Heidelberg AICurriculum

n8n Desk: The Open Source Claude Desktop That Runs on top of n8n's new MCP

n8n ·2026-05-11 ·1 min read

Summary written by us from the video's transcript. The video, and everything in it, is n8n's work.

Learn how n8n Desk lets AI agents build and run workflows securely using the new MCP, replacing local code execution of Claude Desktop.

Takeaways

  • Running AI‑generated code in n8n instead of locally prevents accidental destructive actions and keeps credentials safe.
  • The MCP server lets agents create, edit, and execute workflows, turning LLMs into workflow designers rather than just runners.
  • Connecting a custom n8n connector via OAuth provides secure, token‑based access to the MCP endpoint.
  • Rich context (files, skill .md documents) dramatically improves an agent’s ability to generate correct workflows.
  • Visual debugging in n8n makes it easier to pinpoint and fix broken automations compared to raw code.

Why replace Claude Desktop’s local execution

Claude Desktop runs generated code on a user’s machine, which can cause unintended destructive actions. Marcel experienced this with client‑facing errors and recognized the risk for enterprise deployments.

Running code centrally in n8n removes that risk: credentials are encrypted, actions are logged in the n8n audit trail, and any AI model—including local ones—can be used safely.

Introducing n8n Desk

n8n Desk is a free, open‑source desktop app that mimics Claude Desktop’s UI but delegates code execution to n8n workflows.

The desktop app acts as an interface; the heavy lifting happens inside n8n, giving you visual debugging, version control and secure token handling.

The new MCP server

MCP (Multi‑Connector Protocol) is a beta server that extends the n8n API to let agents not only execute workflows but also create and edit them.

Because MCP ships reference documents and skill files, an AI agent can read detailed instructions on how to build correct n8n nodes, reducing trial‑and‑error.

Building a workflow with Cloud Code / Co‑work mode

Marcel demonstrated using the “Co‑work” mode (a variant of Cloud Code) to connect a custom n8n connector to the MCP endpoint.

After adding the connector URL and authenticating via OAuth, the agent could generate a full support‑ticket workflow that reads Slack messages, processes them with an LLM, and replies in Slack.

Context engineering vs. prompt engineering

Effective agents need rich context—large token windows, uploaded files, and skill documents—to understand the task fully.

Marcel highlighted that placing project files in a pinned folder lets the agent automatically ingest all relevant information, improving workflow generation.

Debugging and iterating visual workflows

Unlike generated code, n8n’s visual canvas shows exactly which node failed, making troubleshooting faster for non‑developers.

Users can edit nodes directly, replace an HTTP node with a more suitable alternative, or add custom logic without rewriting code.