← MCP directory

n8n MCP server

official

n8n has a built-in instance-level MCP server that lets an AI client find, run and build workflows in your own n8n instance. An instance owner or admin enables MCP and then chooses which workflows are exposed.

What an agent can do

  • → Search exposed workflows
  • → Read workflow metadata and triggers
  • → Trigger and test-run workflows
  • → Create and edit workflows
  • → Work with data tables

Access is chosen per workflow, folder or project, but search shows previews of every workflow the user has access to. Building and editing needs n8n 2.13.0 or later.

About the n8n MCP server

n8n has a built-in MCP server at instance level. It is part of n8n itself and runs in your own instance, on n8n Cloud or self-hosted, so its address is your instance's address ending in /mcp-server/http. One connection covers the whole instance, with central sign-in. That sets it apart from the MCP Server Trigger node, which is configured inside a single workflow and only exposes that workflow's tools.

Teams use it to let an AI assistant search the workflows, run those marked as available in MCP and check how the executions went. The assistant can also test-run a workflow with sample data, build new workflows and edit existing ones, publish and unpublish them, and work with data tables. Building and editing needs n8n 2.13.0 or later, and n8n recommends coding agents over chat clients for that work.

Only instance owners or admins can enable MCP. Without an explicit choice, clients only see previews of workflows, and full access is given per workflow, folder or project. Connected clients can be reviewed and revoked, and the whole feature can be disabled, which disconnects every client. In a read-only instance, from n8n 2.47.0 every tool that changes something is blocked, except running published workflows. The credentials list never returns secrets.

Things to ask

  • Which workflows do we have that deal with invoices, and when were they last changed?
  • Run the workflow that syncs new leads and tell me whether the execution succeeded.
  • Which executions failed in the last 24 hours, and in which node?
  • Build a workflow that posts a summary to Slack when a form is submitted, and test-run it with sample data.
  • Show what differs between the published version of the order workflow and the previous one.

Tools the server exposes

  • search_workflows
  • get_workflow_details
  • execute_workflow
  • test_workflow
  • publish_workflow
  • unpublish_workflow
  • get_workflow_history
  • get_workflow_versions_diff
  • get_workflow_execution
  • search_workflow_executions
  • list_credentials
  • search_projects
  • search_folders
  • list_workflow_tags
  • search_nodes
  • validate_workflow
  • create_workflow_from_code
  • update_workflow
  • archive_workflow
  • restore_workflow_version

How to connect n8n

  1. Have the instance owner or an admin go to Settings, Instance-level MCP and select Enable MCP access.
  2. Mark the workflows the assistant may use as Available in MCP, one at a time or for a whole folder or project.
  3. Open Connect a client, copy the server address ending in /mcp-server/http and pick your client in the list.
  4. Sign in with OAuth, which n8n recommends, and approve access when the client redirects you to n8n. The alternative is a personal access token.
  5. If you use a cloud-based client, the instance has to be publicly reachable.

Part of these packages

n8n is part of these Walma packages.

Questions about n8n MCP

Is the n8n MCP server free, and which plan is required?+

The server is built into n8n, and the documentation covers it for both n8n Cloud and self-hosted instances. n8n does not state a plan requirement or a price on the documentation page. There are version requirements, though: building and editing workflows needs n8n 2.13.0 or later.

What is the n8n MCP server URL?+

The address is your own instance's address and ends in /mcp-server/http. You copy it in n8n under Settings, Instance-level MCP and Connect a client. There is no shared address at n8n, because the server runs inside each instance.

How does sign-in work?+

In two ways. n8n recommends OAuth: you pick a client, follow its steps and approve access when the client redirects you to n8n. The alternative is a personal access token that is tied to your user and sent as a Bearer token. It is shown only once. If you generate a new one, the old one is revoked, so every client has to be updated.

Is it read-only?+

No. It can run workflows, create and edit them, publish, unpublish and archive. Access is limited per workflow instead: without the Available in MCP setting the client only sees a preview. In a read-only instance, n8n from version 2.47.0 blocks every tool that does not purely read, with one exception: running published workflows.

Which AI assistants does it work with?+

The connect dialog in n8n lists Claude Code, Codex and Gemini CLI, the web clients Claude.ai, ChatGPT and Mistral Vibe, and the IDEs Cursor, VS Code and Windsurf. The documentation also has examples for Lovable, Claude Desktop and agents built with Google ADK. Cloud-based clients need the instance to be publicly reachable.

How does it differ from the MCP Server Trigger node?+

The instance-level server gives one connection for the whole instance, central sign-in and a per-workflow choice of whether it is available. The MCP Server Trigger node is configured inside a single workflow and only exposes that workflow's tools. The first suits an assistant that should work with many workflows, the second a workflow that should be a server of its own.

How do we turn access off again?+

At three levels. Under Connected clients you revoke a single client, which then has to reconnect and sign in again. Clients that use an access token do not appear there. Setting MCP status to Disable disconnects every client. On self-hosted instances, the environment variable N8N_DISABLED_MODULES=mcp removes the feature entirely.

Missing a connection?

We build and review MCP servers for teams that need them, and run them behind the same policy and log as the rest of the gateway.