How to Build a Personal Assistant in n8n Using MCP

The goal

You want a personal assistant that can check mail-like tasks, files, calendars, and research—not a chat toy that forgets tools exist.

Two complementary pieces:

1. n8n — durable workflows: schedules, webhooks, credentials, multi-app glue.
2. MCP (Model Context Protocol) — standard way for Claude Desktop / Cursor to call tools (filesystem, GitHub, browsers, APIs).

Together: n8n runs background automation; MCP gives your desktop agent live tools; webhooks connect them when needed.

Architecture options

A. Desktop-first assistant (MCP heavy)

• Install MCP servers for filesystem, calendar APIs, browser, etc.
• Chat in Claude Desktop / Cursor.
• Use n8n only for scheduled jobs the agent can trigger via webhook.

Best for: power users at a laptop.

B. Workflow-first assistant (n8n heavy)

• n8n AI Agent / OpenAI nodes orchestrate tools.
• Telegram, Slack, or email as the chat UI.
• MCP optional for local tools.

Best for: always-on bots on a VPS.

C. Hybrid (recommended)

• MCP for interactive desktop work.
• n8n for “every morning brief,” lead alerts, backups.
• Shared secrets in env / credential stores—not in chat history.

Step 1 — Pick assistant jobs

Start with 3 tasks max:

1. Morning brief (calendar + tasks summary).
2. Capture notes → Notion/Drive.
3. Research a URL or topic → short memo.

If a task needs reliability (same time daily), put it in n8n. If it needs conversation, put tools behind MCP.

Step 2 — Install MCP servers

1. Open Skills & MCP catalog.
2. Pick servers (filesystem, GitHub, fetch/browser—start read-only).
3. Copy config JSON into Claude Desktop or Cursor.
4. Follow Install MCP and Getting started with MCP.

Never point a filesystem server at your entire home directory.

Step 3 — Add an n8n “assistant backend”

Typical flow:

1. Webhook or Cron trigger.
2. Gather context (Gmail labels, calendar API, RSS).
3. LLM summarize with a tight system prompt.
4. Deliver — Telegram, Slack DM, or email.
5. Optional: write result to a note app.

Import related recipes from /workflows (search “telegram”, “assistant”, “agent”) and strip unused nodes.

Step 4 — Let the agent call n8n (optional)

Expose a secured webhook in n8n:

• Auth header or secret query param.
• Input schema: { "task": "...", "payload": {} }.
• n8n routes tasks to sub-workflows.

Your MCP-aware client (or a custom MCP server) can POST to that webhook so chat can kick off long-running jobs.

Step 5 — Safety checklist

• Read-only tools first.
• Separate “draft” vs “send” for email/Slack.
• Log every tool call.
• Rate-limit webhooks.
• Prefer Verified blueprints when available; treat community JSON as untrusted code.

Related WorkFlowAI links

• MCP servers for Claude & Cursor
• n8n import docs
• Best MCP servers 2026
• LangChain automation guide (when you need custom agent code)

Bottom line: A real personal assistant is tools + memory + safe automation. Use MCP for interactive tools, n8n for schedules and multi-app side effects, and grow from three jobs—not thirty.