Which n8n node connects to a remote MCP server?

The MCP Client Tool node. Add it as a tool on an AI Agent node, set its endpoint to the server's URL, choose an authentication method, and pick which of the server's tools the agent can call. This is different from the MCP Server Trigger node, which makes n8n itself an MCP server for other clients to call — the opposite direction.

SSE or Streamable HTTP — which transport should I use in n8n?

Use whichever transport the MCP server documents, usually Streamable HTTP. Current versions of the MCP Client Tool node (node version 1.2 and later) have an Endpoint field and a Server Transport option that defaults to HTTP Streamable. Version 1 had only an "SSE Endpoint" field, which is what n8n's docs page still shows, so an older workflow's node may need to be replaced with a fresh one.

Does n8n support OAuth for connecting to an MCP server?

Yes, two ways. If the server is in n8n's built-in MCP servers registry, open the tools panel on an AI Agent node, search for the server, and sign in — n8n creates the MCP credential for you. For a server outside the registry, add the MCP Client Tool node manually and choose OAuth2 as its authentication method, alongside Bearer, header, and no-auth options.

How to Add a Remote MCP Server to n8n

How do I add a remote MCP server to n8n?

Add an AI Agent node to your workflow, then attach an MCP Client Tool node to its tools input. Configure the MCP Client Tool node with the server's endpoint, an authentication method, and which tools to expose — then the agent can call that server's tools during a run. n8n's MCP Client Tool documentation describes the node as "a Model Context Protocol (MCP) client, allowing you to use the tools exposed by an external MCP server."

Two nodes matter here, and they point in opposite directions:

MCP Client Tool — n8n reaches out and consumes an external server's tools. This is the node you want to connect n8n to someone else's MCP server.
MCP Server Trigger — n8n exposes its own tools and workflows to external MCP clients (Claude Desktop, Claude Code, etc.). This is a different node with a different job; per n8n's docs, it "allow[s] n8n to act as a Model Context Protocol (MCP) server."

This guide covers the Client Tool node — connecting n8n to a remote server someone else runs.

Requirements

An n8n instance — self-hosted or n8n Cloud, both support the MCP Client Tool node the same way
An AI Agent node already in your workflow (the MCP Client Tool node attaches to it)
The target MCP server's connection URL, and whatever credential it requires (bearer token, API key headers, or an OAuth app)

Adding the node

On the AI Agent node in your canvas, click Tool + to open the tools panel.
Search for the service. If it's in n8n's built-in MCP servers registry, select it, sign in when prompted, and n8n creates the credential automatically — skip to "Choosing which tools to expose" below.
If the server isn't in the registry, choose the MCP Client Tool node from the panel instead and configure it manually (next section). Per n8n's MCP servers documentation: "If the server you want isn't in the registry yet, use the MCP Client Tool node to connect to any MCP server manually with its connection URL and credentials."

The registry's contents change frequently — n8n's docs say to browse the current list in the node panel rather than rely on a static list. Treat "is my server in the registry" as something to check live, not something to assume from an older guide (including this one).

Configuring the node manually

For a server outside the registry, the MCP Client Tool node's parameters are:

Endpoint and Server Transport — the server's URL, and HTTP Streamable (the default in node version 1.2 and later) or SSE. n8n's docs page still shows the version 1 field, "SSE Endpoint", which only spoke SSE; the current fields come from the node's source code. If you only see "SSE Endpoint", the node in your workflow is version 1 — add a new MCP Client Tool node to get the transport choice.
Authentication — Bearer, generic Header, Multiple Headers (for a server that needs more than one header, such as an API key plus a username), OAuth2, or None.
Tools to Include — All, Selected (pick specific tools), or All Except (expose everything but the ones you exclude).

Authentication options

n8n's MCP Client Tool node supports four ways to authenticate to a remote server, plus no auth:

Bearer — a single bearer token
Header — one custom header, name and value
Multiple Headers — several headers at once, useful when a server wants an API key and a separate identifying header
OAuth2 — a standard OAuth2 credential
None — for a server that doesn't require authentication

If the server is in n8n's registry, you don't configure any of this by hand: selecting the server and signing in creates the credential for you.

Choosing which tools to expose

Whether you connected through the registry or configured the node manually, Tools to Include controls what the agent can call:

All — every tool the server publishes
Selected — you pick specific tools from the list the server returns
All Except — everything except the tools you exclude

Scoping this down matters once a server exposes many tools — a narrower toolset means fewer choices the model has to reason over per turn, and less room for it to reach for the wrong one.

Self-hosted vs. n8n Cloud

The MCP Client Tool node itself works the same on both — it's an outbound connection to a remote server, so there's nothing on the n8n side that differs by hosting type.

The distinction shows up if you're running the other node, MCP Server Trigger, to expose n8n's own tools to a client: n8n's docs note that in queue mode with multiple webhook replicas, MCP traffic (SSE or streamable HTTP) needs to be routed to a single dedicated replica, since those transports require a persistent connection to one server instance. That's a Server Trigger concern, not a Client Tool one — worth knowing if a future workflow has n8n playing both roles.

Common failures

Connection fails immediately. Recheck the endpoint URL against exactly what the server publishes — path and trailing slash included.
Transport mismatch. An old version 1 node speaks SSE only, so a Streamable-HTTP-only server won't connect. Replace it with a new MCP Client Tool node and set Server Transport to HTTP Streamable.
Auth rejected. Confirm you picked the method the server actually expects — Bearer and a generic Header credential are easy to swap by mistake, and a server needing two headers will reject a single-Header credential.
Agent doesn't reach for the tool. Check the agent's system prompt references the task the tool handles, and check Tools to Include — an "All Except" list can silently exclude the tool you meant to expose.
Behind a reverse proxy (self-hosted, if also running MCP Server Trigger). n8n's docs call out disabling proxy buffering, gzip, and chunked transfer encoding on the MCP endpoint in nginx — worth checking if you're running n8n behind one and see connections drop.

A worked example

Tempreon is a hosted MCP server. To try the pattern above against it: add an MCP Client Tool node, set its endpoint to

https://api.tempreon.com/functions/v1/tempreon-mcp/mcp

and set Authentication to OAuth2 with an MCP OAuth2 API credential, leaving Use Dynamic Client Registration on. n8n's OAuth service then discovers the server's endpoints, registers itself, and runs the sign-in with PKCE. Set Tools to Include to whichever of Tempreon's tools the agent's task actually needs, then attach the node to your AI Agent. This is an example of the config shape, not a tested integration. It is one example of the general pattern above, not a claim that n8n has certified or specially optimized this integration — the steps are the same ones you'd follow for any remote MCP server. The account-holder walkthrough is in the bridges setup doc.