n8n and AI: Is there an MCP server for n8n?

n8n describes itself as a workflow automation platform with AI features, built by n8n GmbH in Berlin. You can use it in n8n Cloud or run it yourself. If you want Claude or ChatGPT to work with n8n, the first question is: is there an MCP server for it?

In short

Yes. n8n has an official MCP server that runs in the instance itself. It works in the cloud and when self-hosted. According to n8n, it is included in all plans. According to the product page, it needs n8n 2.18.3 or later. n8n can also act as an MCP server (trigger node) and as an MCP client. According to n8n, n8n Cloud is hosted in Frankfurt.

What n8n offers itself

The instance MCP server. According to the documentation, MCP clients can search workflows, run and test enabled workflows, and create and change workflows and data tables. According to the documentation, building and changing workflows is available from n8n 2.13.0. The product page names n8n 2.18.3 or later for the MCP server and confirms it for the free Community version. The server URL is under Settings > Instance-level MCP and ends in /mcp-server/http. An owner or admin must turn access on. Then the user enables each workflow one by one. The exception is search_workflows: it shows previews of all workflows the user may see. From 2.24.0, access can be switched for a whole project or folder. According to the documentation, n8n is rolling out automatic enabling of new workflows gradually from 2.36.0, off by default.

Sign-in and rights. OAuth is recommended. Alternatively, there is an API key as bearer token. The user grants permissions when connecting. According to the documentation, a client can then, for example, read workflows without being allowed to create or run them. Access can be revoked per client. This does not apply to clients with an API key: they do not appear in the list. All connected clients see the same enabled workflows. According to the documentation, there is no per-client enabling. Admins can restrict the allowed OAuth callback URLs. By default, any URL is allowed.

Tools. The tool list names tools for workflows, executions, credentials (list only), workflow building, agents and data tables. We found no tool there that permanently deletes a workflow. There are archive_workflow (from 2.12.0), delete_agent and delete_data_table_column. According to n8n, agents are a preview feature that may still change. For coding agents, n8n provides guidance at n8n-io/skills (Apache-2.0 according to GitHub).

Claude. On its product page, n8n calls itself an “official Claude Connector”. The Claude directory lists the connector as “Anthropic verified”, made by n8n GmbH, added in January 2026, with sign-in. The page shows 27 tools, fewer than n8n’s documentation names today.

ChatGPT and other clients. According to the documentation, the connection dialog shows Claude.ai, ChatGPT and Mistral Vibe as web clients, Claude Code, Codex and Gemini CLI as CLI clients, and Cursor and VS Code as IDE clients. The selection keeps changing: according to the release notes, Devin replaces Windsurf there, and Antigravity was added. The product page names ChatGPT as a supported client. Our web search found no entry in the ChatGPT directory. We did not check the directory itself.

MCP Server Trigger. The MCP Server Trigger exposes the tools of a single workflow as an MCP server. It supports SSE and streamable HTTP, not stdio. For authentication, choose none, n8n user OAuth2 (node version 2 and later), bearer or header. In queue mode with several webhook replicas, all /mcp* requests must go to one replica. According to the documentation, claude.ai asks for a sign-in to n8n even when authentication is set to none.

MCP client. The MCP Client Tool node connects external MCP servers to n8n agents. It supports bearer, header and OAuth2. According to the documentation, n8n also keeps a registry of MCP servers that connect in the node without your own credentials. Separately, there are two MCP servers for the n8n documentation (GitBook and Kapa.ai). They do not access your instance.

A REST API. According to the documentation, the public API is available on Cloud in Starter, Pro and Enterprise and on all self-hosted editions, not during the trial. Authentication uses an API key in the X-N8N-API-KEY header. According to the documentation, key scopes exist only on the Enterprise plan. There is also an n8n CLI.

AI in the product. The n8n Assistant builds, changes and tests workflows from a chat. According to the documentation, it is available on Cloud in Starter and Pro, and self-hosted in Community, Registered Community and Business. Self-hosted, it needs an API key from Anthropic, OpenAI or OpenRouter and a sandbox. According to the pricing page, it is in preview. According to the AI Terms, n8n does not train ML models on customer content, keeps it only as long as the AI feature needs, and shares it only as far as the AI feature requires at the customer’s direction. According to the AI Terms, n8n may use usage data and prompts internally in aggregated, de-identified form. If you connect your own language models, data goes to those providers.

Cloud or self-hosting

Plans. According to the pricing page, n8n bills by executions: Starter 2,500 and Pro 10,000 per month on Cloud, Business 40,000 self-hosted only, Enterprise by agreement. According to n8n, all plans include unlimited users and workflows. According to n8n, Business is currently not available on Cloud. According to the pricing page, Starter includes 1,600 and Pro up to 9,600 Assistant credits per month.

Data location. According to the pricing page, data on hosted plans is stored in the EU, on servers in Frankfurt. On self-hosted plans, it is stored wherever you run n8n. According to the documentation, the n8n Assistant supports data residency only partly. We found no statement on where the Assistant’s AI models run.

Licence. According to the documentation, the source code is under the Sustainable Use License, and files with .ee. in the name are under the n8n Enterprise License. The Sustainable Use License allows use and modification only for your own internal business purposes or non-commercial use. Distribution is allowed only free of charge and non-commercially. According to n8n, it is not open source in the OSI sense. According to n8n, internal use is fine as long as you do not make n8n available to customers so that they connect their accounts and build workflows. For use the licence does not allow, n8n requires a separate agreement. n8n allows paid consulting and workflow building without an agreement.

What the community offers

czlonkowski/n8n-mcp is an MCP server under the MIT licence, with about 23,000 stars, last updated on 6 November 2026. According to its README, it gives Claude and other assistants documentation on the n8n nodes, plus templates and validation tools. With N8N_API_URL and N8N_API_KEY, it also manages workflows in an instance, including a tool that permanently deletes workflows. According to its README, there is also a hosted variant by the author with a free allowance of 100 calls per day. We did not check further plans. It is not a service of n8n. leonardsellem/n8n-mcp-server (MIT) manages workflows and executions through the n8n API. Its last commit is from 30 June 2025.

Zapier and viaSocket: we found no page for n8n at either. This does not prove that none exists.

What this means for your company

• The instance server can change workflows. When connecting, grant only the rights the client needs, and enable only the workflows it needs.
• According to n8n, cloud clients such as Claude.ai reach only a publicly accessible instance. This also affects self-hosted instances.
• An assistant with an API key or OAuth acts with the user’s rights. Keep test and production workflows apart.
• If you want to run n8n for customers, check the licence first.
• If data must stay in the EU: according to n8n, n8n Cloud is in Frankfurt. What goes to an AI assistant falls under its terms.