Automation
n8n MCP: how to connect your workflows to Claude and ChatGPT, step by step
n8n MCP covers three separate features. Here is which one to use for what you want, how to connect Claude or ChatGPT and which security settings to check before going live.
n8n MCP: three features that often get mixed up
"Can we connect Claude to our n8n workflows?" Yes, and that is what n8n MCP means: n8n's support for the Model Context Protocol, the standard that lets an AI assistant use outside tools. n8n supports it in three different ways, which is where most of the confusion comes from.
In one direction, n8n is the client. An agent built in n8n uses the tools of an external MCP server, such as a project management app's. That is the job of the MCP Client Tool node. The MCP Client node, without "Tool", does the same outside an agent and calls one specific tool as an ordinary workflow step.
In the other direction, n8n is the server. The MCP Server Trigger node turns a workflow into an MCP server exposing a few chosen tools, which Claude or ChatGPT can then call. Instance-level MCP goes further: a connected assistant can search your workflows, run them, read executions and, from version 2.13, create or edit workflows from a description.
This page assumes you already know what MCP is and sticks to n8n. Menu names follow the official n8n, Claude and OpenAI documentation as of November 2026. They change often, so check them before following the steps.
Which n8n MCP feature fits your goal
Start from the outcome you want. Each need maps to one feature, with its own settings.
Your n8n agent needs an external tool
Use the MCP Client Tool. Paste the server address into the SSE Endpoint field, pick the authentication (Bearer, header, multiple headers, OAuth2 or none) and decide which tools the agent sees: all, a selection, or all except some. For services in n8n's MCP registry, you connect from the nodes panel without creating a credential.
An assistant needs to trigger one action
Use the MCP Server Trigger and attach the tools to it, for example a sub-workflow through the Custom n8n Workflow Tool node. The assistant sees only those tools, and the logic, checks and permissions stay in n8n.
An assistant needs to work across the instance
Turn on instance-level MCP. It suits someone who builds and tests workflows from Claude or a coding agent. It gives the assistant far more power, so it calls for more care.
A fixed step needs to call an MCP tool
The MCP Client node calls a tool chosen in advance, with parameters entered by hand or as JSON, without an agent or a language model. It is the most predictable option.
Publish a workflow as an MCP server for Claude
The route with the MCP Server Trigger and Claude, for a first read-only tool.
- 01
Create the workflow and trigger
Add an MCP Server Trigger node. It does not pass data to a next node; it acts as the entry point for the tools you attach to it as sub-nodes.
- 02
Attach one or two tools
For example, one that looks up an order's status in your database. Give each tool a clear name and description, because the assistant relies on them to decide when to call it.
- 03
Lock down access
In the trigger, choose authentication: Bearer, header or n8n User Auth (OAuth2) on recent versions of the node. Leave "None" for testing only.
- 04
Test, then publish
n8n creates a test URL, which shows data in the editor, and a production URL registered when you publish the workflow. Production calls appear in the Executions tab.
- 05
Add the connector in Claude
On Free, Pro, Max and Ultra plans, go to Customize > Connectors, click Add custom connector and paste the production URL. On Team and Enterprise, an Owner adds it in organisation settings and each member connects with their own account. The Free plan allows one custom connector.
- 06
Run a first test
Ask Claude a question that should trigger the tool, then check the execution in n8n: which data came in and what the tool returned.
Connecting Claude or ChatGPT to instance-level MCP
Instance-level MCP is turned on in Settings > Instance-level MCP with the Enable MCP access button, which needs instance owner or admin rights. No workflow is visible by default. You tick Available in MCP in each workflow's settings or from the workflow list. The option that exposes new workflows automatically starts switched off, and it is best left that way.
For authentication, n8n recommends OAuth: in Connection details you pick your client and approve access when it sends you back to n8n. The alternative is a personal access token tied to your account, and generating a new one revokes the old one. The address you give the client ends in /mcp-server/http and is not the editor address in your browser.
In Claude, you add it from the same custom connectors menu. In ChatGPT, OpenAI's developer documentation gives the route Plugins, plus button, Add custom MCP server, then the URL and authentication. According to OpenAI's help centre in November 2026, full access to custom MCP servers is limited to Business, Enterprise and Edu workspaces, and on Business only admins and owners publish those connections. ChatGPT also asks for manual confirmation before any write action.
Every connected client sees every workflow you have exposed, with no per-client separation. The run tool executes the published version of a workflow by default, and turning MCP access off disconnects all clients at once.
Security checks before going live
Five checks that apply to all three features.
Authentication on every entry point
Bearer, header or OAuth on the MCP Server Trigger; OAuth or a token on the instance. Without authentication, anyone who knows the URL can use the server.
As few tools as possible
In the MCP Client Tool, prefer Selected over All. As a server, expose a tool that finds one order instead of full access to the database.
Read before write
Start with tools that only look things up. Put a human approval step inside the workflow before anything is sent, paid or deleted.
Instructions hidden in the data
An email or document the assistant reads can carry malicious instructions. Claude's documentation warns about this: connect only trusted servers and watch for unexpected behaviour.
Infrastructure
Behind a reverse proxy, turn off buffering for the MCP endpoint (proxy_buffering off in nginx). In queue mode with several webhook replicas, route all /mcp requests to a single replica.
Example
A Leeds accountancy practice gives Claude its file statuses
In this illustrative case, an accountancy practice in Leeds with twelve staff answers the same client questions every day: have you received my bank statements, how far along is my tax return. The answers sit in the practice's job tracking tool, which only two people know well.
The practice builds an n8n workflow with an MCP Server Trigger and a single tool that looks up a job's status from the client's name. The tool reads without changing anything and returns three fields. The trigger requires authentication, and the Owner of the Team workspace adds the connector in Claude.
For a month, staff ask Claude, which calls the tool and drafts the reply to the client. A person checks it before it goes out. A second tool, which would chase clients for missing documents, waits until the first has proved itself.
To measure it, the practice counts the questions answered through the tool, compares response times before and after, and reviews the n8n executions every week to spot unexpected calls.
Where to start with n8n MCP
Pick a question your team often asks a piece of software, expose one read-only tool that answers it and watch for a few weeks what the assistant does with it. Then decide whether to add tools or move up to actions with approval.
Obinova is an AI agency based in Barcelona and Paris. Through our n8n automation service, we build AI workflows and agents with your teams, connected to your tools. We agree with them what the assistant may do alone and what needs approval, and we document every connection so they can keep improving it.
Frequently asked questions
What is the difference between the MCP Client Tool and the MCP Server Trigger?
They work in opposite directions. With the MCP Client Tool, the n8n agent reaches out to tools another server offers. The MCP Server Trigger does the reverse, turning an n8n workflow into an MCP server whose tools Claude, ChatGPT or other clients can call. One project can use both.
Does n8n MCP work with ChatGPT?
Yes. n8n lists ChatGPT among the clients for its instance-level MCP server, and an MCP Server Trigger can also be added as a custom server. In ChatGPT it depends on the plan: in November 2026, OpenAI limits full connections to your own servers to Business, Enterprise and Edu, under admin control.
Can you do RAG with n8n and MCP?
Yes. An n8n agent can query a vector database through n8n's vector store nodes before it answers. You can also build a document search workflow and expose it through the MCP Server Trigger, so Claude or ChatGPT can query it. The second route avoids repeating the search logic in every assistant.
Is it safe to connect Claude to n8n?
It can be, with clear rules: authentication on every entry point, only the tools needed, read before write and human approval for actions that cannot be undone. Review the executions weekly at first. Remember that every client connected to the instance-level server sees every exposed workflow.
Why does claude.ai ask me to sign in when I set no authentication?
n8n's documentation covers this: claude.ai custom connectors prompt users to sign in even when the MCP Server Trigger's authentication is set to "None", and it is the only client known to do so. Either way, set up real authentication, such as Bearer or OAuth2, before exposing the server.