection while the MCP Server Trigger node communicates over Server-Sent Events (SSE) or streamable HTTP, you need a small gateway to bridge the two. n8n’s official docs provide the exact configuration to add to your Claude Desktop config file for this, using your trigger’s URL and bearer token.
Step 5: Test the Connection
Restart Claude Desktop after saving the config change, then ask Claude to use the new tool. A successful call means your workflow’s tool executed and returned a result back through the MCP connection.
A Common Connection Quirk to Know About
n8n’s own documentation flags a specific quirk worth knowing before you troubleshoot for an hour: claude.ai’s custom connectors will ask you to sign in to n8n even when your MCP Server Trigger’s Authentication is set to None.
n8n’s docs note that claude.ai is the only client known to behave this way — it assumes every MCP endpoint on your domain requires n8n user authentication, even when the specific endpoint doesn’t. If you hit an unexpected sign-in prompt despite deliberately leaving authentication open, this is the likely explanation rather than a misconfiguration on your end.
A Deployment Detail Worth Knowing
If you’re running n8n in queue mode with multiple webhook replicas, the MCP Server Trigger node needs special routing: n8n’s documentation specifically warns that without routing all MCP-related requests to one single, dedicated webhook replica, SSE and streamable HTTP connections will frequently break or fail to deliver events reliably. For a single-replica setup, this isn’t a concern — it only applies once you’re scaling webhook processing horizontally.
Security Considerations Before You Expose Anything
Before connecting any AI client to your n8n instance, it’s worth thinking through exactly what that client will be able to do once connected. Instance-level access, in particular, can reach every workflow you’ve enabled — so only enable workflows you’re genuinely comfortable having an AI assistant call on your behalf, and avoid enabling anything that performs irreversible actions (deleting records, sending payments, modifying production data) without a manual confirmation step built into the workflow itself.
For the MCP Server Trigger node specifically, always set Bearer Auth rather than leaving authentication open, even during testing — an unauthenticated MCP endpoint on a public URL is reachable by anyone who finds it, not just your intended client.
Instance-Level Access vs. the Trigger Node: Which to Use
- Choose instance-level access if you want Claude to reach several existing workflows without rebuilding each one as a dedicated MCP server.
- Choose the MCP Server Trigger node if you want a single, tightly scoped MCP server exposing only specific tools you’ve deliberately chosen — useful when you don’t want an AI assistant to have broad reach across your entire n8n instance.
This mirrors a similar setup on Make’s own MCP Server — both platforms now let an AI assistant call an automation directly as a tool rather than only running it manually or on a schedule, which pairs naturally with the reasoning capabilities covered in our n8n AI agent tutorial.
A Realistic Use Case
Say you want Claude to be able to check current inventory levels during a conversation, without you switching tabs to look it up. Build a small workflow with an MCP Server Trigger connected to a database or spreadsheet lookup tool, secure it with Bearer Auth, and connect it as a dedicated MCP server. From that point, asking Claude “what’s our current stock of X” triggers the workflow, retrieves the answer, and returns it directly in the conversation — a narrow, purpose-built tool rather than exposing your whole n8n instance.
Frequently Asked Questions
What’s the difference between instance-level MCP access and the MCP Server Trigger node?
Instance-level access is one connection covering your whole n8n instance, with centralized authentication and a list of enabled workflows a client can reach. The MCP Server Trigger node is configured inside a single workflow and only exposes that workflow’s own tools.
Can Claude Desktop connect directly to an n8n MCP Server Trigger?
Not directly — Claude Desktop expects a stdio-based connection, while the MCP Server Trigger node uses SSE or streamable HTTP. n8n’s documentation provides a gateway configuration for your Claude Desktop config file to bridge this gap.
Do I need authentication on my n8n MCP server?
For the MCP Server Trigger node, n8n supports Bearer Auth, which you configure directly on the node — this is the standard way to prevent unauthorized clients from calling your exposed tools. Instance-level access uses OAuth2 through Claude Desktop’s connector setup instead.
Does the MCP Server Trigger node work with n8n running in queue mode?
Yes, but with one caveat: if you run multiple webhook replicas, you must route all MCP-related requests to a single dedicated replica, or SSE and streamable HTTP connections will frequently break.
Can more than one AI client connect to the same n8n MCP server?
Yes — nothing in either setup limits an MCP server to a single connected client. Instance-level access can serve multiple clients, such as Claude Desktop and Lovable. Each client must authenticate using OAuth (recommended) or an API key. A single MCP Server Trigger node can also accept more than one properly authenticated client.
Setting up an n8n MCP server comes down to one decision made early: whether you want broad, instance-wide access for a client like Claude Desktop, or a narrow, purpose-built server scoped to one workflow’s tools. Both are officially supported, and the right choice for your n8n MCP server setup depends on how much reach you actually want an outside AI assistant to have into your automations.
Sources
- n8n — Connect to n8n MCP server (official docs)
- n8n — MCP client connection examples (official docs)
- n8n — MCP Server Trigger node (official docs)
- Model Context Protocol — official introduction
