How do you use n8n as an MCP server for Claude, ChatGPT or Cursor?

Short answer

n8n works with MCP in three ways. The instance-level MCP server, enabled under Settings > Instance-level MCP, lets Claude, ChatGPT, Cursor, Codex and other clients search and run the workflows you mark as available. The MCP Server Trigger node turns a single workflow into its own MCP server. And the MCP Client Tool node lets n8n's own AI agents call outside MCP servers.

Three ways n8n and MCP meet

People searching for “n8n MCP server” mean different things. n8n has an answer for each.

Instance-level MCP server MCP Server Trigger node MCP Client Tool node
Direction AI tool calls n8n AI tool calls n8n n8n’s AI agent calls another server
Set up Once per instance, under Settings > Instance-level MCP Inside one workflow Inside one workflow
What is exposed Every workflow you mark as available Only the tools attached to that trigger n/a
Sign-in OAuth (recommended) or a personal access token None, n8n User Auth (OAuth2), Bearer or Header auth Per external server
Transports Remote, over your instance URL SSE and streamable HTTP, no stdio Per external server

For most teams the instance-level server is the simplest: one connection, central sign-in, and you choose which workflows the AI may run. The trigger node fits when you want one tightly scoped server, for example a single “create a quote” workflow for one agent.

For the third column, the MCP Client Tool node, n8n’s docs list the sign-in choices as Bearer, generic header, multiple headers, OAuth2 or none. You can expose all of a remote server’s tools, only a selected set, or all except the ones you exclude, which is a cheap way to keep write tools away from an agent that only needs to read.

Setting up the instance-level server

  1. Enable it. An instance owner or admin goes to Settings > Instance-level MCP and selects Enable MCP access.
  2. Choose workflows. Access is opt-in per workflow: use the Available in MCP toggle in the workflow’s settings, the workflow card, or the Workflows enabled page. Newer versions add bulk toggles per project or folder; automatic exposure of new workflows is off by default.
  3. Connect a client. In the Connect a client dialog, pick the tool. n8n names Claude Desktop, Claude Code, ChatGPT, Codex, Cursor, Gemini CLI, VS Code, Windsurf and others. OAuth is recommended; an API key (a personal access token) also works.
  4. Check reachability. Cloud-based clients need a publicly reachable instance. Behind a reverse proxy, pass the MCP-Protocol-Version, Mcp-Method and Mcp-Name headers.

The client then gets tools such as search_workflows and execute_workflow. The latter runs the published version by default. From n8n 2.13.0 clients can also create and edit workflows and data tables.

The common misunderstanding

“We only enabled two workflows, so the AI only sees two.” Not quite. search_workflows returns previews of every workflow the user can view, enabled or not; only the full data and execution are limited to enabled ones. And access is not scoped per client: if you connect both Claude and Cursor, both see all MCP-enabled workflows that user can access. Combined with the ability to create and edit workflows in recent versions, that is a reason to think about which user account an AI tool connects with.

Two more caveats from n8n’s docs: clients that connect with an API key do not appear in the list of connected clients you can revoke, and for the trigger node in queue mode with several webhook replicas, all /mcp* requests must go to one replica or connections break.

What this means for a team

n8n through MCP is a practical way to give every AI tool in the team the same actions: “create the invoice”, “add the lead”, “book the meeting” become workflows that Claude, ChatGPT and Cursor all call the same way. That beats each person building their own integration.

The workflows hold the actions, not the judgment around them. Which customer needs a different payment term, or which lead should never get an automatic follow-up, is knowledge the AI still has to get from somewhere. For the security side, read is MCP safe to use in a company. For comparing n8n with Bombos directly, see n8n or Bombos. New to the concept? Start with what is an MCP server.

Where Bombos fits

n8n gives your AI tools actions; Bombos gives them shared memory about when and how to use them. It is a plugin for Claude, Claude Code, ChatGPT, Copilot, Cursor, Codex, Grok and your own agents, with team knowledge and rules the AI learned about how you work, and it reads what already sits in mail, documents and repos, with a trail of what each answer rests on. It runs in our cloud or on your own server, and one person can try it first with their own agents.

Related questions

Which n8n workflows can an AI tool see through MCP?

Only the ones you enable, one by one, for example with the Available in MCP toggle in a workflow's settings. The exception is search_workflows, which returns previews of every workflow the user can view, without the full workflow data.

Can an AI tool create or change n8n workflows through MCP?

From n8n 2.13.0 the instance-level MCP server can also create and edit workflows and data tables, according to n8n's documentation.

Does the MCP Server Trigger support stdio?

No. It supports SSE and streamable HTTP. For a local stdio-only client you need a bridge.

Can I limit one client to certain workflows?

Not with the instance-level server. n8n says access is not scoped per client: every connected client sees all MCP-enabled workflows the user can access. For a narrow scope, use an MCP Server Trigger per workflow.

Does ChatGPT need my n8n instance to be public?

Yes. Cloud-based clients such as ChatGPT and Claude.ai need a publicly reachable instance.

Sources

  1. Accessing n8n MCP server (n8n docs)
  2. MCP Server Trigger node (n8n docs)
  3. MCP Client Tool node (n8n docs)
The first half hour · free

What does your AI forget?

  1. 01You tell us which AI you use and where it keeps starting over.
  2. 02We show what a shared memory changes there, on your own examples.
  3. 03You keep the plan, with or without Bombos.
Piet Baudoin
Piet BaudoinYou talk to Piet, from the Bombos teamDutch team from Utrecht and Nijmegen · data in Europe
We reply the same working day.